You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

452 lines
12 KiB

#ifndef PACK_OBJECTS_H
#define PACK_OBJECTS_H
#include "object-store.h"
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
#include "thread-utils.h"
#include "pack.h"
struct repository;
#define DEFAULT_DELTA_CACHE_SIZE (256 * 1024 * 1024)
#define OE_DFS_STATE_BITS 2
#define OE_DEPTH_BITS 12
#define OE_IN_PACK_BITS 10
#define OE_Z_DELTA_BITS 20
/*
* Note that oe_set_size() becomes expensive when the given size is
* above this limit. Don't lower it too much.
*/
#define OE_SIZE_BITS 31
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
#define OE_DELTA_SIZE_BITS 23
/*
* State flags for depth-first search used for analyzing delta cycles.
*
* The depth is measured in delta-links to the base (so if A is a delta
* against B, then A has a depth of 1, and B a depth of 0).
*/
enum dfs_state {
DFS_NONE = 0,
DFS_ACTIVE,
DFS_DONE,
DFS_NUM_STATES
};
/*
* The size of struct nearly determines pack-objects's memory
* consumption. This struct is packed tight for that reason. When you
* add or reorder something in this struct, think a bit about this.
*
* basic object info
* -----------------
* idx.oid is filled up before delta searching starts. idx.crc32 is
* only valid after the object is written out and will be used for
* generating the index. idx.offset will be both gradually set and
* used in writing phase (base objects get offset first, then deltas
* refer to them)
*
* "size" is the uncompressed object size. Compressed size of the raw
* data for an object in a pack is not stored anywhere but is computed
* and made available when reverse .idx is made. Note that when a
* delta is reused, "size" is the uncompressed _delta_ size, not the
* canonical one after the delta has been applied.
*
* "hash" contains a path name hash which is used for sorting the
* delta list and also during delta searching. Once prepare_pack()
* returns it's no longer needed.
*
* source pack info
* ----------------
* The (in_pack, in_pack_offset) tuple contains the location of the
* object in the source pack. in_pack_header_size allows quickly
* skipping the header and going straight to the zlib stream.
*
* "type" and "in_pack_type" both describe object type. in_pack_type
* may contain a delta type, while type is always the canonical type.
*
* deltas
* ------
* Delta links (delta, delta_child and delta_sibling) are created to
* reflect that delta graph from the source pack then updated or added
* during delta searching phase when we find better deltas.
*
* delta_child and delta_sibling are last needed in
* compute_write_order(). "delta" and "delta_size" must remain valid
* at object writing phase in case the delta is not cached.
*
* If a delta is cached in memory and is compressed, delta_data points
* to the data and z_delta_size contains the compressed size. If it's
* uncompressed [1], z_delta_size must be zero. delta_size is always
* the uncompressed size and must be valid even if the delta is not
* cached.
*
* [1] during try_delta phase we don't bother with compressing because
* the delta could be quickly replaced with a better one.
*/
struct object_entry {
struct pack_idx_entry idx;
void *delta_data; /* cached delta (uncompressed) */
off_t in_pack_offset;
uint32_t hash; /* name hint hash */
unsigned size_:OE_SIZE_BITS;
unsigned size_valid:1;
uint32_t delta_idx; /* delta base object */
uint32_t delta_child_idx; /* deltified objects who bases me */
uint32_t delta_sibling_idx; /* other deltified objects who
* uses the same base as me
*/
unsigned delta_size_:OE_DELTA_SIZE_BITS; /* delta data size (uncompressed) */
unsigned delta_size_valid:1;
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
unsigned char in_pack_header_size;
unsigned in_pack_idx:OE_IN_PACK_BITS; /* already in pack */
unsigned z_delta_size:OE_Z_DELTA_BITS;
unsigned type_valid:1;
unsigned no_try_delta:1;
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
unsigned type_:TYPE_BITS;
unsigned in_pack_type:TYPE_BITS; /* could be delta */
6 years ago
unsigned preferred_base:1; /*
* we do not pack this, but is available
* to be used as the base object to delta
* objects against.
*/
unsigned tagged:1; /* near the very tip of refs */
unsigned filled:1; /* assigned write-order */
unsigned dfs_state:OE_DFS_STATE_BITS;
unsigned depth:OE_DEPTH_BITS;
pack-objects: reuse on-disk deltas for thin "have" objects When we serve a fetch, we pass the "wants" and "haves" from the fetch negotiation to pack-objects. That tells us not only which objects we need to send, but we also use the boundary commits as "preferred bases": their trees and blobs are candidates for delta bases, both for reusing on-disk deltas and for finding new ones. However, this misses some opportunities. Modulo some special cases like shallow or partial clones, we know that every object reachable from the "haves" could be a preferred base. We don't use all of them for two reasons: 1. It's expensive to traverse the whole history and enumerate all of the objects the other side has. 2. The delta search is expensive, so we want to keep the number of candidate bases sane. The boundary commits are the most likely to work. When we have reachability bitmaps, though, reason 1 no longer applies. We can efficiently compute the set of reachable objects on the other side (and in fact already did so as part of the bitmap set-difference to get the list of interesting objects). And using this set conveniently covers the shallow and partial cases, since we have to disable the use of bitmaps for those anyway. The second reason argues against using these bases in the search for new deltas. But there's one case where we can use this information for free: when we have an existing on-disk delta that we're considering reusing, we can do so if we know the other side has the base object. This in fact saves time during the delta search, because it's one less delta we have to compute. And that's exactly what this patch does: when we're considering whether to reuse an on-disk delta, if bitmaps tell us the other side has the object (and we're making a thin-pack), then we reuse it. Here are the results on p5311 using linux.git, which simulates a client fetching after `N` days since their last fetch: Test origin HEAD -------------------------------------------------------------------------- 5311.3: server (1 days) 0.27(0.27+0.04) 0.12(0.09+0.03) -55.6% 5311.4: size (1 days) 0.9M 237.0K -73.7% 5311.5: client (1 days) 0.04(0.05+0.00) 0.10(0.10+0.00) +150.0% 5311.7: server (2 days) 0.34(0.42+0.04) 0.13(0.10+0.03) -61.8% 5311.8: size (2 days) 1.5M 347.7K -76.5% 5311.9: client (2 days) 0.07(0.08+0.00) 0.16(0.15+0.01) +128.6% 5311.11: server (4 days) 0.56(0.77+0.08) 0.13(0.10+0.02) -76.8% 5311.12: size (4 days) 2.8M 566.6K -79.8% 5311.13: client (4 days) 0.13(0.15+0.00) 0.34(0.31+0.02) +161.5% 5311.15: server (8 days) 0.97(1.39+0.11) 0.30(0.25+0.05) -69.1% 5311.16: size (8 days) 4.3M 1.0M -76.0% 5311.17: client (8 days) 0.20(0.22+0.01) 0.53(0.52+0.01) +165.0% 5311.19: server (16 days) 1.52(2.51+0.12) 0.30(0.26+0.03) -80.3% 5311.20: size (16 days) 8.0M 2.0M -74.5% 5311.21: client (16 days) 0.40(0.47+0.03) 1.01(0.98+0.04) +152.5% 5311.23: server (32 days) 2.40(4.44+0.20) 0.31(0.26+0.04) -87.1% 5311.24: size (32 days) 14.1M 4.1M -70.9% 5311.25: client (32 days) 0.70(0.90+0.03) 1.81(1.75+0.06) +158.6% 5311.27: server (64 days) 11.76(26.57+0.29) 0.55(0.50+0.08) -95.3% 5311.28: size (64 days) 89.4M 47.4M -47.0% 5311.29: client (64 days) 5.71(9.31+0.27) 15.20(15.20+0.32) +166.2% 5311.31: server (128 days) 16.15(36.87+0.40) 0.91(0.82+0.14) -94.4% 5311.32: size (128 days) 134.8M 100.4M -25.5% 5311.33: client (128 days) 9.42(16.86+0.49) 25.34(25.80+0.46) +169.0% In all cases we save CPU time on the server (sometimes significant) and the resulting pack is smaller. We do spend more CPU time on the client side, because it has to reconstruct more deltas. But that's the right tradeoff to make, since clients tend to outnumber servers. It just means the thin pack mechanism is doing its job. From the user's perspective, the end-to-end time of the operation will generally be faster. E.g., in the 128-day case, we saved 15s on the server at a cost of 16s on the client. Since the resulting pack is 34MB smaller, this is a net win if the network speed is less than 270Mbit/s. And that's actually the worst case. The 64-day case saves just over 11s at a cost of just under 11s. So it's a slight win at any network speed, and the 40MB saved is pure bonus. That trend continues for the smaller fetches. The implementation itself is mostly straightforward, with the new logic going into check_object(). But there are two tricky bits. The first is that check_object() needs access to the relevant information (the thin flag and bitmap result). We can do this by pushing these into program-lifetime globals. The second is that the rest of the code assumes that any reused delta will point to another "struct object_entry" as its base. But of course the case we are interested in here is the one where don't have such an entry! I looked at a number of options that didn't quite work: - we could use a flag to signal a reused delta, but it's not a single bit. We have to actually store the oid of the base, which is normally done by pointing to the existing object_entry. And we'd have to modify all the code which looks at deltas. - we could add the reused bases to the end of the existing object_entry array. While this does create some extra work as later stages consider the extra entries, it's actually not too bad (we're not sending them, so they don't cost much in the delta search, and at most we'd have 2*N of them). But there's a more subtle problem. Adding to the existing array means we might need to grow it with realloc, which could move the earlier entries around. While many of the references to other entries are done by integer index, some (including ones on the stack) use pointers, which would become invalidated. This isn't insurmountable, but it would require quite a bit of refactoring (and it's hard to know that you've got it all, since it may work _most_ of the time and then fail subtly based on memory allocation patterns). - we could allocate a new one-off entry for the base. In fact, this is what an earlier version of this patch did. However, since the refactoring brought in by ad635e82d6 (Merge branch 'nd/pack-objects-pack-struct', 2018-05-23), the delta_idx code requires that both entries be in the main packing list. So taking all of those options into account, what I ended up with is a separate list of "external bases" that are not part of the main packing list. Each delta entry that points to an external base has a single-bit flag to do so; we have a little breathing room in the bitfield section of object_entry. This lets us limit the change primarily to the oe_delta() and oe_set_delta_ext() functions. And as a bonus, most of the rest of the code does not consider these dummy entries at all, saving both runtime CPU and code complexity. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years ago
