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.

2640 lines
111 KiB

diff -rup numad-0.5git/numad.8 numad-0.5git-new/numad.8
--- numad-0.5git/numad.8 2012-12-03 15:40:40.000000000 +0100
+++ numad-0.5git-new/numad.8 2016-08-30 08:45:19.000000000 +0200
@@ -1,45 +1,56 @@
.TH "numad" "8" "1.0.0" "Bill Gray" "Administration"
-.SH "numad"
-.LP
+.SH "NAME"
+.LP
numad \- A user\-level daemon that provides placement advice and process
management for efficient use of CPUs and memory on systems with NUMA topology.
-.SH "SYNTAX"
-.LP
+.SH "SYNOPSIS"
+.LP
numad [\fI\-dhvV\fP]
-.br
-.LP
-numad [\fI\-D non-standard-cgroup-mount-point\fP]
-.br
-.LP
+.br
+.LP
+numad [\fI\-C 0|1\fP]
+.br
+.LP
+numad [\fI\-H THP_hugepage_scan_sleep_ms\fP]
+.br
+.LP
numad [\fI\-i [min_interval:]max_interval\fP]
-.br
-.LP
+.br
+.LP
numad [\fI\-K 0|1\fP]
-.br
-.LP
+.br
+.LP
numad [\fI\-l log_level\fP]
-.br
-.LP
+.br
+.LP
+numad [\fI\-m target_memory_locality\fP]
+.br
+.LP
numad [\fI\-p PID\fP]
-.br
-.LP
+.br
+.LP
numad [\fI\-r PID\fP]
-.br
-.LP
+.br
+.LP
+numad [\fI\-R reserved-CPU-list\fP]
+.br
+.LP
numad [\fI\-S 0|1\fP]
-.br
-.LP
+.br
+.LP
+numad [\fI\-t logical_CPU_percent\fP]
+.br
+.LP
numad [\fI\-u target_utilization\fP]
-.br
-.LP
+.br
+.LP
numad [\fI\-w NCPUS[:MB]\fP]
-.br
-.LP
+.br
+.LP
numad [\fI\-x PID\fP]
-.br
-
+.br
.SH "DESCRIPTION"
-.LP
+.LP
Numad is a system daemon that monitors NUMA topology and resource usage. It
will attempt to locate processes for efficient NUMA locality and affinity,
dynamically adjusting to changing system conditions. Numad also provides
@@ -53,25 +64,42 @@ large in-memory database application, fo
accesses will likely remain unpredictable -- numad will probably not improve
performance.
.SH "OPTIONS"
-.LP
-.TP
+.LP
+.TP
+\fB\-C\fR <\fI0|1\fP>
+This option controls whether or not numad treats inactive file cache as
+available memory. By default, numad assumes it can count inactive file cache as
+"free" memory when considering resources to match with processes. Specify
+\fI\-C 0\fP if numad should instead consider inactive file cache as a consumed
+resource.
+.TP
\fB\-d\fR
Debug output in log, sets the log level to LOG_DEBUG. Same effect as \fI\-l 7\fP.
.TP
-\fB\-D\fR <\fInon-standard-cgroup-mount-point\fP>
-This option can be used to communicate a non-standard cgroup mount point to
-numad. This is not normally necessary.
-.TP
\fB\-h\fR
Display usage help information and then exit.
-.TP
+.TP
+\fB\-H\fR <\fITHP_scan_sleep_ms\fP>
+Set the desired transparent hugepage scan interval in ms. The
+.na
+/sys/kernel/mm/tranparent_hugepage/khugepaged/scan_sleep_millisecs
+.ad
+tunable is usually set to 10000ms by the operating system. The default is
+changed by numad to be 1000ms since it is helpful for the hugepage daemon to be
+more aggressive when memory moves between nodes. Specifying (\fI\-H 0\fP) will
+cause numad to retain the system default value. You can also make the hugepage
+daemon more or less aggressive by specifying an alternate value with this
+option. For example, setting this value to 100ms (\fI\-H 100\fP) might improve
+the performance of workloads which use many transparent hugepages.
+.TP
\fB\-i\fR <\fI[min_interval:]max_interval\fP>
Sets the time interval that numad waits between system scans, in seconds to
<\fImax_interval\fP>. Default <\fImax_interval\fP> is 15 seconds, default
<\fImin_interval\fP> is 5 seconds. Setting a <\fImax_interval\fP> of zero will
cause the daemon to exit. (This is the normal mechanism to terminate the
daemon.) A bigger <\fImax_interval\fP> will decrease numad overhead but also
-decrease responsiveness to changing loads.
+decrease responsiveness to changing loads. The default numad max_interval can
+be changed in the numad.conf file.
.TP
\fB\-K\fR <\fI0|1\fP>
This option controls whether numad keeps interleaved memory spread across NUMA
@@ -82,10 +110,24 @@ a large, single-instance application tha
the workload will have continuous unpredictable memory access patterns (e.g. a
large in-memory database), you might get better results by specifying \fI\-K
1\fP to instruct numad to keep interleaved memory distributed.
-.TP
+.TP
\fB\-l\fR <\fIlog_level\fP>
Sets the log level to <\fIlog_level\fP>. Reasonable choices are 5, 6, or 7.
-The default value is 5.
+The default value is 5. Note that CPU values are scaled by a factor of 100
+internally and in the numad log files. Unfortunately, you don't actually have
+that many CPUs.
+.TP
+\fB\-m\fR <\fItarget_memory_locality\fP>
+Set the desired memory locality threshold to stop moving process memory. Numad
+might stop retrying to coalesce process memory when more than this percentage
+of the process's memory is already localized in the target node(s). The
+default is 90%. Numad will frequently localize more than the localization
+threshold percent, but it will not necessarily do so. Decrease the threshold
+to allow numad to leave more process memory distributed on various nodes.
+Increase the threshold to instruct numad to try to localize more memory.
+Acceptable values are between 50 and 100 percent. Note that setting the target
+memory locality to 100% might cause numad to continually retry to move memory
+that the kernel will never succesfully move.
.TP
\fB\-p\fR <\fIPID\fP>
Add PID to explicit inclusion list of processes to consider for managing, if
@@ -102,6 +144,12 @@ processes. After daemon start, only one
process lists per subsequent numad invocation. Use with \-S and \-p and \-x to
precisely control the scope of processes numad can manage.
.TP
+\fB\-R\fR <\fICPU_LIST\fP>
+Specify a list of CPUs that numad should assume are reserved for non-numad use.
+No processes will be bound to the specified CPUs by numad. This option is
+effective only when starting numad. You cannot change reserved CPUs
+dynamically while numad is already running.
+.TP
\fB\-S\fR <\fI0|1\fP>
This option controls whether numad scans all system processes or only the
processes on the explicit inclusion PID list. The default is to scan all
@@ -113,18 +161,30 @@ exclusion list). Starting numad as
.br
will limit scanning, and thus also automatic NUMA management, to only those
three explicitly specified processes.
-.TP
+.TP
+\fB\-t\fR <\fIlogical_CPU_percent\fP>
+Specify the resource value of logical CPUs. Hardware threads typically share
+most core resources, and so logical CPUs add only a fraction of CPU power for
+many workloads. By default numad considers logical CPUs to be only 20 percent
+of a dedicated hardware core.
+.TP
\fB\-u\fR <\fItarget_utilization\fP>
Set the desired maximum consumption percentage of a node. Default is 85%.
Decrease the target value to maintain more available resource margin on each
node. Increase the target value to more exhaustively consume node resources.
-.TP
+If you have sized your workloads to precisely fit inside a NUMA node,
+specifying (\fI\-u 100\fP) might improve system performance by telling numad to
+go ahead and consume all the resources in each node. It is possible to specify
+values up to 130 percent to oversubscribe CPUs in the nodes, but memory
+utilization is always capped at 100%. Use oversubscription values very
+carefully.
+.TP
\fB\-v\fR
Verbose output in log, sets the log level to LOG_INFO. Same effect as \fI\-l 6\fP.
-.TP
+.TP
\fB\-V\fR
Display version information and exit.
-.TP
+.TP
\fB\-w\fR <\fINCPUS[:MB]\fP>
Queries numad for the best NUMA nodes to bind an entity that needs
<\fINCPUS\fP>. The amount of memory (in MBs) is optional, but should normally
@@ -145,32 +205,37 @@ Add PID to explicit exclusion list of pr
Multiple \fI\-x PID\fP options can be specified at daemon start, but after
daemon start, only one PID can be added to the exclusion list per subsequent
numad invocation. Use with \-S to precisely control the scope of processes
-numad can manage.
+numad can manage.
.SH "FILES"
-.LP
-\fI/usr/bin/numad\fP
-.br
-\fI/var/log/numad.log\fP
-.br
-\fI/var/run/numad.pid\fP
+.LP
+\fI/usr/bin/numad\fP
+.br
+\fI/etc/numad.conf\fP
+.br
+\fI/var/log/numad.log\fP
+.br
+\fI/var/run/numad.pid\fP
.SH "ENVIRONMENT VARIABLES"
-.LP
-.TP
+.LP
+.TP
None.
.SH "EXAMPLES"
-.LP
-Numad is normally run as a system daemon and should be managed by the
+.LP
+Numad can be run as a system daemon and can be managed by the
standard init mechanisms of the host.
-.LP
+.LP
If interactive (manual) control is desired, you can start the daemon manually by typing:
-.LP
+.LP
/usr/bin/numad
.LP
-Subsequent numad invocations while the daemon is running can be used to dynamically change run-time options.
+Subsequent numad invocations while the daemon is running can be used to dynamically change most run-time options.