unsigned ext_base:1; /* delta_idx points outside packlist */
pack-objects: break delta cycles before delta-search phase We do not allow cycles in the delta graph of a pack (i.e., A is a delta of B which is a delta of A) for the obvious reason that you cannot actually access any of the objects in such a case. There's a last-ditch attempt to notice cycles during the write phase, during which we issue a warning to the user and write one of the objects out in full. However, this is "last-ditch" for two reasons: 1. By this time, it's too late to find another delta for the object, so the resulting pack is larger than it otherwise could be. 2. The warning is there because this is something that _shouldn't_ ever happen. If it does, then either: a. a pack we are reusing deltas from had its own cycle b. we are reusing deltas from multiple packs, and we found a cycle among them (i.e., A is a delta of B in one pack, but B is a delta of A in another, and we choose to use both deltas). c. there is a bug in the delta-search code So this code serves as a final check that none of these things has happened, warns the user, and prevents us from writing a bogus pack. Right now, (2b) should never happen because of the static ordering of packs in want_object_in_pack(). If two objects have a delta relationship, then they must be in the same pack, and therefore we will find them from that same pack. However, a future patch would like to change that static ordering, which will make (2b) a common occurrence. In preparation, we should be able to handle those kinds of cycles better. This patch does by introducing a cycle-breaking step during the get_object_details() phase, when we are deciding which deltas can be reused. That gives us the chance to feed the objects into the delta search as if the cycle did not exist. We'll leave the detection and warning in the write_object() phase in place, as it still serves as a check for case (2c). This does mean we will stop warning for (2a). That case is caused by bogus input packs, and we ideally would warn the user about it. However, since those cycles show up after picking reusable deltas, they look the same as (2b) to us; our new code will break the cycles early and the last-ditch check will never see them. We could do analysis on any cycles that we find to distinguish the two cases (i.e., it is a bogus pack if and only if every delta in the cycle is in the same pack), but we don't need to. If there is a cycle inside a pack, we'll run into problems not only reusing the delta, but accessing the object data at all. So when we try to dig up the actual size of the object, we'll hit that same cycle and kick in our usual complain-and-try-another-source code. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
9 years ago
/*
* pahole results on 64-bit linux (gcc and clang)
pack-objects: enforce --depth limit in reused deltas Since 898b14c (pack-objects: rework check_delta_limit usage, 2007-04-16), we check the delta depth limit only when figuring out whether we should make a new delta. We don't consider it at all when reusing deltas, which means that packing once with --depth=250, and then again with --depth=50, the second pack may still contain chains larger than 50. This is generally considered a feature, as the results of earlier high-depth repacks are carried forward, used for serving fetches, etc. However, since we started using cross-pack deltas in c9af708b1 (pack-objects: use mru list when iterating over packs, 2016-08-11), we are no longer bounded by the length of an existing delta chain in a single pack. Here's one particular pathological case: a sequence of N packs, each with 2 objects, the base of which is stored as a delta in a previous pack. If we chain all the deltas together, we have a cycle of length N. We break the cycle, but the tip delta is still at depth N-1. This is less unlikely than it might sound. See the included test for a reconstruction based on real-world actions. I ran into such a case in the wild, where a client was rapidly sending packs, and we had accumulated 10,000 before doing a server-side repack. The pack that "git repack" tried to generate had a very deep chain, which caused pack-objects to run out of stack space in the recursive write_one(). This patch bounds the length of delta chains in the output pack based on --depth, regardless of whether they are caused by cross-pack deltas or existed in the input packs. This fixes the problem, but does have two possible downsides: 1. High-depth aggressive repacks followed by "normal" repacks will throw away the high-depth chains. In the long run this is probably OK; investigation showed that high-depth repacks aren't actually beneficial, and we dropped the aggressive depth default to match the normal case in 07e7dbf0d (gc: default aggressive depth to 50, 2016-08-11). 2. If you really do want to store high-depth deltas on disk, they may be discarded and new delta computed when serving a fetch, unless you set pack.depth to match your high-depth size. The implementation uses the existing search for delta cycles. That lets us compute the depth of any node based on the depth of its base, because we know the base is DFS_DONE by the time we look at it (modulo any cycles in the graph, but we know there cannot be any because we break them as we see them). There is some subtlety worth mentioning, though. We record the depth of each object as we compute it. It might seem like we could save the per-object storage space by just keeping track of the depth of our traversal (i.e., have break_delta_chains() report how deep it went). But we may visit an object through multiple delta paths, and on subsequent paths we want to know its depth immediately, without having to walk back down to its final base (doing so would make our graph walk quadratic rather than linear). Likewise, one could try to record the depth not from the base, but from our starting point (i.e., start recursion_depth at 0, and pass "recursion_depth + 1" to each invocation of break_delta_chains()). And then when recursion_depth gets too big, we know that we must cut the delta chain. But that technique is wrong if we do not visit the nodes in topological order. In a chain A->B->C, it if we visit "C", then "B", then "A", we will never recurse deeper than 1 link (because we see at each node that we have already visited it). Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
8 years ago
*
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
* size: 80, bit_padding: 9 bits
*
* and on 32-bit (gcc)
*
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
* size: 76, bit_padding: 9 bits
pack-objects: break delta cycles before delta-search phase We do not allow cycles in the delta graph of a pack (i.e., A is a delta of B which is a delta of A) for the obvious reason that you cannot actually access any of the objects in such a case. There's a last-ditch attempt to notice cycles during the write phase, during which we issue a warning to the user and write one of the objects out in full. However, this is "last-ditch" for two reasons: 1. By this time, it's too late to find another delta for the object, so the resulting pack is larger than it otherwise could be. 2. The warning is there because this is something that _shouldn't_ ever happen. If it does, then either: a. a pack we are reusing deltas from had its own cycle b. we are reusing deltas from multiple packs, and we found a cycle among them (i.e., A is a delta of B in one pack, but B is a delta of A in another, and we choose to use both deltas). c. there is a bug in the delta-search code So this code serves as a final check that none of these things has happened, warns the user, and prevents us from writing a bogus pack. Right now, (2b) should never happen because of the static ordering of packs in want_object_in_pack(). If two objects have a delta relationship, then they must be in the same pack, and therefore we will find them from that same pack. However, a future patch would like to change that static ordering, which will make (2b) a common occurrence. In preparation, we should be able to handle those kinds of cycles better. This patch does by introducing a cycle-breaking step during the get_object_details() phase, when we are deciding which deltas can be reused. That gives us the chance to feed the objects into the delta search as if the cycle did not exist. We'll leave the detection and warning in the write_object() phase in place, as it still serves as a check for case (2c). This does mean we will stop warning for (2a). That case is caused by bogus input packs, and we ideally would warn the user about it. However, since those cycles show up after picking reusable deltas, they look the same as (2b) to us; our new code will break the cycles early and the last-ditch check will never see them. We could do analysis on any cycles that we find to distinguish the two cases (i.e., it is a bogus pack if and only if every delta in the cycle is in the same pack), but we don't need to. If there is a cycle inside a pack, we'll run into problems not only reusing the delta, but accessing the object data at all. So when we try to dig up the actual size of the object, we'll hit that same cycle and kick in our usual complain-and-try-another-source code. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
9 years ago
*/
};
struct packing_data {
struct repository *repo;
struct object_entry *objects;
uint32_t nr_objects, nr_alloc;
int32_t *index;
uint32_t index_size;
unsigned int *in_pack_pos;
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
unsigned long *delta_size;
/*
* Only one of these can be non-NULL and they have different
* sizes. if in_pack_by_idx is allocated, oe_in_pack() returns
* the pack of an object using in_pack_idx field. If not,
* in_pack[] array is used the same way as in_pack_pos[]
*/
struct packed_git **in_pack_by_idx;
struct packed_git **in_pack;
/*
* During packing with multiple threads, protect the in-core
* object database from concurrent accesses.