+.LP
+You can terminate numad from running by typing:
+.LP
+/usr/bin/numad -i0
.SH "AUTHORS"
-.LP
+.LP
Bill Gray <bgray@redhat.com>
.SH "SEE ALSO"
-.LP
+.LP
numactl(8)
-
diff -rup numad-0.5git/numad.c numad-0.5git-new/numad.c
--- numad-0.5git/numad.c 2012-12-03 15:40:40.000000000 +0100
+++ numad-0.5git-new/numad.c 2016-08-30 08:45:19.000000000 +0200
@@ -19,7 +19,7 @@ Inc., 59 Temple Place, Suite 330, Boston
*/
-// Compile with: gcc -O -std=gnu99 -Wall -pthread -o numad numad.c -lrt
+// Compile with: gcc -std=gnu99 -g -Wall -pthread -o numad numad.c -lrt -lm
#define _GNU_SOURCE
@@ -40,6 +40,10 @@ Inc., 59 Temple Place, Suite 330, Boston
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
+#include <time.h>
+#include <unistd.h>
+#include <values.h>
+
#include <sys/ipc.h>
#include <sys/mman.h>
#include <sys/msg.h>
@@ -49,26 +53,16 @@ Inc., 59 Temple Place, Suite 330, Boston
#include <sys/syslog.h>
#include <sys/time.h>
#include <sys/types.h>
-#include <time.h>
-#include <unistd.h>
-#include <values.h>
+
+#include <asm/unistd.h>
-#define VERSION_STRING "20121130"
+#define VERSION_STRING "20150602"
#define VAR_RUN_FILE "/var/run/numad.pid"
#define VAR_LOG_FILE "/var/log/numad.log"
-char *cpuset_dir = NULL;
-char *cpuset_dir_list[] = {
- NULL,
- "/sys/fs/cgroup/cpuset",
- "/cgroup/cpuset",
- NULL
-};
-
-
#define KILOBYTE (1024)
#define MEGABYTE (1024 * 1024)
@@ -86,14 +80,11 @@ char *cpuset_dir_list[] = {
#define MAX_INTERVAL 15
#define CPU_THRESHOLD 50
#define MEMORY_THRESHOLD 300
-#define TARGET_UTILIZATION_PERCENT 85
-#define IMPROVEMENT_THRESHOLD_PERCENT 5
-
+#define DEFAULT_HTT_PERCENT 20
+#define DEFAULT_THP_SCAN_SLEEP_MS 1000
+#define DEFAULT_UTILIZATION_PERCENT 85
+#define DEFAULT_MEMLOCALITY_PERCENT 90
-#define ELIM_NEW_LINE(s) \
- if (s[strlen(s) - 1] == '\n') { \
- s[strlen(s) - 1] = '\0'; \
- }
#define CONVERT_DIGITS_TO_NUM(p, n) \
n = *p++ - '0'; \
@@ -105,19 +96,36 @@ char *cpuset_dir_list[] = {
int num_cpus = 0;
int num_nodes = 0;
-int page_size_in_bytes = 0;
-int huge_page_size_in_bytes = 0;
+int threads_per_core = 0;
+uint64_t page_size_in_bytes = 0;
+uint64_t huge_page_size_in_bytes = 0;
int min_interval = MIN_INTERVAL;
int max_interval = MAX_INTERVAL;
-int target_utilization = TARGET_UTILIZATION_PERCENT;
+int htt_percent = DEFAULT_HTT_PERCENT;
+int thp_scan_sleep_ms = DEFAULT_THP_SCAN_SLEEP_MS;
+int target_utilization = DEFAULT_UTILIZATION_PERCENT;
+int target_memlocality = DEFAULT_MEMLOCALITY_PERCENT;
int scan_all_processes = 1;
int keep_interleaved_memory = 0;
+int use_inactive_file_cache = 1;
pthread_mutex_t pid_list_mutex;
pthread_mutex_t node_info_mutex;
+long sum_CPUs_total = 0;
int requested_mbs = 0;
int requested_cpus = 0;
+int got_sighup = 0;
+int got_sigterm = 0;
+int got_sigquit = 0;
+
+void sig_handler(int signum) {
+ switch (signum) {
+ case SIGHUP: got_sighup = 1; break;
+ case SIGTERM: got_sigterm = 1; break;
+ case SIGQUIT: got_sigquit = 1; break;
+ }
+}
@@ -139,7 +147,7 @@ void numad_log(int level, const char *fm
}
char buf[BUF_SIZE];
time_t ts = time(NULL);
- sprintf(buf, ctime(&ts));
+ strncpy(buf, ctime(&ts), sizeof(buf));
char *p = &buf[strlen(buf) - 1];
*p++ = ':';
*p++ = ' ';
@@ -155,13 +163,16 @@ void open_log_file() {
log_fs = fopen(VAR_LOG_FILE, "a");
if (log_fs == NULL) {
log_fs = stderr;
- numad_log(LOG_ERR, "Cannot open numad log file -- using stderr\n");
+ numad_log(LOG_ERR, "Cannot open numad log file (errno: %d) -- using stderr\n", errno);
}
}
+
void close_log_file() {
if (log_fs != NULL) {
- fclose(log_fs);
+ if (log_fs != stderr) {
+ fclose(log_fs);
+ }
log_fs = NULL;
}
}
@@ -235,23 +246,32 @@ void send_msg(long dst_pid, long cmd, lo
typedef struct id_list {
- // Use CPU_SET(3) <sched.h> cpuset bitmasks,
+ // Use CPU_SET(3) <sched.h> bitmasks,
// but bundle size and pointer together
// and genericize for both CPU and Node IDs
cpu_set_t *set_p;
size_t bytes;
} id_list_t, *id_list_p;
-#define INIT_ID_LIST(list_p) \
+#define ID_LIST_SET_P(list_p) (list_p->set_p)
+#define ID_LIST_BYTES(list_p) (list_p->bytes)
+
+#define INIT_ID_LIST(list_p, num_elements) \
list_p = malloc(sizeof(id_list_t)); \
if (list_p == NULL) { numad_log(LOG_CRIT, "INIT_ID_LIST malloc failed\n"); exit(EXIT_FAILURE); } \
- list_p->set_p = CPU_ALLOC(num_cpus); \
+ list_p->set_p = CPU_ALLOC(num_elements); \
if (list_p->set_p == NULL) { numad_log(LOG_CRIT, "CPU_ALLOC failed\n"); exit(EXIT_FAILURE); } \
- list_p->bytes = CPU_ALLOC_SIZE(num_cpus);
+ list_p->bytes = CPU_ALLOC_SIZE(num_elements);
-#define CLEAR_LIST(list_p) \
+#define CLEAR_CPU_LIST(list_p) \
if (list_p == NULL) { \
- INIT_ID_LIST(list_p); \
+ INIT_ID_LIST(list_p, num_cpus); \
+ } \
+ CPU_ZERO_S(list_p->bytes, list_p->set_p)
+
+#define CLEAR_NODE_LIST(list_p) \
+ if (list_p == NULL) { \
+ INIT_ID_LIST(list_p, num_nodes); \
} \
CPU_ZERO_S(list_p->bytes, list_p->set_p)
@@ -262,6 +282,9 @@ typedef struct id_list {
list_p = NULL; \
}
+#define COPY_LIST(orig_list_p, copy_list_p) \
+ memcpy(copy_list_p->set_p, orig_list_p->set_p, orig_list_p->bytes)
+
#define NUM_IDS_IN_LIST(list_p) CPU_COUNT_S(list_p->bytes, list_p->set_p)
#define ADD_ID_TO_LIST(k, list_p) CPU_SET_S(k, list_p->bytes, list_p->set_p)
#define CLR_ID_IN_LIST(k, list_p) CPU_CLR_S(k, list_p->bytes, list_p->set_p)
@@ -272,6 +295,25 @@ typedef struct id_list {
#define OR_LISTS( or_list_p, list_1_p, list_2_p) CPU_OR_S( or_list_p->bytes, or_list_p->set_p, list_1_p->set_p, list_2_p->set_p)
#define XOR_LISTS(xor_list_p, list_1_p, list_2_p) CPU_XOR_S(xor_list_p->bytes, xor_list_p->set_p, list_1_p->set_p, list_2_p->set_p)
+int negate_cpu_list(id_list_p list_p) {
+ if (list_p == NULL) {
+ numad_log(LOG_CRIT, "Cannot negate a NULL list\n");
+ exit(EXIT_FAILURE);
+ }
+ if (num_cpus < 1) {
+ numad_log(LOG_CRIT, "No CPUs to negate in list!\n");
+ exit(EXIT_FAILURE);
+ }
+ for (int ix = 0; (ix < num_cpus); ix++) {
+ if (ID_IS_IN_LIST(ix, list_p)) {
+ CLR_ID_IN_LIST(ix, list_p);
+ } else {
+ ADD_ID_TO_LIST(ix, list_p);
+ }
+ }
+ return NUM_IDS_IN_LIST(list_p);
+}
+
int add_ids_to_list_from_str(id_list_p list_p, char *s) {
if (list_p == NULL) {
numad_log(LOG_CRIT, "Cannot add to NULL list\n");
@@ -352,9 +394,21 @@ typedef struct node_data {
uint8_t *distance;
id_list_p cpu_list_p;
} node_data_t, *node_data_p;
-
node_data_p node = NULL;
+int min_node_CPUs_free_ix = -1;
+int min_node_MBs_free_ix = -1;
+long min_node_CPUs_free = MAXINT;
+long min_node_MBs_free = MAXINT;
+long max_node_CPUs_free = 0;
+long max_node_MBs_free = 0;
+long avg_node_CPUs_free = 0;
+long avg_node_MBs_free = 0;
+double stddev_node_CPUs_free = 0.0;
+double stddev_node_MBs_free = 0.0;
+
+
+
// RING_BUF_SIZE must be a power of two
#define RING_BUF_SIZE 8
@@ -366,14 +420,15 @@ typedef struct process_data {
uint64_t data_time_stamp; // hundredths of seconds
uint64_t bind_time_stamp;
uint64_t num_threads;
+ uint64_t MBs_size;
uint64_t MBs_used;
uint64_t cpu_util;
uint64_t CPUs_used; // scaled * ONE_HUNDRED
uint64_t CPUs_used_ring_buf[RING_BUF_SIZE];
int ring_buf_ix;
- int dup_bind_count;
char *comm;
- char *cpuset_name;
+ id_list_p node_list_p;
+ uint64_t *process_MBs;
} process_data_t, *process_data_p;
@@ -433,7 +488,8 @@ int process_hash_insert(int pid) {
}
int process_hash_update(process_data_p newp) {
- // This updates hash table stats for processes we are monitoring
+ // This updates hash table stats for processes we are monitoring. Only the
+ // scalar resource consumption stats need to be updated here.
int new_hash_table_entry = 1;
int ix = process_hash_insert(newp->pid);
if (ix >= 0) {
@@ -460,6 +516,7 @@ int process_hash_update(process_data_p n
}
p->comm = strdup(newp->comm);
}
+ p->MBs_size = newp->MBs_size;
p->MBs_used = newp->MBs_used;
p->cpu_util = newp->cpu_util;
p->num_threads = newp->num_threads;
@@ -468,6 +525,11 @@ int process_hash_update(process_data_p n
return new_hash_table_entry;
}
+void process_hash_clear_all_bind_time_stamps() {
+ for (int ix = 0; (ix < process_hash_table_size); ix++) {
+ process_hash_table[ix].bind_time_stamp = 0;
+ }
+}
int process_hash_rehash(int old_ix) {
// Given the index of a table entry that would otherwise be orphaned by
@@ -489,7 +551,8 @@ int process_hash_remove(int pid) {
// remove the target
process_data_p dp = &process_hash_table[ix];
if (dp->comm) { free(dp->comm); }
- if (dp->cpuset_name) { free(dp->cpuset_name); }
+ if (dp->process_MBs) { free(dp->process_MBs); }
+ FREE_LIST(dp->node_list_p);
memset(dp, 0, sizeof(process_data_t));
// bubble up the collision chain and rehash if neeeded
for (;;) {
@@ -543,15 +606,15 @@ void process_hash_table_dump() {
process_data_p p = &process_hash_table[ix];
if (p->pid) {
numad_log(LOG_DEBUG,
- "ix: %d PID: %d %s Thds: %d CPU %ld MBs: %ld Data TS: %ld Bind TS: %ld\n",
+ "ix: %d PID: %d %s Thds: %d CPU %ld MBs: %ld/%ld Data TS: %ld Bind TS: %ld\n",
ix, p->pid, ((p->comm != NULL) ? p->comm : "(Null)"), p->num_threads,
- p->CPUs_used, p->MBs_used, p->data_time_stamp, p->bind_time_stamp);
+ p->CPUs_used, p->MBs_used, p->MBs_size, p->data_time_stamp, p->bind_time_stamp);
+ // FIXME: make this dump every field, but this is not even currently used
}
}
}
void process_hash_table_cleanup(uint64_t update_time) {
- int cpusets_removed = 0;
int num_hash_entries_used = 0;
for (int ix = 0; (ix < process_hash_table_size); ix++) {
process_data_p p = &process_hash_table[ix];
@@ -562,34 +625,14 @@ void process_hash_table_cleanup(uint64_t
p->data_time_stamp = 0;
p->CPUs_used = 0;
// Check for dead pids and remove them...
- char fname[FNAME_SIZE];
- snprintf(fname, FNAME_SIZE, "/proc/%d", p->pid);
- if (access(fname, F_OK) < 0) {
- // Seems dead. Forget this pid -- after first checking
- // and removing obsolete numad.PID cpuset directories.
- snprintf(fname, FNAME_SIZE, "%s/numad.%d", cpuset_dir, p->pid);
- if (access(fname, F_OK) == 0) {
- numad_log(LOG_NOTICE, "Removing obsolete cpuset: %s\n", fname);
- int rc = rmdir(fname);
- if (rc >= 0) {
- cpusets_removed += 1;
- } else {
- numad_log(LOG_ERR, "bad cpuset rmdir\n");
- // exit(EXIT_FAILURE);
- }
- }
+ if ((kill(p->pid, 0) == -1) && (errno == ESRCH)) {
+ // Seems dead. Forget this pid
process_hash_remove(p->pid);
num_hash_entries_used -= 1;
}
}
}
}
- if (cpusets_removed > 0) {
- // Expire all the duplicate bind counts so things will be re-evaluated sooner.