*/
pthread_mutex_t odb_lock;
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
pack-objects: reuse on-disk deltas for thin "have" objects When we serve a fetch, we pass the "wants" and "haves" from the fetch negotiation to pack-objects. That tells us not only which objects we need to send, but we also use the boundary commits as "preferred bases": their trees and blobs are candidates for delta bases, both for reusing on-disk deltas and for finding new ones. However, this misses some opportunities. Modulo some special cases like shallow or partial clones, we know that every object reachable from the "haves" could be a preferred base. We don't use all of them for two reasons: 1. It's expensive to traverse the whole history and enumerate all of the objects the other side has. 2. The delta search is expensive, so we want to keep the number of candidate bases sane. The boundary commits are the most likely to work. When we have reachability bitmaps, though, reason 1 no longer applies. We can efficiently compute the set of reachable objects on the other side (and in fact already did so as part of the bitmap set-difference to get the list of interesting objects). And using this set conveniently covers the shallow and partial cases, since we have to disable the use of bitmaps for those anyway. The second reason argues against using these bases in the search for new deltas. But there's one case where we can use this information for free: when we have an existing on-disk delta that we're considering reusing, we can do so if we know the other side has the base object. This in fact saves time during the delta search, because it's one less delta we have to compute. And that's exactly what this patch does: when we're considering whether to reuse an on-disk delta, if bitmaps tell us the other side has the object (and we're making a thin-pack), then we reuse it. Here are the results on p5311 using linux.git, which simulates a client fetching after `N` days since their last fetch: Test origin HEAD -------------------------------------------------------------------------- 5311.3: server (1 days) 0.27(0.27+0.04) 0.12(0.09+0.03) -55.6% 5311.4: size (1 days) 0.9M 237.0K -73.7% 5311.5: client (1 days) 0.04(0.05+0.00) 0.10(0.10+0.00) +150.0% 5311.7: server (2 days) 0.34(0.42+0.04) 0.13(0.10+0.03) -61.8% 5311.8: size (2 days) 1.5M 347.7K -76.5% 5311.9: client (2 days) 0.07(0.08+0.00) 0.16(0.15+0.01) +128.6% 5311.11: server (4 days) 0.56(0.77+0.08) 0.13(0.10+0.02) -76.8% 5311.12: size (4 days) 2.8M 566.6K -79.8% 5311.13: client (4 days) 0.13(0.15+0.00) 0.34(0.31+0.02) +161.5% 5311.15: server (8 days) 0.97(1.39+0.11) 0.30(0.25+0.05) -69.1% 5311.16: size (8 days) 4.3M 1.0M -76.0% 5311.17: client (8 days) 0.20(0.22+0.01) 0.53(0.52+0.01) +165.0% 5311.19: server (16 days) 1.52(2.51+0.12) 0.30(0.26+0.03) -80.3% 5311.20: size (16 days) 8.0M 2.0M -74.5% 5311.21: client (16 days) 0.40(0.47+0.03) 1.01(0.98+0.04) +152.5% 5311.23: server (32 days) 2.40(4.44+0.20) 0.31(0.26+0.04) -87.1% 5311.24: size (32 days) 14.1M 4.1M -70.9% 5311.25: client (32 days) 0.70(0.90+0.03) 1.81(1.75+0.06) +158.6% 5311.27: server (64 days) 11.76(26.57+0.29) 0.55(0.50+0.08) -95.3% 5311.28: size (64 days) 89.4M 47.4M -47.0% 5311.29: client (64 days) 5.71(9.31+0.27) 15.20(15.20+0.32) +166.2% 5311.31: server (128 days) 16.15(36.87+0.40) 0.91(0.82+0.14) -94.4% 5311.32: size (128 days) 134.8M 100.4M -25.5% 5311.33: client (128 days) 9.42(16.86+0.49) 25.34(25.80+0.46) +169.0% In all cases we save CPU time on the server (sometimes significant) and the resulting pack is smaller. We do spend more CPU time on the client side, because it has to reconstruct more deltas. But that's the right tradeoff to make, since clients tend to outnumber servers. It just means the thin pack mechanism is doing its job. From the user's perspective, the end-to-end time of the operation will generally be faster. E.g., in the 128-day case, we saved 15s on the server at a cost of 16s on the client. Since the resulting pack is 34MB smaller, this is a net win if the network speed is less than 270Mbit/s. And that's actually the worst case. The 64-day case saves just over 11s at a cost of just under 11s. So it's a slight win at any network speed, and the 40MB saved is pure bonus. That trend continues for the smaller fetches. The implementation itself is mostly straightforward, with the new logic going into check_object(). But there are two tricky bits. The first is that check_object() needs access to the relevant information (the thin flag and bitmap result). We can do this by pushing these into program-lifetime globals. The second is that the rest of the code assumes that any reused delta will point to another "struct object_entry" as its base. But of course the case we are interested in here is the one where don't have such an entry! I looked at a number of options that didn't quite work: - we could use a flag to signal a reused delta, but it's not a single bit. We have to actually store the oid of the base, which is normally done by pointing to the existing object_entry. And we'd have to modify all the code which looks at deltas. - we could add the reused bases to the end of the existing object_entry array. While this does create some extra work as later stages consider the extra entries, it's actually not too bad (we're not sending them, so they don't cost much in the delta search, and at most we'd have 2*N of them). But there's a more subtle problem. Adding to the existing array means we might need to grow it with realloc, which could move the earlier entries around. While many of the references to other entries are done by integer index, some (including ones on the stack) use pointers, which would become invalidated. This isn't insurmountable, but it would require quite a bit of refactoring (and it's hard to know that you've got it all, since it may work _most_ of the time and then fail subtly based on memory allocation patterns). - we could allocate a new one-off entry for the base. In fact, this is what an earlier version of this patch did. However, since the refactoring brought in by ad635e82d6 (Merge branch 'nd/pack-objects-pack-struct', 2018-05-23), the delta_idx code requires that both entries be in the main packing list. So taking all of those options into account, what I ended up with is a separate list of "external bases" that are not part of the main packing list. Each delta entry that points to an external base has a single-bit flag to do so; we have a little breathing room in the bitfield section of object_entry. This lets us limit the change primarily to the oe_delta() and oe_set_delta_ext() functions. And as a bonus, most of the rest of the code does not consider these dummy entries at all, saving both runtime CPU and code complexity. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years ago
/*
* This list contains entries for bases which we know the other side
* has (e.g., via reachability bitmaps), but which aren't in our
* "objects" list.
*/
struct object_entry *ext_bases;
uint32_t nr_ext, alloc_ext;
uintmax_t oe_size_limit;
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
uintmax_t oe_delta_size_limit;
/* delta islands */
unsigned int *tree_depth;
unsigned char *layer;
};
void prepare_packing_data(struct repository *r, struct packing_data *pdata);
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
/* Protect access to object database */
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
static inline void packing_data_lock(struct packing_data *pdata)
{
pthread_mutex_lock(&pdata->odb_lock);
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
}
static inline void packing_data_unlock(struct packing_data *pdata)
{
pthread_mutex_unlock(&pdata->odb_lock);
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
}
struct object_entry *packlist_alloc(struct packing_data *pdata,
pack-objects: drop packlist index_pos optimization Once upon a time, the code to add an object to our packing list in pack-objects all lived in a single function. It computed the position within the hash table once, then used it to check if the object was already present, and if not, to add it. Later, in 2834bc27c1 (pack-objects: refactor the packing list, 2013-10-24), this was split into two functions: packlist_find() and packlist_alloc(). We ended up with an "index_pos" variable that gets passed through several functions to make it from one to the other. The resulting code is rather confusing to follow. The "index_pos" variable is sometimes undefined, if we don't yet have a hash table. This works out in practice because in that case packlist_alloc() won't use it at all, since it will have to create/grow the hash table. But it's hard to verify that, and it does cause gcc 9.2.1's -Wmaybe-uninitialized to complain when compiled with "-flto -O3" (rightfully, since we do pass the uninitialized value as a function parameter, even if nobody ends up using it). All of this is to save computing the hash index again when we're inserting into the hash table, which I found doesn't make a measurable difference in the program runtime (which is not surprising, since we're doing all kinds of other heavyweight things for each object). Let's just drop this index_pos variable entirely, simplifying the code (and pleasing the compiler). We might be better still refactoring this custom hash table to use one of our existing implementations (an oidmap, or a kh_oid_map). I stopped short of that here, but this would be the likely first step towards that anyway. Reported-by: Stephan Beyer <s-beyer@gmx.net> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
5 years ago
const struct object_id *oid);
struct object_entry *packlist_find(struct packing_data *pdata,
pack-objects: drop packlist index_pos optimization Once upon a time, the code to add an object to our packing list in pack-objects all lived in a single function. It computed the position within the hash table once, then used it to check if the object was already present, and if not, to add it. Later, in 2834bc27c1 (pack-objects: refactor the packing list, 2013-10-24), this was split into two functions: packlist_find() and packlist_alloc(). We ended up with an "index_pos" variable that gets passed through several functions to make it from one to the other. The resulting code is rather confusing to follow. The "index_pos" variable is sometimes undefined, if we don't yet have a hash table. This works out in practice because in that case packlist_alloc() won't use it at all, since it will have to create/grow the hash table. But it's hard to verify that, and it does cause gcc 9.2.1's -Wmaybe-uninitialized to complain when compiled with "-flto -O3" (rightfully, since we do pass the uninitialized value as a function parameter, even if nobody ends up using it). All of this is to save computing the hash index again when we're inserting into the hash table, which I found doesn't make a measurable difference in the program runtime (which is not surprising, since we're doing all kinds of other heavyweight things for each object). Let's just drop this index_pos variable entirely, simplifying the code (and pleasing the compiler). We might be better still refactoring this custom hash table to use one of our existing implementations (an oidmap, or a kh_oid_map). I stopped short of that here, but this would be the likely first step towards that anyway. Reported-by: Stephan Beyer <s-beyer@gmx.net> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
5 years ago
const struct object_id *oid);
static inline uint32_t pack_name_hash(const char *name)
{
uint32_t c, hash = 0;
if (!name)
return 0;
/*
* This effectively just creates a sortable number from the
* last sixteen non-whitespace characters. Last characters
* count "most", so things that end in ".c" sort together.
*/
while ((c = *name++) != 0) {
if (isspace(c))
continue;
hash = (hash >> 2) + (c << 24);
}
return hash;
}
static inline enum object_type oe_type(const struct object_entry *e)
{
return e->type_valid ? e->type_ : OBJ_BAD;
}
static inline void oe_set_type(struct object_entry *e,
enum object_type type)
{
if (type >= OBJ_ANY)
BUG("OBJ_ANY cannot be set in pack-objects code");
e->type_valid = type >= OBJ_NONE;
e->type_ = (unsigned)type;
}
static inline unsigned int oe_in_pack_pos(const struct packing_data *pack,
const struct object_entry *e)
{
return pack->in_pack_pos[e - pack->objects];
}
static inline void oe_set_in_pack_pos(const struct packing_data *pack,
const struct object_entry *e,
unsigned int pos)
{
pack->in_pack_pos[e - pack->objects] = pos;
}
static inline struct packed_git *oe_in_pack(const struct packing_data *pack,
const struct object_entry *e)
{
if (pack->in_pack_by_idx)
return pack->in_pack_by_idx[e->in_pack_idx];
else
return pack->in_pack[e - pack->objects];
}
void oe_map_new_pack(struct packing_data *pack);
static inline void oe_set_in_pack(struct packing_data *pack,
struct object_entry *e,
struct packed_git *p)
{
pack-objects: avoid pointless oe_map_new_pack() calls This patch fixes an extreme slowdown in pack-objects when you have more than 1023 packs. See below for numbers. Since 43fa44fa3b (pack-objects: move in_pack out of struct object_entry, 2018-04-14), we use a complicated system to save some per-object memory. Each object_entry structs gets a 10-bit field to store the index of the pack it's in. We map those indices into pointers using packing_data->in_pack_by_idx, which we initialize at the start of the program. If we have 2^10 or more packs, then we instead create an array of pack pointers, one per object. This is packing_data->in_pack. So far so good. But there's one other tricky case: if a new pack arrives after we've initialized in_pack_by_idx, it won't have an index yet. We solve that by calling oe_map_new_pack(), which just switches on the fly to the less-optimal in_pack mechanism, allocating the array and back-filling it for already-seen objects. But that logic kicks in even when we've switched to it already (whether because we really did see a new pack, or because we had too many packs in the first place). The result doesn't produce a wrong outcome, but it's very slow. What happens is this: - imagine you have a repo with 500k objects and 2000 packs that you want to repack. - before looking at any objects, we call prepare_in_pack_by_idx(). It starts allocating an index for each pack. On the 1024th pack, it sees there are too many, so it bails, leaving in_pack_by_idx as NULL. - while actually adding objects to the packing list, we call oe_set_in_pack(), which checks whether the pack already has an index. If it's one of the packs after the first 1023, then it doesn't have one, and we'll call oe_map_new_pack(). But there's no useful work for that function to do. We're already using in_pack, so it just uselessly walks over the complete list of objects, trying to backfill in_pack. And we end up doing this for almost 1000 packs (each of which may be triggered by more than one object). And each time it triggers, we may iterate over up to 500k objects. So in the absolute worst case, this is quadratic in the number of objects. The solution is simple: we don't need to bother checking whether the pack has an index if we've already converted to using in_pack, since by definition we're not going to use it. So we can just push the "does the pack have a valid index" check down into that half of the conditional, where we know we're going to use it. The current test in p5303 sadly doesn't notice this problem, since it maxes out at 1000 packs. If we add a new test to it at 2000 packs, it does show the improvement: Test HEAD^ HEAD ---------------------------------------------------------------------- 5303.12: repack (2000) 26.72(39.68+0.67) 15.70(28.70+0.66) -41.2% However, these many-pack test cases are rather expensive to run, so adding larger and larger numbers isn't appealing. Instead, we can show it off more easily by using GIT_TEST_FULL_IN_PACK_ARRAY, which forces us into the absolute worst case: no pack has an index, so we'll trigger oe_map_new_pack() pointlessly for every single object, making it truly quadratic. Here are the numbers (on git.git) with the included change to p5303: Test HEAD^ HEAD ---------------------------------------------------------------------- 5303.3: rev-list (1) 2.05(1.98+0.06) 2.06(1.99+0.06) +0.5% 5303.4: repack (1) 33.45(33.46+0.19) 2.75(2.73+0.22) -91.8% 5303.6: rev-list (50) 2.07(2.01+0.06) 2.06(2.01+0.05) -0.5% 5303.7: repack (50) 34.21(35.18+0.16) 3.49(4.50+0.12) -89.8% 5303.9: rev-list (1000) 2.87(2.78+0.08) 2.88(2.80+0.07) +0.3% 5303.10: repack (1000) 41.26(51.30+0.47) 10.75(20.75+0.44) -73.9% Again, those improvements aren't realistic for the 1-pack case (because in the real world, the full-array solution doesn't kick in), but it's more useful to be testing the more-complicated code path. While we're looking at this issue, we'll tweak one more thing: in oe_map_new_pack(), we call REALLOC_ARRAY(pack->in_pack). But we'd never expect to get here unless we're back-filling it for the first time, in which case it would be NULL. So let's switch that to ALLOC_ARRAY() for clarity, and add a BUG() to document the expectation. Unfortunately this code isn't well-covered in the test suite because it's inherently racy (it only kicks in if somebody else adds a new pack while we're in the middle of repacking). Signed-off-by: Jeff King <peff@peff.net> Reviewed-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
5 years ago
if (pack->in_pack_by_idx) {
if (p->index) {
e->in_pack_idx = p->index;
return;
}
/*
* We're accessing packs by index, but this pack doesn't have
* an index (e.g., because it was added since we created the
* in_pack_by_idx array). Bail to oe_map_new_pack(), which
* will convert us to using the full in_pack array, and then
* fall through to our in_pack handling.