- for (int ix = 0; (ix < process_hash_table_size); ix++) {
- process_hash_table[ix].dup_bind_count = 0;
- }
- }
// Keep hash table approximately half empty
if ((num_hash_entries_used * 7) / 4 > process_hash_table_size) {
process_hash_table_expand();
@@ -610,9 +653,7 @@ pid_list_p insert_pid_into_pid_list(pid_
if (process_hash_table != NULL) {
int hash_ix = process_hash_lookup(pid);
if ((hash_ix >= 0) && (list_ptr == include_pid_list)) {
- // Clear dup_bind_count and interleaved flag,
- // in case user wants it to be re-evaluated soon
- process_hash_table[hash_ix].dup_bind_count = 0;
+ // Clear interleaved flag, in case user wants it to be re-evaluated
process_hash_table[hash_ix].flags &= ~PROCESS_FLAG_INTERLEAVED;
}
}
@@ -678,18 +719,23 @@ void print_version_and_exit(char *prog_n
void print_usage_and_exit(char *prog_name) {
fprintf(stderr, "Usage: %s <options> ...\n", prog_name);
+ fprintf(stderr, "-C 1 to count inactive file cache as available memory (default 1)\n");
+ fprintf(stderr, "-C 0 to count inactive file cache memory as unavailable (default 1)\n");
fprintf(stderr, "-d for debug logging (same effect as '-l 7')\n");
- fprintf(stderr, "-D <CGROUP_MOUNT_POINT> to specify cgroup mount point\n");
fprintf(stderr, "-h to print this usage info\n");
+ fprintf(stderr, "-H <N> to set THP scan_sleep_ms (default %d)\n", DEFAULT_THP_SCAN_SLEEP_MS);
fprintf(stderr, "-i [<MIN>:]<MAX> to specify interval seconds\n");
- fprintf(stderr, "-K 1 to keep interleaved memory spread across nodes\n");
- fprintf(stderr, "-K 0 to merge interleaved memory to local NUMA nodes\n");
- fprintf(stderr, "-l <N> to specify logging level (usually 5, 6, or 7)\n");
+ fprintf(stderr, "-K 1 to keep interleaved memory spread across nodes (default 0)\n");
+ fprintf(stderr, "-K 0 to merge interleaved memory to local NUMA nodes (default 0)\n");
+ fprintf(stderr, "-l <N> to specify logging level (usually 5, 6, or 7 -- default 5)\n");
+ fprintf(stderr, "-m <N> to specify memory locality target percent (default %d)\n", DEFAULT_MEMLOCALITY_PERCENT);
fprintf(stderr, "-p <PID> to add PID to inclusion pid list\n");
fprintf(stderr, "-r <PID> to remove PID from explicit pid lists\n");
- fprintf(stderr, "-S 1 to scan all processes\n");
- fprintf(stderr, "-S 0 to scan only explicit PID list processes\n");
- fprintf(stderr, "-u <N> to specify target utilization percent (default 85)\n");
+ fprintf(stderr, "-R <CPU_LIST> to reserve some CPUs for non-numad use\n");
+ fprintf(stderr, "-S 1 to scan all processes (default 1)\n");
+ fprintf(stderr, "-S 0 to scan only explicit PID list processes (default 1)\n");
+ fprintf(stderr, "-t <N> to specify thread / logical CPU valuation percent (default %d)\n", DEFAULT_HTT_PERCENT);
+ fprintf(stderr, "-u <N> to specify utilization target percent (default %d)\n", DEFAULT_UTILIZATION_PERCENT);
fprintf(stderr, "-v for verbose (same effect as '-l 6')\n");
fprintf(stderr, "-V to show version info\n");
fprintf(stderr, "-w <CPUs>[:<MBs>] for NUMA node suggestions\n");
@@ -698,62 +744,35 @@ void print_usage_and_exit(char *prog_nam
}
-void check_prereqs(char *prog_name) {
- // Verify cpusets are available on this system.
- char **dir = &cpuset_dir_list[0];
- if (*dir == NULL) { dir++; }
- while (*dir != NULL) {
- cpuset_dir = *dir;
- char fname[FNAME_SIZE];
- snprintf(fname, FNAME_SIZE, "%s/cpuset.cpus", cpuset_dir);
- if (access(fname, F_OK) == 0) {
- break;
- }
- dir++;
- }
- if (*dir == NULL) {
- fprintf(stderr, "\n");
- fprintf(stderr, "Are CPUSETs enabled on this system?\n");
- fprintf(stderr, "They are required for %s to function.\n\n", prog_name);
- fprintf(stderr, "Check manpage CPUSET(7). You might need to do something like:\n");
- fprintf(stderr, " # mkdir <DIRECTORY_MOUNT_POINT>\n");
- fprintf(stderr, " # mount cgroup -t cgroup -o cpuset <DIRECTORY_MOUNT_POINT>\n");
- fprintf(stderr, " where <DIRECTORY_MOUNT_POINT> is something like:\n");
- dir = &cpuset_dir_list[0];
- if (*dir == NULL) { dir++; }
- while (*dir != NULL) {
- fprintf(stderr, " - %s\n", *dir);
- dir++;
- }
- fprintf(stderr, "and then try again...\n");
- fprintf(stderr, "Or, use '-D <DIRECTORY_MOUNT_POINT>' to specify the correct mount point\n");
- fprintf(stderr, "\n");
- exit(EXIT_FAILURE);
+void set_thp_scan_sleep_ms(int new_ms) {
+ if (new_ms < 1) {
+ // 0 means do not change the system default
+ return;
}
- // Check on THP scan sleep time.
- char *thp_scan_fname = "/sys/kernel/mm/redhat_transparent_hugepage/khugepaged/scan_sleep_millisecs";
- int fd = open(thp_scan_fname, O_RDONLY, 0);
+ char *thp_scan_fname = "/sys/kernel/mm/transparent_hugepage/khugepaged/scan_sleep_millisecs";
+ int fd = open(thp_scan_fname, O_RDWR, 0);
if (fd >= 0) {
- int ms;
char buf[BUF_SIZE];
int bytes = read(fd, buf, BUF_SIZE);
- close(fd);
if (bytes > 0) {
+ buf[bytes] = '\0';
+ int cur_ms;
char *p = buf;
- CONVERT_DIGITS_TO_NUM(p, ms);
- if (ms > 150) {
- fprintf(stderr, "\n");
- numad_log(LOG_NOTICE, "Looks like transparent hugepage scan time in %s is %d ms.\n", thp_scan_fname, ms);
- fprintf(stderr, "Looks like transparent hugepage scan time in %s is %d ms.\n", thp_scan_fname, ms);
- fprintf(stderr, "Consider increasing the frequency of THP scanning,\n");
- fprintf(stderr, "by echoing a smaller number (e.g. 100) to %s\n", thp_scan_fname);
- fprintf(stderr, "to more aggressively (re)construct THPs. For example:\n");
- fprintf(stderr, "# echo 100 > /sys/kernel/mm/redhat_transparent_hugepage/khugepaged/scan_sleep_millisecs\n");
- fprintf(stderr, "\n");
+ CONVERT_DIGITS_TO_NUM(p, cur_ms);
+ if (cur_ms != new_ms) {
+ lseek(fd, 0, SEEK_SET);
+ numad_log(LOG_NOTICE, "Changing THP scan time in %s from %d to %d ms.\n", thp_scan_fname, cur_ms, new_ms);
+ sprintf(buf, "%d\n", new_ms);
+ write(fd, buf, strlen(buf));
}
}
+ close(fd);
}
- // FIXME: ?? check for enabled ksmd, and recommend disabling ksm?
+}
+
+void check_prereqs(char *prog_name) {
+ // Adjust kernel tunable to scan for THP more frequently...
+ set_thp_scan_sleep_ms(thp_scan_sleep_ms);
}
@@ -785,7 +804,6 @@ int get_daemon_pid() {
return pid;
}
-
int register_numad_pid() {
int pid;
char buf[BUF_SIZE];
@@ -831,6 +849,43 @@ fail_numad_run_file:
}
+int count_set_bits_in_hex_list_file(char *fname) {
+ int sum = 0;
+ int fd = open(fname, O_RDONLY, 0);
+ if (fd >= 0) {
+ char buf[BUF_SIZE];
+ int bytes = read(fd, buf, BUF_SIZE);
+ close(fd);
+ for (int ix = 0; (ix < bytes); ix++) {
+ char c = tolower(buf[ix]);
+ switch (c) {
+ case '0' : sum += 0; break;
+ case '1' : sum += 1; break;
+ case '2' : sum += 1; break;
+ case '3' : sum += 2; break;
+ case '4' : sum += 1; break;
+ case '5' : sum += 2; break;
+ case '6' : sum += 2; break;
+ case '7' : sum += 3; break;
+ case '8' : sum += 1; break;
+ case '9' : sum += 2; break;
+ case 'a' : sum += 2; break;
+ case 'b' : sum += 3; break;
+ case 'c' : sum += 2; break;
+ case 'd' : sum += 3; break;
+ case 'e' : sum += 3; break;
+ case 'f' : sum += 4; break;
+ case ' ' : sum += 0; break;
+ case ',' : sum += 0; break;
+ case '\n' : sum += 0; break;
+ default : numad_log(LOG_CRIT, "Unexpected character in list\n"); exit(EXIT_FAILURE);
+ }
+ }
+ }
+ return sum;
+}
+
+
int get_num_cpus() {
int n1 = sysconf(_SC_NPROCESSORS_CONF);
int n2 = sysconf(_SC_NPROCESSORS_ONLN);
@@ -848,7 +903,7 @@ int get_num_cpus() {
int get_num_kvm_vcpu_threads(int pid) {
// Try to return the number of vCPU threads for this VM guest,
// excluding the IO threads. All failures return MAXINT.
- // FIXME: figure out some better way to do this...
+ // FIXME: someday figure out some better way to do this...
char fname[FNAME_SIZE];
snprintf(fname, FNAME_SIZE, "/proc/%d/cmdline", pid);
int fd = open(fname, O_RDONLY, 0);
@@ -876,8 +931,8 @@ int get_num_kvm_vcpu_threads(int pid) {
}
-int get_huge_page_size_in_bytes() {
- int huge_page_size = 0;;
+uint64_t get_huge_page_size_in_bytes() {
+ uint64_t huge_page_size = 0;;
FILE *fs = fopen("/proc/meminfo", "r");
if (!fs) {
numad_log(LOG_CRIT, "Can't open /proc/meminfo\n");
@@ -890,7 +945,7 @@ int get_huge_page_size_in_bytes() {
while ((!isdigit(*p)) && (p < buf + BUF_SIZE)) {
p++;
}
- huge_page_size = atoi(p);
+ huge_page_size = atol(p);
break;
}
}
@@ -916,143 +971,134 @@ static int name_starts_with_digit(const
}
-int bind_process_and_migrate_memory(int pid, char *cpuset_name, id_list_p node_list_p, id_list_p cpu_list_p) {
- // Check basic parameter validity.
- if (pid <= 0) {
+
+#define BITS_IN_LONG (CHAR_BIT * sizeof(unsigned long))
+#define SET_BIT(i,a) (a)[(i) / BITS_IN_LONG] |= (1u << ((i) % BITS_IN_LONG))
+#define TEST_BIT(i,a) (((a)[(i) / BITS_IN_LONG] & (1u << ((i) % BITS_IN_LONG))) != 0)
+#define CLEAR_BIT(i,a) (a)[(i) / BITS_IN_LONG] &= ~(1u << ((i) % BITS_IN_LONG))
+
+int bind_process_and_migrate_memory(process_data_p p) {
+ uint64_t t0 = get_time_stamp();
+ // Parameter p is a pointer to an element in the hash table
+ if ((!p) || (p->pid < 1)) {
numad_log(LOG_CRIT, "Bad PID to bind\n");
exit(EXIT_FAILURE);
}
- if ((cpuset_name == NULL) || (strlen(cpuset_name) == 0)) {
- numad_log(LOG_CRIT, "Bad cpuset name to bind\n");
- exit(EXIT_FAILURE);
- }
- int nodes;
- if ((node_list_p == NULL) || ((nodes = NUM_IDS_IN_LIST(node_list_p)) == 0)) {
- numad_log(LOG_CRIT, "Cannot bind to unspecified node\n");
+ if (!p->node_list_p) {
+ numad_log(LOG_CRIT, "Cannot bind to unspecified node(s)\n");
exit(EXIT_FAILURE);
}
- // Cpu_list_p is optional and may be NULL...