*/
oe_map_new_pack(pack);
pack-objects: avoid pointless oe_map_new_pack() calls This patch fixes an extreme slowdown in pack-objects when you have more than 1023 packs. See below for numbers. Since 43fa44fa3b (pack-objects: move in_pack out of struct object_entry, 2018-04-14), we use a complicated system to save some per-object memory. Each object_entry structs gets a 10-bit field to store the index of the pack it's in. We map those indices into pointers using packing_data->in_pack_by_idx, which we initialize at the start of the program. If we have 2^10 or more packs, then we instead create an array of pack pointers, one per object. This is packing_data->in_pack. So far so good. But there's one other tricky case: if a new pack arrives after we've initialized in_pack_by_idx, it won't have an index yet. We solve that by calling oe_map_new_pack(), which just switches on the fly to the less-optimal in_pack mechanism, allocating the array and back-filling it for already-seen objects. But that logic kicks in even when we've switched to it already (whether because we really did see a new pack, or because we had too many packs in the first place). The result doesn't produce a wrong outcome, but it's very slow. What happens is this: - imagine you have a repo with 500k objects and 2000 packs that you want to repack. - before looking at any objects, we call prepare_in_pack_by_idx(). It starts allocating an index for each pack. On the 1024th pack, it sees there are too many, so it bails, leaving in_pack_by_idx as NULL. - while actually adding objects to the packing list, we call oe_set_in_pack(), which checks whether the pack already has an index. If it's one of the packs after the first 1023, then it doesn't have one, and we'll call oe_map_new_pack(). But there's no useful work for that function to do. We're already using in_pack, so it just uselessly walks over the complete list of objects, trying to backfill in_pack. And we end up doing this for almost 1000 packs (each of which may be triggered by more than one object). And each time it triggers, we may iterate over up to 500k objects. So in the absolute worst case, this is quadratic in the number of objects. The solution is simple: we don't need to bother checking whether the pack has an index if we've already converted to using in_pack, since by definition we're not going to use it. So we can just push the "does the pack have a valid index" check down into that half of the conditional, where we know we're going to use it. The current test in p5303 sadly doesn't notice this problem, since it maxes out at 1000 packs. If we add a new test to it at 2000 packs, it does show the improvement: Test HEAD^ HEAD ---------------------------------------------------------------------- 5303.12: repack (2000) 26.72(39.68+0.67) 15.70(28.70+0.66) -41.2% However, these many-pack test cases are rather expensive to run, so adding larger and larger numbers isn't appealing. Instead, we can show it off more easily by using GIT_TEST_FULL_IN_PACK_ARRAY, which forces us into the absolute worst case: no pack has an index, so we'll trigger oe_map_new_pack() pointlessly for every single object, making it truly quadratic. Here are the numbers (on git.git) with the included change to p5303: Test HEAD^ HEAD ---------------------------------------------------------------------- 5303.3: rev-list (1) 2.05(1.98+0.06) 2.06(1.99+0.06) +0.5% 5303.4: repack (1) 33.45(33.46+0.19) 2.75(2.73+0.22) -91.8% 5303.6: rev-list (50) 2.07(2.01+0.06) 2.06(2.01+0.05) -0.5% 5303.7: repack (50) 34.21(35.18+0.16) 3.49(4.50+0.12) -89.8% 5303.9: rev-list (1000) 2.87(2.78+0.08) 2.88(2.80+0.07) +0.3% 5303.10: repack (1000) 41.26(51.30+0.47) 10.75(20.75+0.44) -73.9% Again, those improvements aren't realistic for the 1-pack case (because in the real world, the full-array solution doesn't kick in), but it's more useful to be testing the more-complicated code path. While we're looking at this issue, we'll tweak one more thing: in oe_map_new_pack(), we call REALLOC_ARRAY(pack->in_pack). But we'd never expect to get here unless we're back-filling it for the first time, in which case it would be NULL. So let's switch that to ALLOC_ARRAY() for clarity, and add a BUG() to document the expectation. Unfortunately this code isn't well-covered in the test suite because it's inherently racy (it only kicks in if somebody else adds a new pack while we're in the middle of repacking). Signed-off-by: Jeff King <peff@peff.net> Reviewed-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
5 years ago
}
pack->in_pack[e - pack->objects] = p;
}
static inline struct object_entry *oe_delta(
const struct packing_data *pack,
const struct object_entry *e)
{
pack-objects: reuse on-disk deltas for thin "have" objects When we serve a fetch, we pass the "wants" and "haves" from the fetch negotiation to pack-objects. That tells us not only which objects we need to send, but we also use the boundary commits as "preferred bases": their trees and blobs are candidates for delta bases, both for reusing on-disk deltas and for finding new ones. However, this misses some opportunities. Modulo some special cases like shallow or partial clones, we know that every object reachable from the "haves" could be a preferred base. We don't use all of them for two reasons: 1. It's expensive to traverse the whole history and enumerate all of the objects the other side has. 2. The delta search is expensive, so we want to keep the number of candidate bases sane. The boundary commits are the most likely to work. When we have reachability bitmaps, though, reason 1 no longer applies. We can efficiently compute the set of reachable objects on the other side (and in fact already did so as part of the bitmap set-difference to get the list of interesting objects). And using this set conveniently covers the shallow and partial cases, since we have to disable the use of bitmaps for those anyway. The second reason argues against using these bases in the search for new deltas. But there's one case where we can use this information for free: when we have an existing on-disk delta that we're considering reusing, we can do so if we know the other side has the base object. This in fact saves time during the delta search, because it's one less delta we have to compute. And that's exactly what this patch does: when we're considering whether to reuse an on-disk delta, if bitmaps tell us the other side has the object (and we're making a thin-pack), then we reuse it. Here are the results on p5311 using linux.git, which simulates a client fetching after `N` days since their last fetch: Test origin HEAD -------------------------------------------------------------------------- 5311.3: server (1 days) 0.27(0.27+0.04) 0.12(0.09+0.03) -55.6% 5311.4: size (1 days) 0.9M 237.0K -73.7% 5311.5: client (1 days) 0.04(0.05+0.00) 0.10(0.10+0.00) +150.0% 5311.7: server (2 days) 0.34(0.42+0.04) 0.13(0.10+0.03) -61.8% 5311.8: size (2 days) 1.5M 347.7K -76.5% 5311.9: client (2 days) 0.07(0.08+0.00) 0.16(0.15+0.01) +128.6% 5311.11: server (4 days) 0.56(0.77+0.08) 0.13(0.10+0.02) -76.8% 5311.12: size (4 days) 2.8M 566.6K -79.8% 5311.13: client (4 days) 0.13(0.15+0.00) 0.34(0.31+0.02) +161.5% 5311.15: server (8 days) 0.97(1.39+0.11) 0.30(0.25+0.05) -69.1% 5311.16: size (8 days) 4.3M 1.0M -76.0% 5311.17: client (8 days) 0.20(0.22+0.01) 0.53(0.52+0.01) +165.0% 5311.19: server (16 days) 1.52(2.51+0.12) 0.30(0.26+0.03) -80.3% 5311.20: size (16 days) 8.0M 2.0M -74.5% 5311.21: client (16 days) 0.40(0.47+0.03) 1.01(0.98+0.04) +152.5% 5311.23: server (32 days) 2.40(4.44+0.20) 0.31(0.26+0.04) -87.1% 5311.24: size (32 days) 14.1M 4.1M -70.9% 5311.25: client (32 days) 0.70(0.90+0.03) 1.81(1.75+0.06) +158.6% 5311.27: server (64 days) 11.76(26.57+0.29) 0.55(0.50+0.08) -95.3% 5311.28: size (64 days) 89.4M 47.4M -47.0% 5311.29: client (64 days) 5.71(9.31+0.27) 15.20(15.20+0.32) +166.2% 5311.31: server (128 days) 16.15(36.87+0.40) 0.91(0.82+0.14) -94.4% 5311.32: size (128 days) 134.8M 100.4M -25.5% 5311.33: client (128 days) 9.42(16.86+0.49) 25.34(25.80+0.46) +169.0% In all cases we save CPU time on the server (sometimes significant) and the resulting pack is smaller. We do spend more CPU time on the client side, because it has to reconstruct more deltas. But that's the right tradeoff to make, since clients tend to outnumber servers. It just means the thin pack mechanism is doing its job. From the user's perspective, the end-to-end time of the operation will generally be faster. E.g., in the 128-day case, we saved 15s on the server at a cost of 16s on the client. Since the resulting pack is 34MB smaller, this is a net win if the network speed is less than 270Mbit/s. And that's actually the worst case. The 64-day case saves just over 11s at a cost of just under 11s. So it's a slight win at any network speed, and the 40MB saved is pure bonus. That trend continues for the smaller fetches. The implementation itself is mostly straightforward, with the new logic going into check_object(). But there are two tricky bits. The first is that check_object() needs access to the relevant information (the thin flag and bitmap result). We can do this by pushing these into program-lifetime globals. The second is that the rest