- // Generate CPU id list from the specified node list if necessary
- if (cpu_list_p == NULL) {
- static id_list_p tmp_cpu_list_p;
- CLEAR_LIST(tmp_cpu_list_p);
- int node_id = 0;
- while (nodes) {
- if (ID_IS_IN_LIST(node_id, node_list_p)) {
- OR_LISTS(tmp_cpu_list_p, tmp_cpu_list_p, node[node_id].cpu_list_p);
- nodes -= 1;
- }
- node_id += 1;
- }
- cpu_list_p = tmp_cpu_list_p;
- }
- // Make the cpuset directory if necessary
- char cpuset_name_buf[FNAME_SIZE];
- snprintf(cpuset_name_buf, FNAME_SIZE, "%s%s", cpuset_dir, cpuset_name);
- char *p = &cpuset_name_buf[strlen(cpuset_dir)];
- if (!strcmp(p, "/")) {
- // Make a cpuset directory for this process
- snprintf(cpuset_name_buf, FNAME_SIZE, "%s/numad.%d", cpuset_dir, pid);
- numad_log(LOG_NOTICE, "Making new cpuset: %s\n", cpuset_name_buf);
- int rc = mkdir(cpuset_name_buf, S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH);
- if (rc == -1) {
- numad_log(LOG_CRIT, "Bad cpuset mkdir -- errno: %d\n", errno);
- return 0;
+ // Generate CPU list derived from target node list.
+ static id_list_p cpu_bind_list_p;
+ CLEAR_CPU_LIST(cpu_bind_list_p);
+ int nodes = NUM_IDS_IN_LIST(p->node_list_p);
+ int node_id = 0;
+ while (nodes) {
+ if (ID_IS_IN_LIST(node_id, p->node_list_p)) {
+ OR_LISTS(cpu_bind_list_p, cpu_bind_list_p, node[node_id].cpu_list_p);
+ nodes -= 1;
}
+ node_id += 1;
}
- cpuset_name = cpuset_name_buf;
- // Now that we have a cpuset for pid and a populated cpulist,
- // start the actual binding and migration.
- uint64_t t0 = get_time_stamp();
-
- // Write "1" out to cpuset.memory_migrate file
char fname[FNAME_SIZE];
- snprintf(fname, FNAME_SIZE, "%s/cpuset.memory_migrate", cpuset_name);
- int fd = open(fname, O_WRONLY | O_TRUNC, 0);
- if (fd == -1) {
- numad_log(LOG_CRIT, "Could not open cpuset.memory_migrate -- errno: %d\n", errno);
- return 0;
- }
- write(fd, "1", 1);
- close(fd);
-
- // Write node IDs out to cpuset.mems file
- char node_list_buf[BUF_SIZE];
- snprintf(fname, FNAME_SIZE, "%s/cpuset.mems", cpuset_name);
- fd = open(fname, O_WRONLY | O_TRUNC, 0);
- if (fd == -1) {
- numad_log(LOG_CRIT, "Could not open cpuset.mems -- errno: %d\n", errno);
- return 0;
- }
- int len = str_from_id_list(node_list_buf, BUF_SIZE, node_list_p);
- write(fd, node_list_buf, len);
- close(fd);
-
- // Write CPU IDs out to cpuset.cpus file
- char cpu_list_buf[BUF_SIZE];
- snprintf(fname, FNAME_SIZE, "%s/cpuset.cpus", cpuset_name);
- fd = open(fname, O_WRONLY | O_TRUNC, 0);
- if (fd == -1) {
- numad_log(LOG_CRIT, "Could not open cpuset.cpus -- errno: %d\n", errno);
- return 0;
- }
- len = str_from_id_list(cpu_list_buf, BUF_SIZE, cpu_list_p);
- write(fd, cpu_list_buf, len);
- close(fd);
-
- // Copy pid tasks one at a time to tasks file
- snprintf(fname, FNAME_SIZE, "%s/tasks", cpuset_name);
- fd = open(fname, O_WRONLY | O_TRUNC, 0);
- if (fd == -1) {
- numad_log(LOG_CRIT, "Could not open tasks -- errno: %d\n", errno);
- return 0;
- }
- snprintf(fname, FNAME_SIZE, "/proc/%d/task", pid);
struct dirent **namelist;
- int files = scandir(fname, &namelist, name_starts_with_digit, NULL);
- if (files < 0) {
- numad_log(LOG_WARNING, "Could not scandir task list\n");
+ snprintf(fname, FNAME_SIZE, "/proc/%d/task", p->pid);
+ int num_tasks = scandir(fname, &namelist, name_starts_with_digit, NULL);
+ if (num_tasks <= 0) {
+ numad_log(LOG_WARNING, "Could not scandir task list for PID: %d\n", p->pid);
return 0; // Assume the process terminated
}
- for (int ix = 0; (ix < files); ix++) {
- // copy pid tasks, one at a time
- numad_log(LOG_NOTICE, "Including task: %s\n", namelist[ix]->d_name);
- write(fd, namelist[ix]->d_name, strlen(namelist[ix]->d_name));
- free(namelist[ix]);
+ // Set the affinity of each task in the process...
+ for (int namelist_ix = 0; (namelist_ix < num_tasks); namelist_ix++) {
+ int tid = atoi(namelist[namelist_ix]->d_name);
+ int rc = sched_setaffinity(tid, ID_LIST_BYTES(cpu_bind_list_p), ID_LIST_SET_P(cpu_bind_list_p));
+ if (rc < 0) {
+ // Check errno
+ if (errno == ESRCH) {
+ numad_log(LOG_WARNING, "Tried to move PID %d, TID %d, but it apparently went away.\n", p->pid, tid);
+ }
+ numad_log(LOG_ERR, "Bad sched_setaffinity() on PID %d, TID %d -- errno: %d\n", p->pid, tid, errno);
+ }
+ free(namelist[namelist_ix]);
}
free(namelist);
- close(fd);
-
- uint64_t t1 = get_time_stamp();
+ // Now move the memory to the target nodes....
+ static unsigned long *dest_mask;
+ static unsigned long *from_mask;
+ static int allocated_bytes_in_masks;
+ // Lie about num_nodes being one bigger because of kernel bug...
+ int num_bytes_in_masks = (1 + ((num_nodes + 1) / BITS_IN_LONG)) * sizeof(unsigned long);
+ if (allocated_bytes_in_masks < num_bytes_in_masks) {
+ allocated_bytes_in_masks = num_bytes_in_masks;
+ dest_mask = realloc(dest_mask, num_bytes_in_masks);
+ from_mask = realloc(from_mask, num_bytes_in_masks);
+ if ((dest_mask == NULL) || (from_mask == NULL)) {
+ numad_log(LOG_CRIT, "bit mask malloc failed\n");
+ exit(EXIT_FAILURE);
+ }
+ }
+ // In an effort to put semi-balanced memory in each target node, move the
+ // contents from the source node with the max amount of memory to the
+ // destination node with the least amount of memory. Repeat until done.
+ int prev_from_node_id = -1;
+ for (;;) {
+ int min_dest_node_id = -1;
+ int max_from_node_id = -1;
+ for (int node_ix = 0; (node_ix < num_nodes); node_ix++) {
+ node_id = node[node_ix].node_id;
+ if (ID_IS_IN_LIST(node_id, p->node_list_p)) {
+ if ((min_dest_node_id < 0) || (p->process_MBs[min_dest_node_id] >= p->process_MBs[node_id])) {
+ // The ">=" above is intentional, so we tend to move memory to higher numbered nodes
+ min_dest_node_id = node_id;
+ }
+ } else {
+ if ((max_from_node_id < 0) || (p->process_MBs[max_from_node_id] < p->process_MBs[node_id])) {
+ max_from_node_id = node_id;
+ }
+ }
+ }
+ if ((p->process_MBs[max_from_node_id] == 0) || (max_from_node_id == prev_from_node_id)) {
+ break;
+ }
+ memset(dest_mask, 0, num_bytes_in_masks);
+ memset(from_mask, 0, num_bytes_in_masks);
+ SET_BIT(max_from_node_id, from_mask);
+ SET_BIT(min_dest_node_id, dest_mask);
+ numad_log(LOG_DEBUG, "Moving memory from node: %d to node %d\n", max_from_node_id, min_dest_node_id);
+ // Lie about num_nodes being one bigger because of kernel bug...
+ int rc = syscall(__NR_migrate_pages, p->pid, num_nodes + 1, from_mask, dest_mask);
+ if (rc > 2) {
+ // rc == the number of pages that could not be moved.
+ // A couple pages not moving is probably not a problem, hence ignoring rc == 1 or 2.
+ numad_log(LOG_WARNING, "Tried to move PID %d, but %d pages would not move.\n", p->pid, rc);
+ } else if (rc < 0) {
+ // Check errno
+ if (errno == ESRCH) {
+ numad_log(LOG_WARNING, "Tried to move PID %d, but it apparently went away.\n", p->pid);
+ return 0; // Assume the process terminated
+ }
+ }
+ // Assume memory did move for current accounting purposes...
+ p->process_MBs[min_dest_node_id] += p->process_MBs[max_from_node_id];
+ p->process_MBs[max_from_node_id] = 0;
+ prev_from_node_id = max_from_node_id;
+ }
// Check pid still active
- snprintf(fname, FNAME_SIZE, "/proc/%d", pid);
+ snprintf(fname, FNAME_SIZE, "/proc/%d", p->pid);
if (access(fname, F_OK) < 0) {
- numad_log(LOG_WARNING, "Could not migrate pid\n");
- return 0; // Assume the process terminated
+ numad_log(LOG_WARNING, "Could not migrate pid %d. Apparently it went away.\n", p->pid);
+ return 0;
+ } else {
+ uint64_t t1 = get_time_stamp();
+ p->bind_time_stamp = t1;
+ char node_list_str[BUF_SIZE];
+ str_from_id_list(node_list_str, BUF_SIZE, p->node_list_p);
+ numad_log(LOG_NOTICE, "PID %d moved to node(s) %s in %d.%d seconds\n", p->pid, node_list_str, (t1-t0)/100, (t1-t0)%100);
+ return 1;
}
- numad_log(LOG_NOTICE, "PID %d moved to node(s) %s in %d.%d seconds\n", pid, node_list_buf, (t1-t0)/100, (t1-t0)%100);
- return 1;
}
-void show_nodes() {
- time_t ts = time(NULL);
- fprintf(log_fs, "%s", ctime(&ts));
- fprintf(log_fs, "Nodes: %d\n", num_nodes);
- for (int ix = 0; (ix < num_nodes); ix++) {
- fprintf(log_fs, "Node %d: MBs_total %ld, MBs_free %6ld, CPUs_total %ld, CPUs_free %4ld, Distance: ",
- ix, node[ix].MBs_total, node[ix].MBs_free, node[ix].CPUs_total, node[ix].CPUs_free);
- for (int d = 0; (d < num_nodes); d++) {
- fprintf(log_fs, "%d ", node[ix].distance[d]);
- }
- char buf[BUF_SIZE];
- str_from_id_list(buf, BUF_SIZE, node[ix].cpu_list_p);
- fprintf(log_fs, " CPUs: %s\n", buf);
- }
- fprintf(log_fs, "\n");
- fflush(log_fs);
-}
-
typedef struct cpu_data {
uint64_t time_stamp;
@@ -1062,10 +1108,9 @@ typedef struct cpu_data {
cpu_data_t cpu_data_buf[2]; // Two sets, to calc deltas
int cur_cpu_data_buf = 0;
-
void update_cpu_data() {
// Parse idle percents from CPU stats in /proc/stat cpu<N> lines
- static FILE *fs = NULL;
+ static FILE *fs;
if (fs != NULL) {
rewind(fs);
} else {
@@ -1107,14 +1152,14 @@ void update_cpu_data() {
while (!isdigit(*p)) { p++; } while (isdigit(*p)) { p++; } // skip nice
while (!isdigit(*p)) { p++; } while (isdigit(*p)) { p++; } // skip system
while (!isdigit(*p)) { p++; }
- uint64_t idle = *p++ - '0'; while (isdigit(*p)) { idle *= 10; idle += (*p++ - '0'); }
+ uint64_t idle;
+ CONVERT_DIGITS_TO_NUM(p, idle);
cpu_data_buf[new].idle[cpu_id] = idle;
}
}
cur_cpu_data_buf = new;
}
-
int node_and_digits(const struct dirent *dptr) {
char *p = (char *)(dptr->d_name);
if (*p++ != 'n') return 0;
@@ -1129,10 +1174,31 @@ int node_and_digits(const struct dirent
}
+uint64_t node_info_time_stamp = 0;
id_list_p all_cpus_list_p = NULL;
id_list_p all_nodes_list_p = NULL;
-uint64_t node_info_time_stamp = 0;
+id_list_p reserved_cpu_mask_list_p = NULL;
+char *reserved_cpu_str = NULL;
+void show_nodes() {
+ fprintf(log_fs, "\n");
+ numad_log(LOG_INFO, "Nodes: %d\n", num_nodes);
+ fprintf(log_fs, "Min CPUs free: %ld, Max CPUs: %ld, Avg CPUs: %ld, StdDev: %lg\n",
+ min_node_CPUs_free, max_node_CPUs_free, avg_node_CPUs_free, stddev_node_CPUs_free);
+ fprintf(log_fs, "Min MBs free: %ld, Max MBs: %ld, Avg MBs: %ld, StdDev: %lg\n",
+ min_node_MBs_free, max_node_MBs_free, avg_node_MBs_free, stddev_node_MBs_free);
+ for (int ix = 0; (ix < num_nodes); ix++) {
+ fprintf(log_fs, "Node %d: MBs_total %ld, MBs_free %6ld, CPUs_total %ld, CPUs_free %4ld, Distance: ",
+ ix, node[ix].MBs_total, node[ix].MBs_free, node[ix].CPUs_total, node[ix].CPUs_free);
+ for (int d = 0; (d < num_nodes); d++) {
+ fprintf(log_fs, "%d ", node[ix].distance[d]);
+ }
+ char buf[BUF_SIZE];
+ str_from_id_list(buf, BUF_SIZE, node[ix].cpu_list_p);
+ fprintf(log_fs, " CPUs: %s\n", buf);
+ }
+ fflush(log_fs);
+}
int update_nodes() {
char fname[FNAME_SIZE];
@@ -1141,6 +1207,7 @@ int update_nodes() {
uint64_t time_stamp = get_time_stamp();
#define STATIC_NODE_INFO_DELAY (600 * ONE_HUNDRED)
if ((num_nodes == 0) || (node_info_time_stamp + STATIC_NODE_INFO_DELAY < time_stamp)) {
+ node_info_time_stamp = time_stamp;
// Count directory names of the form: /sys/devices/system/node/node<N>
struct dirent **namelist;
int num_files = scandir ("/sys/devices/system/node", &namelist, node_and_digits, NULL);
@@ -1167,8 +1234,15 @@ int update_nodes() {
}
num_nodes = num_files;
}
- CLEAR_LIST(all_cpus_list_p);
- CLEAR_LIST(all_nodes_list_p);
+ sum_CPUs_total = 0;
+ CLEAR_CPU_LIST(all_cpus_list_p);
+ CLEAR_NODE_LIST(all_nodes_list_p);
+ // Figure out how many threads per core there are (for later discounting of hyper-threads)
+ threads_per_core = count_set_bits_in_hex_list_file("/sys/devices/system/cpu/cpu0/topology/thread_siblings");
+ if (threads_per_core < 1) {
+ numad_log(LOG_CRIT, "Could not count threads per core\n");
+ exit(EXIT_FAILURE);
+ }
// For each "node<N>" filename present, save <N> in node[ix].node_id
// Note that the node id might not necessarily match the node ix.
// Also populate the cpu lists and distance vectors for this node.
@@ -1184,11 +1258,24 @@ int update_nodes() {
snprintf(fname, FNAME_SIZE, "/sys/devices/system/node/node%d/cpulist", node_id);
int fd = open(fname, O_RDONLY, 0);
if ((fd >= 0) && (read(fd, buf, BIG_BUF_SIZE) > 0)) {
+ buf[BIG_BUF_SIZE - 1] = '\0';
// get cpulist from the cpulist string
- CLEAR_LIST(node[node_ix].cpu_list_p);
+ CLEAR_CPU_LIST(node[node_ix].cpu_list_p);
int n = add_ids_to_list_from_str(node[node_ix].cpu_list_p, buf);
+ if (reserved_cpu_str != NULL) {
+ AND_LISTS(node[node_ix].cpu_list_p, node[node_ix].cpu_list_p, reserved_cpu_mask_list_p);
+ n = NUM_IDS_IN_LIST(node[node_ix].cpu_list_p);
+ }
OR_LISTS(all_cpus_list_p, all_cpus_list_p, node[node_ix].cpu_list_p);
- node[node_ix].CPUs_total = n * ONE_HUNDRED;
+ // Calculate total CPUs, but possibly discount hyper-threads
+ if ((threads_per_core == 1) || (htt_percent >= 100)) {
+ node[node_ix].CPUs_total = n * ONE_HUNDRED;
+ } else {
+ n /= threads_per_core;
+ node[node_ix].CPUs_total = n * ONE_HUNDRED;
+ node[node_ix].CPUs_total += n * (threads_per_core - 1) * htt_percent;
+ }
+ sum_CPUs_total += node[node_ix].CPUs_total;
close(fd);
} else {
numad_log(LOG_CRIT, "Could not get node cpu list\n");
@@ -1220,15 +1307,30 @@ int update_nodes() {
}
free(namelist);
}
- // Second, get the dynamic free memory and available CPU capacity
+ // Second, update the dynamic free memory and available CPU capacity
+ while (cpu_data_buf[cur_cpu_data_buf].time_stamp + 7 >= time_stamp) {
+ // Make sure at least 7/100 of a second has passed.
+ // Otherwise sleep for 1/10 second.
+ struct timespec ts = { 0, 100000000 };
+ nanosleep(&ts, &ts);
+ time_stamp = get_time_stamp();
+ }
update_cpu_data();
+ max_node_MBs_free = 0;
+ max_node_CPUs_free = 0;
+ min_node_MBs_free = MAXINT;
+ min_node_CPUs_free = MAXINT;
+ uint64_t sum_of_node_MBs_free = 0;
+ uint64_t sum_of_node_CPUs_free = 0;
for (int node_ix = 0; (node_ix < num_nodes); node_ix++) {
int node_id = node[node_ix].node_id;
// Get available memory info from node<N>/meminfo file
snprintf(fname, FNAME_SIZE, "/sys/devices/system/node/node%d/meminfo", node_id);
int fd = open(fname, O_RDONLY, 0);
if ((fd >= 0) && (read(fd, buf, BIG_BUF_SIZE) > 0)) {
+ close(fd);
uint64_t KB;
+ buf[BIG_BUF_SIZE - 1] = '\0';
char *p = strstr(buf, "MemTotal:");
if (p != NULL) {
p += 9;
@@ -1238,7 +1340,11 @@ int update_nodes() {
}
while (!isdigit(*p)) { p++; }
CONVERT_DIGITS_TO_NUM(p, KB);
- node[node_ix].MBs_total = KB / KILOBYTE;
+ node[node_ix].MBs_total = (KB / KILOBYTE);
+ if (node[node_ix].MBs_total < 1) {
+ // If a node has zero memory, remove it from the all_nodes_list...
+ CLR_ID_IN_LIST(node_id, all_nodes_list_p);
+ }
p = strstr(p, "MemFree:");
if (p != NULL) {
p += 8;
@@ -1248,8 +1354,28 @@ int update_nodes() {
}
while (!isdigit(*p)) { p++; }
CONVERT_DIGITS_TO_NUM(p, KB);
- node[node_ix].MBs_free = KB / KILOBYTE;
- close(fd);
+ node[node_ix].MBs_free = (KB / KILOBYTE);
+ if (use_inactive_file_cache) {
+ // Add inactive file cache quantity to "free" memory
+ p = strstr(p, "Inactive(file):");
+ if (p != NULL) {
+ p += 15;
+ } else {
+ numad_log(LOG_CRIT, "Could not get node Inactive(file)\n");
+ exit(EXIT_FAILURE);
+ }
+ while (!isdigit(*p)) { p++; }
+ CONVERT_DIGITS_TO_NUM(p, KB);
+ node[node_ix].MBs_free += (KB / KILOBYTE);
+ }
+ sum_of_node_MBs_free += node[node_ix].MBs_free;
+ if (min_node_MBs_free > node[node_ix].MBs_free) {
+ min_node_MBs_free = node[node_ix].MBs_free;
+ min_node_MBs_free_ix = node[node_ix].node_id;
+ }
+ if (max_node_MBs_free < node[node_ix].MBs_free) {
+ max_node_MBs_free = node[node_ix].MBs_free;
+ }
} else {
numad_log(LOG_CRIT, "Could not get node meminfo\n");
exit(EXIT_FAILURE);
@@ -1260,7 +1386,8 @@ int update_nodes() {
if (cpu_data_buf[old_cpu_data_buf].time_stamp > 0) {
uint64_t idle_ticks = 0;
int cpu = 0;
- int num_cpus_to_process = node[node_ix].CPUs_total / ONE_HUNDRED;
+ int num_lcpus = NUM_IDS_IN_LIST(node[node_ix].cpu_list_p);
+ int num_cpus_to_process = num_lcpus;
while (num_cpus_to_process) {
if (ID_IS_IN_LIST(cpu, node[node_ix].cpu_list_p)) {
idle_ticks += cpu_data_buf[cur_cpu_data_buf].idle[cpu]
@@ -1274,15 +1401,46 @@ int update_nodes() {
// printf("Node: %d CPUs: %ld time diff %ld Idle ticks %ld\n", node_id, node[node_ix].CPUs_total, time_diff, idle_ticks);
// assert(time_diff > 0);
node[node_ix].CPUs_free = (idle_ticks * ONE_HUNDRED) / time_diff;
+ // Possibly discount hyper-threads
+ if ((threads_per_core > 1) && (htt_percent < 100)) {
+ uint64_t htt_discount = (num_lcpus - (num_lcpus / threads_per_core)) * (100 - htt_percent);
+ if (node[node_ix].CPUs_free > htt_discount) {
+ node[node_ix].CPUs_free -= htt_discount;
+ } else {
+ node[node_ix].CPUs_free = 0;
+ }
+ }
if (node[node_ix].CPUs_free > node[node_ix].CPUs_total) {
node[node_ix].CPUs_free = node[node_ix].CPUs_total;
}
+ sum_of_node_CPUs_free += node[node_ix].CPUs_free;
+ if (min_node_CPUs_free > node[node_ix].CPUs_free) {
+ min_node_CPUs_free = node[node_ix].CPUs_free;
+ min_node_CPUs_free_ix = node[node_ix].node_id;
+ }
+ if (max_node_CPUs_free < node[node_ix].CPUs_free) {
+ max_node_CPUs_free = node[node_ix].CPUs_free;
+ }
node[node_ix].magnitude = node[node_ix].CPUs_free * node[node_ix].MBs_free;
} else {
node[node_ix].CPUs_free = 0;
node[node_ix].magnitude = 0;
}
}
+ avg_node_MBs_free = sum_of_node_MBs_free / num_nodes;
+ avg_node_CPUs_free = sum_of_node_CPUs_free / num_nodes;
+ double MBs_variance_sum = 0.0;
+ double CPUs_variance_sum = 0.0;
+ for (int node_ix = 0; (node_ix < num_nodes); node_ix++) {
+ double MBs_diff = (double)node[node_ix].MBs_free - (double)avg_node_MBs_free;
+ double CPUs_diff = (double)node[node_ix].CPUs_free - (double)avg_node_CPUs_free;
+ MBs_variance_sum += MBs_diff * MBs_diff;
+ CPUs_variance_sum += CPUs_diff * CPUs_diff;
+ }
+ double MBs_variance = MBs_variance_sum / (num_nodes);
+ double CPUs_variance = CPUs_variance_sum / (num_nodes);
+ stddev_node_MBs_free = sqrt(MBs_variance);
+ stddev_node_CPUs_free = sqrt(CPUs_variance);
if (log_level >= LOG_INFO) {
show_nodes();
}
@@ -1316,7 +1474,7 @@ typedef struct stat_data {
int64_t num_threads; // 19
int64_t itrealvalue;
uint64_t starttime;
- uint64_t vsize;
+ uint64_t vsize; // 22
int64_t rss; // 23
uint64_t rsslim;
uint64_t startcode;
@@ -1356,15 +1514,16 @@ process_data_p get_stat_data_for_pid(int
}
static char buf[BUF_SIZE];
int bytes = read(fd, buf, BUF_SIZE);
+ close(fd);
if (bytes < 50) {
numad_log(LOG_WARNING, "Could not read stat file: %s\n", fname);
return NULL;
}
- close(fd);
+ uint64_t val;
char *p = buf;
static process_data_t data;
// Get PID from field 0
- uint64_t val = *p++ - '0'; while (isdigit(*p)) { val *= 10; val += (*p++ - '0'); }
+ CONVERT_DIGITS_TO_NUM(p, val);
data.pid = val;
// Copy comm from field 1
while (*p == ' ') { p++; }
@@ -1373,23 +1532,27 @@ process_data_p get_stat_data_for_pid(int
// Skip fields 2 through 12
for (int ix = 0; (ix < 11); ix++) { while (*p != ' ') { p++; } while (*p == ' ') { p++; } }
// Get utime from field 13 for cpu_util
- val = *p++ - '0'; while (isdigit(*p)) { val *= 10; val += (*p++ - '0'); }
+ CONVERT_DIGITS_TO_NUM(p, val);
data.cpu_util = val;
// Get stime from field 14 to add on to cpu_util (which already has utime)
while (*p == ' ') { p++; }
- val = *p++ - '0'; while (isdigit(*p)) { val *= 10; val += (*p++ - '0'); }
+ CONVERT_DIGITS_TO_NUM(p, val);
data.cpu_util += val;
// Skip fields 15 through 18
while (*p == ' ') { p++; }
for (int ix = 0; (ix < 4); ix++) { while (*p != ' ') { p++; } while (*p == ' ') { p++; } }
// Get num_threads from field 19
- val = *p++ - '0'; while (isdigit(*p)) { val *= 10; val += (*p++ - '0'); }
+ CONVERT_DIGITS_TO_NUM(p, val);
data.num_threads = val;
- // Skip fields 20 through 22
+ // Skip fields 20 through 21
while (*p == ' ') { p++; }
- for (int ix = 0; (ix < 3); ix++) { while (*p != ' ') { p++; } while (*p == ' ') { p++; } }
+ for (int ix = 0; (ix < 2); ix++) { while (*p != ' ') { p++; } while (*p == ' ') { p++; } }
+ // Get vsize from field 22 to compute MBs_size
+ CONVERT_DIGITS_TO_NUM(p, val);
+ data.MBs_size = val / MEGABYTE;
// Get rss from field 23 to compute MBs_used
- val = *p++ - '0'; while (isdigit(*p)) { val *= 10; val += (*p++ - '0'); }
+ while (*p == ' ') { p++; }
+ CONVERT_DIGITS_TO_NUM(p, val);