of the code assumes that any reused delta will point to another "struct object_entry" as its base. But of course the case we are interested in here is the one where don't have such an entry! I looked at a number of options that didn't quite work: - we could use a flag to signal a reused delta, but it's not a single bit. We have to actually store the oid of the base, which is normally done by pointing to the existing object_entry. And we'd have to modify all the code which looks at deltas. - we could add the reused bases to the end of the existing object_entry array. While this does create some extra work as later stages consider the extra entries, it's actually not too bad (we're not sending them, so they don't cost much in the delta search, and at most we'd have 2*N of them). But there's a more subtle problem. Adding to the existing array means we might need to grow it with realloc, which could move the earlier entries around. While many of the references to other entries are done by integer index, some (including ones on the stack) use pointers, which would become invalidated. This isn't insurmountable, but it would require quite a bit of refactoring (and it's hard to know that you've got it all, since it may work _most_ of the time and then fail subtly based on memory allocation patterns). - we could allocate a new one-off entry for the base. In fact, this is what an earlier version of this patch did. However, since the refactoring brought in by ad635e82d6 (Merge branch 'nd/pack-objects-pack-struct', 2018-05-23), the delta_idx code requires that both entries be in the main packing list. So taking all of those options into account, what I ended up with is a separate list of "external bases" that are not part of the main packing list. Each delta entry that points to an external base has a single-bit flag to do so; we have a little breathing room in the bitfield section of object_entry. This lets us limit the change primarily to the oe_delta() and oe_set_delta_ext() functions. And as a bonus, most of the rest of the code does not consider these dummy entries at all, saving both runtime CPU and code complexity. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years ago
if (!e->delta_idx)
return NULL;
if (e->ext_base)
return &pack->ext_bases[e->delta_idx - 1];
else
return &pack->objects[e->delta_idx - 1];
}
static inline void oe_set_delta(struct packing_data *pack,
struct object_entry *e,
struct object_entry *delta)
{
if (delta)
e->delta_idx = (delta - pack->objects) + 1;
else
e->delta_idx = 0;
}
pack-objects: reuse on-disk deltas for thin "have" objects When we serve a fetch, we pass the "wants" and "haves" from the fetch negotiation to pack-objects. That tells us not only which objects we need to send, but we also use the boundary commits as "preferred bases": their trees and blobs are candidates for delta bases, both for reusing on-disk deltas and for finding new ones. However, this misses some opportunities. Modulo some special cases like shallow or partial clones, we know that every object reachable from the "haves" could be a preferred base. We don't use all of them for two reasons: 1. It's expensive to traverse the whole history and enumerate all of the objects the other side has. 2. The delta search is expensive, so we want to keep the number of candidate bases sane. The boundary commits are the most likely to work. When we have reachability bitmaps, though, reason 1 no longer applies. We can efficiently compute the set of reachable objects on the other side (and in fact already did so as part of the bitmap set-difference to get the list of interesting objects). And using this set conveniently covers the shallow and partial cases, since we have to disable the use of bitmaps for those anyway. The second reason argues against using these bases in the search for new deltas. But there's one case where we can use this information for free: when we have an existing on-disk delta that we're considering reusing, we can do so if we know the other side has the base object. This in fact saves time during the delta search, because it's one less delta we have to compute. And that's exactly what this patch does: when we're considering whether to reuse an on-disk delta, if bitmaps tell us the other side has the object (and we're making a thin-pack), then we reuse it. Here are the results on p5311 using linux.git, which simulates a client fetching after `N` days since their last fetch: Test origin HEAD -------------------------------------------------------------------------- 5311.3: server (1 days) 0.27(0.27+0.04) 0.12(0.09+0.03) -55.6% 5311.4: size (1 days) 0.9M 237.0K -73.7% 5311.5: client (1 days) 0.04(0.05+0.00) 0.10(0.10+0.00) +150.0% 5311.7: server (2 days) 0.34(0.42+0.04) 0.13(0.10+0.03) -61.8% 5311.8: size (2 days) 1.5M 347.7K -76.5% 5311.9: client (2 days) 0.07(0.08+0.00) 0.16(0.15+0.01) +128.6% 5311.11: server (4 days) 0.56(0.77+0.08) 0.13(0.10+0.02) -76.8% 5311.12: size (4 days) 2.8M 566.6K -79.8% 5311.13: client (4 days) 0.13(0.15+0.00) 0.34(0.31+0.02) +161.5% 5311.15: server (8 days) 0.97(1.39+0.11) 0.30(0.25+0.05) -69.1% 5311.16: size (8 days) 4.3M 1.0M -76.0% 5311.17: client (8 days) 0.20(0.22+0.01) 0.53(0.52+0.01) +165.0% 5311.19: server (16 days) 1.52(2.51+0.12) 0.30(0.26+0.03) -80.3% 5311.20: size (16 days) 8.0M 2.0M -74.5% 5311.21: client (16 days) 0.40(0.47+0.03) 1.01(0.98+0.04) +152.5% 5311.23: server (32 days) 2.40(4.44+0.20) 0.31(0.26+0.04) -87.1% 5311.24: size (32 days) 14.1M 4.1M -70.9% 5311.25: client (32 days) 0.70(0.90+0.03) 1.81(1.75+0.06) +158.6% 5311.27: server (64 days) 11.76(26.57+0.29) 0.55(0.50+0.08) -95.3% 5311.28: size (64 days) 89.4M 47.4M -47.0% 5311.29: client (64 days) 5.71(9.31+0.27) 15.20(15.20+0.32) +166.2% 5311.31: server (128 days) 16.15(36.87+0.40) 0.91(0.82+0.14) -94.4% 5311.32: size (128 days) 134.8M 100.4M -25.5% 5311.33: client (128 days) 9.42(16.86+0.49) 25.34(25.80+0.46) +169.0% In all cases we save CPU time on the server (sometimes significant) and the resulting pack is smaller. We do spend more CPU time on the client side, because it has to reconstruct more deltas. But that's the right tradeoff to make, since clients tend to outnumber servers. It just means the thin pack mechanism is doing its job. From the user's perspective, the end-to-end time of the operation will generally be faster. E.g., in the 128-day case, we saved 15s on the server at a cost of 16s on the client. Since the resulting pack is 34MB smaller, this is a net win if the network speed is less than 270Mbit/s. And that's actually the worst case. The 64-day case saves just over 11s at a cost of just under 11s. So it's a slight win at any network speed, and the 40MB saved is pure bonus. That trend continues for the smaller fetches. The implementation itself is mostly straightforward, with the new logic going into check_object(). But there are two tricky bits. The first is that check_object() needs access to the relevant information (the thin flag and bitmap result). We can do this by pushing these into program-lifetime globals. The second is that the rest of the code assumes that any reused delta will point to another "struct object_entry" as its base. But of course the case we are interested in here is the one where don't have such an entry! I looked at a number of options that didn't quite work: - we could use a flag to signal a reused delta, but it's not a single bit. We have to actually store the oid of the base, which is normally done by pointing to the existing object_entry. And we'd have to modify all the code which looks at deltas. - we could add the reused bases to the end of the existing object_entry array. While this does create some extra work as later stages consider the extra entries, it's actually not too bad (we're not sending them, so they don't cost much in the delta search, and at most we'd have 2*N of them). But there's a more subtle problem. Adding to the existing array means we might need to grow it with realloc, which could move the earlier entries around. While many of the references to other entries are done by integer index, some (including ones on the stack) use pointers, which would become invalidated. This isn't insurmountable, but it would require quite a bit of refactoring (and it's hard to know that you've got it all, since it may work _most_ of the time and then fail subtly based on memory allocation patterns). - we could allocate a new one-off entry for the base. In fact, this is what an earlier version of this patch did. However, since the refactoring brought in by ad635e82d6 (Merge branch 'nd/pack-objects-pack-struct', 2018-05-23), the delta_idx code requires that both entries be in the main packing list. So taking all of those options into account, what I ended up with is a separate list of "external bases" that are not part of the main packing list. Each delta entry that points to an external base has a single-bit flag to do so; we have a little breathing room in the bitfield section of object_entry. This lets us limit the change primarily to the oe_delta() and oe_set_delta_ext() functions. And as a bonus, most of the rest of the code does not consider these dummy entries at all, saving both runtime CPU and code complexity. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years ago
void oe_set_delta_ext(struct packing_data *pack,
struct object_entry *e,
const struct object_id *oid);
pack-objects: reuse on-disk deltas for thin "have" objects When we serve a fetch, we pass the "wants" and "haves" from the fetch negotiation to pack-objects. That tells us not only which objects we need to send, but we also use the boundary commits as "preferred bases": their trees and blobs are candidates for delta bases, both for reusing on-disk deltas and for finding new ones. However, this misses some opportunities. Modulo some special cases like shallow or partial clones, we know that every object reachable from the "haves" could be a preferred base. We don't use all of them for two reasons: 1. It's expensive to traverse the whole history and enumerate all of the objects the other side has. 2. The delta search is expensive, so we want to keep the number of candidate bases sane. The boundary commits are the most likely to work. When we have reachability bitmaps, though, reason 1 no longer applies. We can efficiently compute the set of reachable objects on the other side (and in fact already did so as part of the bitmap set-difference to get the list of interesting objects). And using this set conveniently covers the shallow and partial cases, since we have to disable the use of bitmaps for those anyway. The second reason argues against using these bases in the search for new deltas. But there's one case where we can use this information for free: when we have an existing on-disk delta that we're considering reusing, we can do so if we know the other side has the base object. This in fact saves time during the delta search, because it's one less delta we have to compute. And that's exactly what this patch does: when we're considering whether to reuse an on-disk delta, if bitmaps tell us the other side has the object (and we're making a thin-pack), then we reuse it. Here are the results on p5311 using linux.git, which simulates a client fetching after `N` days since their last fetch: Test origin HEAD -------------------------------------------------------------------------- 5311.3: server (1 days) 0.27(0.27+0.04) 0.12(0.09+0.03) -55.6% 5311.4: size (1 days) 0.9M 237.0K -73.7% 5311.5: client (1 days) 0.04(0.05+0.00) 0.10(0.10+0.00) +150.0% 5311.7: server (2 days) 0.34(0.42+0.04) 0.13(0.10+0.03) -61.8% 5311.8: size (2 days) 1.5M 347.7K -76.5% 5311.9: client (2 days) 0.07(0.08+0.00) 0.16(0.15+0.01) +128.6% 5311.11: server (4 days) 0.56(0.77+0.08) 0.13(0.10+0.02) -76.8% 5311.12: size (4 days) 2.8M 566.6K -79.8% 5311.13: client (4 days) 0.13(0.15+0.00) 0.34(0.31+0.02) +161.5% 5311.15: server (8 days) 0.97(1.39+0.11) 0.30(0.25+0.05) -69.1% 5311.16: size (8 days) 4.3M 1.0M -76.0% 5311.17: client (8 days) 0.20(0.22+0.01) 0.53(0.52+0.01) +165.0% 5311.19: server (16 days) 1.52(2.51+0.12) 0.30(0.26+0.03) -80.3% 5311.20: size (16 days) 8.0M 2.0M -74.5% 5311.21: client (16 days) 0.40(0.47+0.03) 1.01(0.98+0.04) +152.5% 5311.23: server (32 days) 2.40(4.44+0.20) 0.31(0.26+0.04) -87.1% 5311.24: size (32 days) 14.1M 4.1M -70.9% 5311.25: client (32 days) 0.70(0.90+0.03) 1.81(1.75+0.06) +158.6% 5311.27: server (64 days) 11.76(26.57+0.29) 0.55(0.50+0.08) -95.3% 5311.28: size (64 days) 89.4M 47.4M -47.0% 5311.29: client (64 days) 5.71(9.31+0.27) 15.20(15.20+0.32) +166.2% 5311.31: server (128 days) 16.15(36.87+0.40) 0.91(0.82+0.14) -94.4% 5311.32: size (128 days) 134.8M 100.4M -25.5% 5311.33: client (128 days) 9.42(16.86+0.49) 25.34(25.80+0.46) +169.0% In all cases we save CPU time on the server (sometimes significant) and the resulting pack is smaller. We do spend more CPU time on the client side, because it has to reconstruct more deltas. But that's the right tradeoff to make, since clients tend to outnumber servers. It just means the thin pack mechanism is doing its job. From the user's perspective, the end-to-end time of the operation will generally be faster. E.g., in the 128-day case, we saved 15s on the server at a cost of 16s on the client. Since the resulting pack is 34MB smaller, this is a net win if the network speed is less than 270Mbit/s. And that's actually the worst case. The 64-day case saves just over 11s at a cost of just under 11s. So it's a slight win at any network speed, and the 40MB saved is pure bonus. That trend continues for the smaller fetches. The implementation itself is mostly straightforward, with the new logic going into check_object(). But there are two tricky bits. The first is that check_object() needs access to the relevant information (the thin flag and bitmap result). We can do this by pushing these into program-lifetime globals. The second is that the rest of the code assumes that any reused delta will point to another "struct object_entry" as its base. But of course the case we are interested in here is the one where don't have such an entry! I looked at a number of options that didn't quite work: - we could use a flag to signal a reused delta, but it's not a single bit. We have to actually store the oid of the base, which is normally done by pointing to the existing object_entry. And we'd have to modify all the code which looks at deltas. - we could add the reused bases to the end of the existing object_entry array. While this does create some extra work as later stages consider the extra entries, it's actually not too bad (we're not sending them, so they don't cost much in the delta search, and at most we'd have 2*N of them). But there's a more subtle problem. Adding to the existing array means we might need to grow it with realloc, which could move the earlier entries around. While many of the references to other entries are done by integer index, some (including ones on the stack) use pointers, which would become invalidated. This isn't insurmountable, but it would require quite a bit of refactoring (and it's hard to know that you've got it all, since it may work _most_ of the time and then fail subtly based on memory allocation patterns). - we could allocate a new one-off entry for the base. In fact, this is what an earlier version of this patch did. However, since the refactoring brought in by ad635e82d6 (Merge branch 'nd/pack-objects-pack-struct', 2018-05-23), the delta_idx code requires that both entries be in the main packing list. So taking all of those options into account, what I ended up with is a separate list of "external bases" that are not part of the main packing list. Each delta entry that points to an external base has a single-bit flag to do so; we have a little breathing room in the bitfield section of object_entry. This lets us limit the change primarily to the oe_delta() and oe_set_delta_ext() functions. And as a bonus, most of the rest of the code does not consider these dummy entries at all, saving both runtime CPU and code complexity. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years ago
static inline struct object_entry *oe_delta_child(
const struct packing_data *pack,
const struct object_entry *e)
{
if (e->delta_child_idx)
return &pack->objects[e->delta_child_idx - 1];
return NULL;
}
static inline void oe_set_delta_child(struct packing_data *pack,
struct object_entry *e,
struct object_entry *delta)
{
if (delta)
e->delta_child_idx = (delta - pack->objects) + 1;
else
e->delta_child_idx = 0;
}
static inline struct object_entry *oe_delta_sibling(
const struct packing_data *pack,
const struct object_entry *e)
{
if (e->delta_sibling_idx)
return &pack->objects[e->delta_sibling_idx - 1];
return NULL;
}
static inline void oe_set_delta_sibling(struct packing_data *pack,
struct object_entry *e,
struct object_entry *delta)
{
if (delta)
e->delta_sibling_idx = (delta - pack->objects) + 1;
else
e->delta_sibling_idx = 0;
}
unsigned long oe_get_size_slow(struct packing_data *pack,
const struct object_entry *e);
static inline unsigned long oe_size(struct packing_data *pack,
const struct object_entry *e)
{
if (e->size_valid)
return e->size_;
return oe_get_size_slow(pack, e);
}
static inline int oe_size_less_than(struct packing_data *pack,
const struct object_entry *lhs,
unsigned long rhs)
{
if (lhs->size_valid)
return lhs->size_ < rhs;
if (rhs < pack->oe_size_limit) /* rhs < 2^x <= lhs ? */
return 0;
return oe_get_size_slow(pack, lhs) < rhs;
}
static inline int oe_size_greater_than(struct packing_data *pack,
const struct object_entry *lhs,
unsigned long rhs)
{
if (lhs->size_valid)
return lhs->size_ > rhs;
if (rhs < pack->oe_size_limit) /* rhs < 2^x <= lhs ? */
return 1;
return oe_get_size_slow(pack, lhs) > rhs;
}
static inline void oe_set_size(struct packing_data *pack,
struct object_entry *e,
unsigned long size)
{
if (size < pack->oe_size_limit) {
e->size_ = size;
e->size_valid = 1;
} else {
e->size_valid = 0;
if (oe_get_size_slow(pack, e) != size)
BUG("'size' is supposed to be the object size!");
}
}
static inline unsigned long oe_delta_size(struct packing_data *pack,
const struct object_entry *e)
{
if (e->delta_size_valid)
return e->delta_size_;
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
/*
* pack->delta_size[] can't be NULL because oe_set_delta_size()
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
* must have been called when a new delta is saved with
* oe_set_delta().