data.MBs_used = (val * page_size_in_bytes) / MEGABYTE;
// Return pointer to data
return &data;
@@ -1471,446 +1634,409 @@ int update_processes() {
}
+int initialize_mem_node_list(process_data_p p) {
+ // Parameter p is a pointer to an element in the hash table
+ if ((!p) || (p->pid < 1)) {
+ numad_log(LOG_CRIT, "Cannot initialize mem node lists with bad PID\n");
+ exit(EXIT_FAILURE);
+ }
+ int n = 0;
+ char fname[FNAME_SIZE];
+ char buf[BIG_BUF_SIZE];
+ p->process_MBs = NULL;
+ CLEAR_NODE_LIST(p->node_list_p);
+ snprintf(fname, FNAME_SIZE, "/proc/%d/status", p->pid);
+ int fd = open(fname, O_RDONLY, 0);
+ if (fd < 0) {
+ numad_log(LOG_WARNING, "Tried to research PID %d, but it apparently went away.\n", p->pid);
+ return 0; // Assume the process terminated
+ }
+ int bytes = read(fd, buf, BIG_BUF_SIZE);
+ close(fd);
+ if (bytes <= 0) {
+ numad_log(LOG_WARNING, "Tried to research PID %d, but cannot read status file.\n", p->pid);
+ return 0; // Assume the process terminated
+ } else if (bytes >= BIG_BUF_SIZE) {
+ buf[BIG_BUF_SIZE - 1] = '\0';
+ } else {
+ buf[bytes] = '\0';
+ }
+ char *list_str_p = strstr(buf, "Mems_allowed_list:");
+ if (!list_str_p) {
+ numad_log(LOG_CRIT, "Could not get node Mems_allowed_list\n");
+ exit(EXIT_FAILURE);
+ }
+ list_str_p += 18;
+ while (!isdigit(*list_str_p)) { list_str_p++; }
+ n = add_ids_to_list_from_str(p->node_list_p, list_str_p);
+ if (n < num_nodes) {
+ // If process already bound to a subset of nodes when we discover it,
+ // set initial bind_time_stamp to 30 minutes ago...
+ p->bind_time_stamp = get_time_stamp() - (1800 * ONE_HUNDRED);
+ }
+ return n;
+}
-id_list_p pick_numa_nodes(int pid, int cpus, int mbs) {
- char buf[BUF_SIZE];
- char buf2[BUF_SIZE];
+
+uint64_t combined_value_of_weighted_resources(int ix, int mbs, int cpus, uint64_t MBs_free, uint64_t CPUs_free) {
+ int64_t needed_mem;
+ int64_t needed_cpu;
+ int64_t excess_mem;
+ int64_t excess_cpu;
+ if (MBs_free > mbs) {
+ needed_mem = mbs;
+ excess_mem = MBs_free - mbs;
+ } else {
+ needed_mem = MBs_free;
+ excess_mem = 0;
+ }
+ if (CPUs_free > cpus) {
+ needed_cpu = cpus;
+ excess_cpu = CPUs_free - cpus;
+ } else {
+ needed_cpu = CPUs_free;
+ excess_cpu = 0;
+ }
+ // Weight the available resources, and then calculate magnitude as
+ // product of available CPUs and available MBs.
+ int64_t memfactor = (needed_mem * 10 + excess_mem * 4);
+ int64_t cpufactor = (needed_cpu * 6 + excess_cpu * 1);
+ numad_log(LOG_DEBUG, " Node[%d]: mem: %ld cpu: %ld\n", ix, memfactor, cpufactor);
+ return (memfactor * cpufactor);
+}
+
+
+id_list_p pick_numa_nodes(int pid, int cpus, int mbs, int assume_enough_cpus) {
if (log_level >= LOG_DEBUG) {
numad_log(LOG_DEBUG, "PICK NODES FOR: PID: %d, CPUs %d, MBs %d\n", pid, cpus, mbs);
}
- int num_existing_mems = 0;
- static id_list_p existing_mems_list_p;
- CLEAR_LIST(existing_mems_list_p);
- uint64_t time_stamp = get_time_stamp();
- static node_data_p tmp_node;
- static uint64_t *process_MBs;
- static uint64_t *saved_magnitude_for_node;
- static int process_MBs_num_nodes;
- // See if dynamic structures need to grow.
- if (process_MBs_num_nodes < num_nodes + 1) {
- process_MBs_num_nodes = num_nodes + 1;
- // The "+1 node" is for accumulating interleaved memory
- process_MBs = realloc(process_MBs, process_MBs_num_nodes * sizeof(uint64_t));
- tmp_node = realloc(tmp_node, num_nodes * sizeof(node_data_t) );
- saved_magnitude_for_node = realloc(saved_magnitude_for_node, num_nodes * sizeof(uint64_t));
- if ((process_MBs == NULL) || (tmp_node == NULL) || (saved_magnitude_for_node == NULL)) {
- numad_log(LOG_CRIT, "process_MBs realloc failed\n");
- exit(EXIT_FAILURE);
- }
- }
+ char buf[BUF_SIZE];
+ uint64_t proc_avg_node_CPUs_free = 0;
// For existing processes, get miscellaneous process specific details
int pid_ix;
process_data_p p = NULL;
if ((pid > 0) && ((pid_ix = process_hash_lookup(pid)) >= 0)) {
p = &process_hash_table[pid_ix];
- // Quick rejection if this process has interleaved memory, but recheck it once an hour...
-#define MIN_DELAY_FOR_INTERLEAVE (3600 * ONE_HUNDRED)
- if (((p->flags & PROCESS_FLAG_INTERLEAVED) > 0)
- && (p->bind_time_stamp + MIN_DELAY_FOR_INTERLEAVE > time_stamp)) {
- if (log_level >= LOG_DEBUG) {
- numad_log(LOG_DEBUG, "Skipping evaluation because of interleaved memory.\n");
- }
- return NULL;
- }
- // Get cpuset name for this process, and existing mems binding, if any.
+ // Add up per-node memory in use by this process.
+ // This scanning is expensive and should be minimized.
char fname[FNAME_SIZE];
- snprintf(fname, FNAME_SIZE, "/proc/%d/cpuset", pid);
- FILE *fs = fopen(fname, "r");
- if (!fs) {
- numad_log(LOG_WARNING, "Tried to research PID %d cpuset, but it apparently went away.\n", p->pid);
- return NULL; // Assume the process terminated?
- }
- if (!fgets(buf, BUF_SIZE, fs)) {
- numad_log(LOG_WARNING, "Tried to research PID %d cpuset, but it apparently went away.\n", p->pid);
- fclose(fs);
- return NULL; // Assume the process terminated?
- }
- fclose(fs);
- ELIM_NEW_LINE(buf);
- if ((!p->cpuset_name) || (strcmp(p->cpuset_name, buf))) {
- if (p->cpuset_name != NULL) {
- free(p->cpuset_name);
- }
- p->cpuset_name = strdup(buf);
- }
- if (log_level >= LOG_DEBUG) {
- numad_log(LOG_DEBUG, "CPUSET_NAME: %s\n", p->cpuset_name);
- }
- snprintf(fname, FNAME_SIZE, "%s%s/cpuset.mems", cpuset_dir, p->cpuset_name);
- fs = fopen(fname, "r");
- if ((fs) && (fgets(buf, BUF_SIZE, fs))) {
- fclose(fs);
- num_existing_mems = add_ids_to_list_from_str(existing_mems_list_p, buf);
- if (log_level >= LOG_DEBUG) {
- str_from_id_list(buf, BUF_SIZE, existing_mems_list_p);
- numad_log(LOG_DEBUG, "EXISTING CPUSET NODE LIST: %s\n", buf);
- }
- }
- // If this process was just recently bound, enforce a minimum delay
- // period between repeated attempts to potentially move the memory.
- // FIXME: ?? might this retard appropriate process expansion too much?
-#define MIN_DELAY_FOR_REEVALUATION (30 * ONE_HUNDRED)
- if (p->bind_time_stamp + MIN_DELAY_FOR_REEVALUATION > time_stamp) {
- // Skip re-evaluation because we just did it recently.
- if (log_level >= LOG_DEBUG) {
- numad_log(LOG_DEBUG, "Skipping evaluation because done too recently.\n");
- }
- return NULL;
- }
- // Look for short cut because of duplicate bindings. If we have bound
- // this process to the same nodes multiple times already, and the load
- // on those nodes still seems acceptable, skip the rest of this and
- // just return NULL to indicate no change needed. FIXME: should figure
- // out what can change that would make a rebinding desirable (e.g. (1)
- // some process gets sub-optimal allocation on busy machine which
- // subsequently becomes less busy leaving disadvantaged process. (2)
- // node load imbalance, (3) any process split across nodes which should
- // fit within a single node.) For now, just expire the dup_bid_count
- // occasionally, which is a reasonably good mitigation.
- // So, check to see if we should decay the dup_bind_count...
-#define DUP_BIND_TIME_OUT (300 * ONE_HUNDRED)
- if ((p->dup_bind_count > 0) && (p->bind_time_stamp + DUP_BIND_TIME_OUT < time_stamp)) {
- p->dup_bind_count -= 1;
- }
- // Now, look for short cut because of duplicate bindings
- if (p->dup_bind_count > 0) {
- int node_id = 0;
- int nodes_have_cpu = 1;
- int nodes_have_ram = 1;
- int n = num_existing_mems;
- int min_resource_pct = 100 - target_utilization;
- if (min_resource_pct < 5) {
- min_resource_pct = 5;
- }
- while (n) {
- if (ID_IS_IN_LIST(node_id, existing_mems_list_p)) {
- nodes_have_cpu &= ((100 * node[node_id].CPUs_free / node[node_id].CPUs_total) >= (min_resource_pct));
- nodes_have_ram &= ((100 * node[node_id].MBs_free / node[node_id].MBs_total) >= (min_resource_pct));
- n -= 1;
- }
- node_id += 1;
- }
- if ((nodes_have_cpu) && (nodes_have_ram)) {
- if (log_level >= LOG_DEBUG) {
- numad_log(LOG_DEBUG, "Skipping evaluation because of repeat binding\n");
- }
- return NULL;
- }
- if (log_level >= LOG_DEBUG) {
- numad_log(LOG_DEBUG, "Evaluated for skipping by repeat binding, but CPUS: %d, RAM: %d\n", nodes_have_cpu, nodes_have_ram);
- }
- }
- // Fourth, add up per-node memory in use by this process. This scanning
- // is expensive and should be minimized. Also, old kernels dismantle
- // transparent huge pages while producing the numa_maps memory
- // information!
- memset(process_MBs, 0, process_MBs_num_nodes * sizeof(uint64_t));
snprintf(fname, FNAME_SIZE, "/proc/%d/numa_maps", pid);
- fs = fopen(fname, "r");
+ FILE *fs = fopen(fname, "r");
if (!fs) {
numad_log(LOG_WARNING, "Tried to research PID %d numamaps, but it apparently went away.\n", p->pid);
return NULL; // Assume the process terminated
}
+ // Allocate and zero per node memory array.
+ // The "+1 node" is for accumulating interleaved memory
+ p->process_MBs = realloc(p->process_MBs, (num_nodes + 1) * sizeof(uint64_t));
+ if (p->process_MBs == NULL) {
+ numad_log(LOG_CRIT, "p->process_MBs realloc failed\n");
+ exit(EXIT_FAILURE);
+ }
+ memset(p->process_MBs, 0, (num_nodes + 1) * sizeof(uint64_t));
int process_has_interleaved_memory = 0;
while (fgets(buf, BUF_SIZE, fs)) {
int interleaved_memory = 0;
uint64_t page_size = page_size_in_bytes;
const char *delimiters = " \n";
- char *p = strtok(buf, delimiters);
- while (p) {
- if (!strncmp(p, "interleave", 10)) {
+ char *str_p = strtok(buf, delimiters);
+ while (str_p) {
+ if (!strncmp(str_p, "interleave", 10)) {
interleaved_memory = 1;
process_has_interleaved_memory = 1;
- } else if (!strcmp(p, "huge")) {
+ } else if (!strcmp(str_p, "huge")) {
page_size = huge_page_size_in_bytes;
- } else if (*p++ == 'N') {
+ } else if (*str_p++ == 'N') {
int node;
uint64_t pages;
- CONVERT_DIGITS_TO_NUM(p, node);
- if (*p++ != '=') {
+ CONVERT_DIGITS_TO_NUM(str_p, node);
+ if (*str_p++ != '=') {
numad_log(LOG_CRIT, "numa_maps node number parse error\n");
exit(EXIT_FAILURE);
}
- CONVERT_DIGITS_TO_NUM(p, pages);
- process_MBs[node] += (pages * page_size);
+ CONVERT_DIGITS_TO_NUM(str_p, pages);
+ p->process_MBs[node] += (pages * page_size);
if (interleaved_memory) {
// sum interleaved quantity in "extra node"
- process_MBs[num_nodes] += (pages * page_size);
+ p->process_MBs[num_nodes] += (pages * page_size);
}
}
// Get next token on the line
- p = strtok(NULL, delimiters);
+ str_p = strtok(NULL, delimiters);
}
}
fclose(fs);
+ proc_avg_node_CPUs_free = p->CPUs_used;
for (int ix = 0; (ix <= num_nodes); ix++) {
- process_MBs[ix] /= MEGABYTE;
- if (log_level >= LOG_DEBUG) {
- numad_log(LOG_DEBUG, "PROCESS_MBs[%d]: %ld\n", ix, process_MBs[ix]);
+ p->process_MBs[ix] /= MEGABYTE;
+ if ((log_level >= LOG_DEBUG) && (p->process_MBs[ix] > 0)) {
+ if (ix == num_nodes) {
+ numad_log(LOG_DEBUG, "Interleaved MBs: %ld\n", ix, p->process_MBs[ix]);
+ } else {
+ numad_log(LOG_DEBUG, "PROCESS_MBs[%d]: %ld\n", ix, p->process_MBs[ix]);
+ }
+ }
+ if (ID_IS_IN_LIST(ix, p->node_list_p)) {
+ proc_avg_node_CPUs_free += node[ix].CPUs_free;
}
}
+ proc_avg_node_CPUs_free /= NUM_IDS_IN_LIST(p->node_list_p);
if ((process_has_interleaved_memory) && (keep_interleaved_memory)) {
// Mark this process as having interleaved memory so we do not
- // merge the interleaved memory. Time stamp it as done.
+ // merge the interleaved memory. Time stamp it as done and return.
p->flags |= PROCESS_FLAG_INTERLEAVED;
p->bind_time_stamp = get_time_stamp();
if (log_level >= LOG_DEBUG) {
- numad_log(LOG_DEBUG, "Skipping evaluation because of interleaved memory.\n");
+ numad_log(LOG_DEBUG, "Skipping evaluation of PID %d because of interleaved memory.\n", p->pid);
}
return NULL;
}
} // end of existing PID conditional
// Make a copy of node available resources array. Add in info specific to
// this process to equalize available resource quantities wrt locations of
- // resources already in use by this process. Inflate the value of already
- // assigned memory by approximately 3/2, because moving memory is
- // expensive. Average the amount of CPUs_free across the existing nodes
- // used, because the threads are free to move around in that domain. After
- // calculating combined magnitude of available resources, bias the values
- // towards existing locations for this process.
- int target_using_all_nodes = 0;
- uint64_t node_CPUs_free_for_this_process = 0;
- memcpy(tmp_node, node, num_nodes * sizeof(node_data_t) );
- if (num_existing_mems > 0) {
- node_CPUs_free_for_this_process = cpus; // ?? Correct for utilization target inflation?
- int node_id = 0;
- int n = num_existing_mems;
- while (n) {
- if (ID_IS_IN_LIST(node_id, existing_mems_list_p)) {
- node_CPUs_free_for_this_process += tmp_node[node_id].CPUs_free;
- n -= 1;
- }
- node_id += 1;
- }
- // Divide to get average CPUs_free for the nodes in use by process
- node_CPUs_free_for_this_process /= num_existing_mems;
+ // resources already in use by this process.
+ static node_data_p tmp_node;
+ tmp_node = realloc(tmp_node, num_nodes * sizeof(node_data_t) );
+ if (tmp_node == NULL) {
+ numad_log(LOG_CRIT, "tmp_node realloc failed\n");
+ exit(EXIT_FAILURE);
}
+ memcpy(tmp_node, node, num_nodes * sizeof(node_data_t) );
+ uint64_t sum_of_node_CPUs_free = 0;
for (int ix = 0; (ix < num_nodes); ix++) {
if (pid > 0) {
- tmp_node[ix].MBs_free += ((process_MBs[ix] * 12) / 8);
- }
- if ((num_existing_mems > 0) && (ID_IS_IN_LIST(ix, existing_mems_list_p))) {
- tmp_node[ix].CPUs_free = node_CPUs_free_for_this_process;
- }
- if (tmp_node[ix].CPUs_free > tmp_node[ix].CPUs_total) {
- tmp_node[ix].CPUs_free = tmp_node[ix].CPUs_total;
- }
- if (log_level >= LOG_DEBUG) {
- numad_log(LOG_DEBUG, "PROCESS_CPUs[%d]: %ld\n", ix, tmp_node[ix].CPUs_free);
+ if (NUM_IDS_IN_LIST(p->node_list_p) >= num_nodes) {
+ // Process not yet bound to a subset of nodes.
+ // Add back memory used by this process on this node.
+ tmp_node[ix].MBs_free += ((p->process_MBs[ix] * 17) / 16); // Apply light mem bias
+ // Add back CPU used by this process in proportion to the memory used on this node.
+ tmp_node[ix].CPUs_free += ((p->CPUs_used * p->process_MBs[ix]) / p->MBs_used);
+ } else {
+ // If the process is currently running on less than all the
+ // nodes, first add back (biased) memory already used by this
+ // process on this node, then assign average process CPU / node
+ // for this process iff the process is present on this node.
+ tmp_node[ix].MBs_free += ((p->process_MBs[ix] * 5) / 4); // Apply heavy mem bias
+ if (ID_IS_IN_LIST(ix, p->node_list_p)) {
+ tmp_node[ix].CPUs_free = proc_avg_node_CPUs_free;
+ }
+ }
+ sum_of_node_CPUs_free += tmp_node[ix].CPUs_free;
+ if (tmp_node[ix].CPUs_free > tmp_node[ix].CPUs_total) {
+ tmp_node[ix].CPUs_free = tmp_node[ix].CPUs_total;
+ }
+ if (tmp_node[ix].MBs_free > tmp_node[ix].MBs_total) {
+ tmp_node[ix].MBs_free = tmp_node[ix].MBs_total;
+ }
}
- // Calculate magnitude as product of available CPUs and available MBs
- tmp_node[ix].magnitude = tmp_node[ix].CPUs_free * tmp_node[ix].MBs_free;
- // Bias combined magnitude towards already assigned nodes
- if (ID_IS_IN_LIST(ix, existing_mems_list_p)) {
- tmp_node[ix].magnitude *= 9;
- tmp_node[ix].magnitude /= 8;
+ // Enforce 1/100th CPU minimum
+ if (tmp_node[ix].CPUs_free < 1) {
+ tmp_node[ix].CPUs_free = 1;
}
- // Save the current magnitudes
- saved_magnitude_for_node[ix] = tmp_node[ix].magnitude;
+ // numad_log(LOG_DEBUG, "Raw Node[%d]: mem: %ld cpu: %ld\n", ix, tmp_node[ix].MBs_free, tmp_node[ix].CPUs_free);
+ tmp_node[ix].magnitude = combined_value_of_weighted_resources(ix, mbs, cpus, tmp_node[ix].MBs_free, tmp_node[ix].CPUs_free);
}
- // OK, figure out where to get resources for this request.
+ // Now figure out where to get resources for this request....
static id_list_p target_node_list_p;
- CLEAR_LIST(target_node_list_p);
- int prev_node_used = -1;
- // Continue to allocate more resources until request are met.
- // OK if not not quite all the CPU request is met.
- // FIXME: ?? Is half of the utilization margin a good amount of CPU flexing?
- int cpu_flex = ((100 - target_utilization) * tmp_node[0].CPUs_total) / 200;
- if (pid <= 0) {
- // If trying to find resources for pre-placement advice request, do not
- // underestimate the amount of CPUs needed. Instead, err on the side
- // of providing too many resources. So, no flexing here...
- cpu_flex = 0;
- }
- while ((mbs > 0) || (cpus > cpu_flex)) {
- if (log_level >= LOG_DEBUG) {
- numad_log(LOG_DEBUG, "MBs: %d, CPUs: %d\n", mbs, cpus);
+ CLEAR_NODE_LIST(target_node_list_p);
+ if ((pid > 0) && (cpus > sum_of_node_CPUs_free)) {
+ // System CPUs might be oversubscribed, but...
+ assume_enough_cpus = 1;
+ // and rely on available memory for placement.
+ }
+ // Establish a CPU flex fudge factor, on the presumption it is OK if not
+ // quite all the CPU request is met. However, if trying to find resources
+ // for pre-placement advice request, do not underestimate the amount of
+ // CPUs needed. Instead, err on the side of providing too many resources.
+ int cpu_flex = 0;
+ if ((pid > 0) && (target_utilization < 100)) {
+ // FIXME: Is half of the utilization margin a good amount of CPU flexing?
+ cpu_flex = ((100 - target_utilization) * node[0].CPUs_total) / 200;
+ }
+ // Figure out minimum number of nodes required
+ int mem_req_nodes = ceil((double)mbs / (double)node[0].MBs_total);
+ int cpu_req_nodes = ceil((double)(cpus - cpu_flex) / (double)node[0].CPUs_total);
+ int min_req_nodes = mem_req_nodes;
+ if (min_req_nodes < cpu_req_nodes) {
+ min_req_nodes = cpu_req_nodes;
+ }
+ if (min_req_nodes > num_nodes) {
+ min_req_nodes = num_nodes;
+ }
+ // Use an index to sort NUMA connected resource chain for each node
+ int index[num_nodes];
+ uint64_t totmag[num_nodes];
+ for (int ix = 0; (ix < num_nodes); ix++) {
+ // Reset the index each time
+ for (int n = 0; (n < num_nodes); n++) {
+ index[n] = n;
}
- // Sort nodes by magnitude of available resources. Note that
- // inter-node distances (to the previous node used) are factored into
- // the sort.