* If oe_delta() returns NULL (i.e. default state, which means
* delta_size_valid is also false), then the caller must never
* call oe_delta_size().
*/
return pack->delta_size[e - pack->objects];
}
static inline void oe_set_delta_size(struct packing_data *pack,
struct object_entry *e,
unsigned long size)
{
pack-objects: fix performance issues on packing large deltas Let's start with some background about oe_delta_size() and oe_set_delta_size(). If you already know, skip the next paragraph. These two are added in 0aca34e826 (pack-objects: shrink delta_size field in struct object_entry - 2018-04-14) to help reduce 'struct object_entry' size. The delta size field in this struct is reduced to only contain max 1MB. So if any new delta is produced and larger than 1MB, it's dropped because we can't really save such a large size anywhere. Fallback is provided in case existing packfiles already have large deltas, then we can retrieve it from the pack. While this should help small machines repacking large repos without large deltas (i.e. less memory pressure), dropping large deltas during the delta selection process could end up with worse pack files. And if existing packfiles already have >1MB delta and pack-objects is instructed to not reuse deltas, all of them will be dropped on the floor, and the resulting pack would be definitely bigger. There is also a regression in terms of CPU/IO if we have large on-disk deltas because fallback code needs to parse the pack every time the delta size is needed and just access to the mmap'd pack data is enough for extra page faults when memory is under pressure. Both of these issues were reported on the mailing list. Here's some numbers for comparison. Version Pack (MB) MaxRSS(kB) Time (s) ------- --------- ---------- -------- 2.17.0 5498 43513628 2494.85 2.18.0 10531 40449596 4168.94 This patch provides a better fallback that is - cheaper in terms of cpu and io because we won't have to read existing pack files as much - better in terms of pack size because the pack heuristics is back to 2.17.0 time, we do not drop large deltas at all If we encounter any delta (on-disk or created during try_delta phase) that is larger than the 1MB limit, we stop using delta_size_ field for this because it can't contain such size anyway. A new array of delta size is dynamically allocated and can hold all the deltas that 2.17.0 can. This array only contains delta sizes that delta_size_ can't contain. With this, we do not have to drop deltas in try_delta() anymore. Of course the downside is we use slightly more memory, even compared to 2.17.0. But since this is considered an uncommon case, a bit more memory consumption should not be a problem. Delta size limit is also raised from 1MB to 16MB to better cover common case and avoid that extra memory consumption (99.999% deltas in this reported repo are under 12MB; Jeff noted binary artifacts topped out at about 3MB in some other private repos). Other fields are shuffled around to keep this struct packed tight. We don't use more memory in common case even with this limit update. A note about thread synchronization. Since this code can be run in parallel during delta searching phase, we need a mutex. The realloc part in packlist_alloc() is not protected because it only happens during the object counting phase, which is always single-threaded. Access to e->delta_size_ (and by extension pack->delta_size[e - pack->objects]) is unprotected as before, the thread scheduler in pack-objects must make sure "e" is never updated by two different threads. The area under the new lock is as small as possible, avoiding locking at all in common case, since lock contention with high thread count could be expensive (most blobs are small enough that delta compute time is short and we end up taking the lock very often). The previous attempt to always hold a lock in oe_delta_size() and oe_set_delta_size() increases execution time by 33% when repacking linux.git with with 40 threads. Reported-by: Elijah Newren <newren@gmail.com> Helped-by: Elijah Newren <newren@gmail.com> Helped-by: Jeff King <peff@peff.net> Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
7 years ago
if (size < pack->oe_delta_size_limit) {
e->delta_size_ = size;
e->delta_size_valid = 1;
} else {
packing_data_lock(pack);
if (!pack->delta_size)
ALLOC_ARRAY(pack->delta_size, pack->nr_alloc);
packing_data_unlock(pack);
pack->delta_size[e - pack->objects] = size;
e->delta_size_valid = 0;
}
}
static inline unsigned int oe_tree_depth(struct packing_data *pack,
struct object_entry *e)
{
if (!pack->tree_depth)
return 0;
return pack->tree_depth[e - pack->objects];
}
static inline void oe_set_tree_depth(struct packing_data *pack,
struct object_entry *e,
unsigned int tree_depth)
{
if (!pack->tree_depth)
pack-objects: zero-initialize tree_depth/layer arrays Commit 108f530385 (pack-objects: move tree_depth into 'struct packing_data', 2018-08-16) started maintaining a tree_depth array that matches the "objects" array. We extend the array when: 1. The objects array is extended, in which case we use realloc to extend the tree_depth array. 2. A caller asks to store a tree_depth for object N, and this is the first such request; we create the array from scratch and store the value for N. In the latter case, though, we use regular xmalloc(), and the depth values for any objects besides N is undefined. This happens to not trigger a bug with the current code, but the reasons are quite subtle: - we never ask about the depth for any object with index i < N. This is because we store the depth immediately for all trees and blobs. So any such "i" must be a non-tree, and therefore we will never need to care about its depth (in fact, we really only care about the depth of trees). - there are no objects at this point with index i > N, because we always fill in the depth for a tree immediately after its object entry is created (we may still allocate uninitialized depth entries, but they'll be initialized by packlist_alloc() when it initializes the entry in the "objects" array). So it works, but only by chance. To be defensive, let's zero the array, which matches the "unset" values which would be handed out by oe_tree_depth() already. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years ago
CALLOC_ARRAY(pack->tree_depth, pack->nr_alloc);
pack->tree_depth[e - pack->objects] = tree_depth;
}
static inline unsigned char oe_layer(struct packing_data *pack,
struct object_entry *e)
{
if (!pack->layer)
return 0;
return pack->layer[e - pack->objects];
}
static inline void oe_set_layer(struct packing_data *pack,
struct object_entry *e,
unsigned char layer)
{
if (!pack->layer)
pack-objects: zero-initialize tree_depth/layer arrays Commit 108f530385 (pack-objects: move tree_depth into 'struct packing_data', 2018-08-16) started maintaining a tree_depth array that matches the "objects" array. We extend the array when: 1. The objects array is extended, in which case we use realloc to extend the tree_depth array. 2. A caller asks to store a tree_depth for object N, and this is the first such request; we create the array from scratch and store the value for N. In the latter case, though, we use regular xmalloc(), and the depth values for any objects besides N is undefined. This happens to not trigger a bug with the current code, but the reasons are quite subtle: - we never ask about the depth for any object with index i < N. This is because we store the depth immediately for all trees and blobs. So any such "i" must be a non-tree, and therefore we will never need to care about its depth (in fact, we really only care about the depth of trees). - there are no objects at this point with index i > N, because we always fill in the depth for a tree immediately after its object entry is created (we may still allocate uninitialized depth entries, but they'll be initialized by packlist_alloc() when it initializes the entry in the "objects" array). So it works, but only by chance. To be defensive, let's zero the array, which matches the "unset" values which would be handed out by oe_tree_depth() already. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years ago
CALLOC_ARRAY(pack->layer, pack->nr_alloc);
pack->layer[e - pack->objects] = layer;
}
#endif