+ // Sort by minimum relative NUMA distance from node[ix],
+ // breaking distance ties with magnitude of available resources
for (int ij = 0; (ij < num_nodes); ij++) {
- int big_ix = ij;
+ int best_ix = ij;
for (int ik = ij + 1; (ik < num_nodes); ik++) {
- uint64_t ik_dist = 1;
- uint64_t big_ix_dist = 1;
- if (prev_node_used >= 0) {
- ik_dist = tmp_node[ik].distance[prev_node_used];
- big_ix_dist = tmp_node[big_ix].distance[prev_node_used];
- }
- // Scale magnitude comparison by distances to previous node used...
- if ((tmp_node[big_ix].magnitude / big_ix_dist) < (tmp_node[ik].magnitude / ik_dist)) {
- big_ix = ik;
- }
- }
- if (big_ix != ij) {
- node_data_t tmp;
- memcpy((void *)&tmp, (void *)&tmp_node[ij], sizeof(node_data_t) );
- memcpy((void *)&tmp_node[ij], (void *)&tmp_node[big_ix], sizeof(node_data_t) );
- memcpy((void *)&tmp_node[big_ix], (void *)&tmp, sizeof(node_data_t) );
+ int ik_dist = tmp_node[index[ik]].distance[ix];
+ int best_ix_dist = tmp_node[index[best_ix]].distance[ix];
+ if (best_ix_dist > ik_dist) {
+ best_ix = ik;
+ } else if (best_ix_dist == ik_dist) {
+ if (tmp_node[index[best_ix]].magnitude < tmp_node[index[ik]].magnitude ) {
+ best_ix = ik;
+ }
+ }
+ }
+ if (best_ix != ij) {
+ int tmp = index[ij];
+ index[ij] = index[best_ix];
+ index[best_ix] = tmp;
}
}
+#if 0
if (log_level >= LOG_DEBUG) {
- for (int ix = 0; (ix < num_nodes); ix++) {
- numad_log(LOG_DEBUG, "Sorted magnitude[%d]: %ld\n", tmp_node[ix].node_id, tmp_node[ix].magnitude);
+ for (int iq = 0; (iq < num_nodes); iq++) {
+ numad_log(LOG_DEBUG, "Node: %d Dist: %d Magnitude: %ld\n",
+ tmp_node[index[iq]].node_id, tmp_node[index[iq]].distance[ix], tmp_node[index[iq]].magnitude);
+ }
+ }
+#endif
+ // Save the totmag[] sum of the magnitudes of expected needed nodes,
+ // "normalized" by NUMA distance (by dividing each magnitude by the
+ // relative distance squared).
+ totmag[ix] = 0;
+ for (int ij = 0; (ij < min_req_nodes); ij++) {
+ int dist = tmp_node[index[ij]].distance[ix];
+ totmag[ix] += (tmp_node[index[ij]].magnitude / (dist * dist));
+ }
+ numad_log(LOG_DEBUG, "Totmag[%d]: %ld\n", ix, totmag[ix]);
+ }
+ // Now find the best NUMA node based on the normalized sum of node
+ // magnitudes expected to be used.
+ int best_node_ix = 0;
+ for (int ix = 0; (ix < num_nodes); ix++) {
+ if (totmag[best_node_ix] < totmag[ix]) {
+ best_node_ix = ix;
+ }
+ }
+ numad_log(LOG_DEBUG, "best_node_ix: %d\n", best_node_ix);
+ // Reset sorting index again
+ for (int n = 0; (n < num_nodes); n++) {
+ index[n] = n;
+ }
+ // Sort index by distance from node[best_node_ix],
+ // breaking distance ties with magnitude
+ for (int ij = 0; (ij < num_nodes); ij++) {
+ int best_ix = ij;
+ for (int ik = ij + 1; (ik < num_nodes); ik++) {
+ int ik_dist = tmp_node[index[ik]].distance[best_node_ix];
+ int best_ix_dist = tmp_node[index[best_ix]].distance[best_node_ix];
+ if (best_ix_dist > ik_dist) {
+ best_ix = ik;
+ } else if (best_ix_dist == ik_dist) {
+ if (tmp_node[index[best_ix]].magnitude < tmp_node[index[ik]].magnitude ) {
+ best_ix = ik;
+ }
}
}
- if (tmp_node[0].node_id == prev_node_used) {
- // Hmmm. Looks like the best node for more resources, is also the
- // last one we used. This is not going to make progress... So
- // just punt and use everything.
- OR_LISTS(target_node_list_p, target_node_list_p, all_nodes_list_p);
- target_using_all_nodes = 1;
- break;
+ if (best_ix != ij) {
+ int tmp = index[ij];
+ index[ij] = index[best_ix];
+ index[best_ix] = tmp;
+ }
+ }
+ if (log_level >= LOG_DEBUG) {
+ for (int iq = 0; (iq < num_nodes); iq++) {
+ numad_log(LOG_DEBUG, "Node: %d Dist: %d Magnitude: %ld\n",
+ tmp_node[index[iq]].node_id, tmp_node[index[iq]].distance[best_node_ix], tmp_node[index[iq]].magnitude);
}
- prev_node_used = tmp_node[0].node_id;
- ADD_ID_TO_LIST(tmp_node[0].node_id, target_node_list_p);
+ }
+ // Allocate more resources until request is met.
+ best_node_ix = 0;
+ while ((min_req_nodes > 0) || (mbs > 0) || ((cpus > cpu_flex) && (!assume_enough_cpus))) {
if (log_level >= LOG_DEBUG) {
- str_from_id_list(buf, BUF_SIZE, existing_mems_list_p);
- str_from_id_list(buf2, BUF_SIZE, target_node_list_p);
- numad_log(LOG_DEBUG, "Existing nodes: %s Target nodes: %s\n", buf, buf2);
+ numad_log(LOG_DEBUG, "MBs: %d, CPUs: %d\n", mbs, cpus);
}
+ numad_log(LOG_DEBUG, "Assigning resources from node %d\n", index[best_node_ix]);
+ ADD_ID_TO_LIST(tmp_node[index[best_node_ix]].node_id, target_node_list_p);
+ min_req_nodes -= 1;
if (EQUAL_LISTS(target_node_list_p, all_nodes_list_p)) {
// Apparently we must use all resource nodes...
- target_using_all_nodes = 1;
break;
}
-#define MBS_MARGIN 10
- if (tmp_node[0].MBs_free >= (mbs + MBS_MARGIN)) {
- tmp_node[0].MBs_free -= mbs;
+ // "Consume" the resources on this node
+#define CPUS_MARGIN 0
+#define MBS_MARGIN 100
+ if (tmp_node[index[best_node_ix]].MBs_free >= (mbs + MBS_MARGIN)) {
+ tmp_node[index[best_node_ix]].MBs_free -= mbs;
mbs = 0;
} else {
- mbs -= (tmp_node[0].MBs_free - MBS_MARGIN);
- tmp_node[0].MBs_free = MBS_MARGIN;
+ mbs -= (tmp_node[index[best_node_ix]].MBs_free - MBS_MARGIN);
+ tmp_node[index[best_node_ix]].MBs_free = MBS_MARGIN;
}
-#define CPUS_MARGIN 0
- if (tmp_node[0].CPUs_free >= (cpus + CPUS_MARGIN)) {
- tmp_node[0].CPUs_free -= cpus;
+ if (tmp_node[index[best_node_ix]].CPUs_free >= (cpus + CPUS_MARGIN)) {
+ tmp_node[index[best_node_ix]].CPUs_free -= cpus;
cpus = 0;
} else {
- cpus -= (tmp_node[0].CPUs_free - CPUS_MARGIN);
- tmp_node[0].CPUs_free = CPUS_MARGIN;
- }
- tmp_node[0].magnitude = tmp_node[0].CPUs_free * tmp_node[0].MBs_free;
- }
- // If this existing process is already located where we want it, and almost
- // all memory is already moved to those nodes, then return NULL indicating
- // no need to change binding this time.
- if ((pid > 0) && (EQUAL_LISTS(target_node_list_p, existing_mems_list_p))) {
- // May not need to change binding. However, if there is any significant
- // memory still on non-target nodes, advise the bind anyway because
- // there are some scenarios when the kernel will not move it all the
- // first time.
- if (!target_using_all_nodes) {
- p->dup_bind_count += 1;
- for (int ix = 0; (ix < num_nodes); ix++) {
- if ((process_MBs[ix] > 10) && (!ID_IS_IN_LIST(ix, target_node_list_p))) {
- goto try_memory_move_again;
- }
- }
- // We will accept these memory locations. Stamp it as done.
- p->bind_time_stamp = get_time_stamp();
- }
- // Skip rebinding either because practically all memory is in the
- // target nodes, or because we are stuck using all the nodes.
- if (log_level >= LOG_DEBUG) {
- numad_log(LOG_DEBUG, "Skipping evaluation because memory is reasonably situated.\n");
+ cpus -= (tmp_node[index[best_node_ix]].CPUs_free - CPUS_MARGIN);
+ tmp_node[index[best_node_ix]].CPUs_free = CPUS_MARGIN;
}
- return NULL;
- } else {
- // Either a non-existing process, or a new binding for an existing process.
- if (p != NULL) {
- // Must be a new binding for an existing process, so reset dup_bind_count.
- p->dup_bind_count = 0;
- }
- }
- // See if this proposed move will make a significant difference.
- // If not, return null instead of advising the move.
- uint64_t target_magnitude = 0;
- uint64_t existing_magnitude = 0;
- int num_target_nodes = NUM_IDS_IN_LIST(target_node_list_p);
- int num_existing_nodes = NUM_IDS_IN_LIST(existing_mems_list_p);
- /* FIXME: this expansion seems to cause excessive growth
- * So calculate the improvement before hastily expanding nodes.
- if (num_target_nodes > num_existing_nodes) { goto try_memory_move_again; }
- */
- int node_id = 0;
- int n = num_existing_nodes + num_target_nodes;
- while (n) {
- if (ID_IS_IN_LIST(node_id, target_node_list_p)) {
- target_magnitude += saved_magnitude_for_node[node_id];
- n -= 1;
- }
- if (ID_IS_IN_LIST(node_id, existing_mems_list_p)) {
- existing_magnitude += saved_magnitude_for_node[node_id];
- n -= 1;
- }
- node_id += 1;
- }
- if (existing_magnitude > 0) {
- uint64_t magnitude_change = ((target_magnitude - existing_magnitude) * 100) / existing_magnitude;
- if (magnitude_change < 0) {
- magnitude_change = -(magnitude_change);
- }
- if (magnitude_change <= IMPROVEMENT_THRESHOLD_PERCENT) {
- // Not significant enough percentage change to do rebind
+ // Next line optional, since we will not look at that node again
+ tmp_node[index[best_node_ix]].magnitude = combined_value_of_weighted_resources(0, mbs, cpus, tmp_node[index[best_node_ix]].MBs_free, tmp_node[index[best_node_ix]].CPUs_free);
+ best_node_ix += 1;
+ }
+ // For existing processes, calculate the non-local memory percent to see if
+ // process is already in the right place.
+ if ((pid > 0) && (p != NULL)) {
+ uint64_t nonlocal_memory = 0;
+ for (int ix = 0; (ix < num_nodes); ix++) {
+ if (!ID_IS_IN_LIST(ix, target_node_list_p)) {
+ // Accumulate total of nonlocal memory
+ nonlocal_memory += p->process_MBs[ix];
+ }
+ }
+ int disp_percent = (100 * nonlocal_memory) / p->MBs_used;
+ // If this existing process is already located where we want it, then just
+ // return NULL indicating no need to change binding this time. Check the
+ // ammount of nonlocal memory against the target_memlocality_perecent.
+ if ((disp_percent <= (100 - target_memlocality)) && (p->bind_time_stamp) && (EQUAL_LISTS(target_node_list_p, p->node_list_p))) {
+ // Already bound to targets, and enough of the memory is located where we want it, so no need to rebind
if (log_level >= LOG_DEBUG) {
- str_from_id_list(buf, BUF_SIZE, existing_mems_list_p);
- str_from_id_list(buf2, BUF_SIZE, target_node_list_p);
- numad_log(LOG_DEBUG, "Moving pid %d from nodes (%s) to nodes (%s) skipped as insignificant improvement: %ld percent.\n",
- pid, buf, buf2, magnitude_change);
+ numad_log(LOG_DEBUG, "Process %d already %d percent localized to target nodes.\n", p->pid, 100 - disp_percent);
}
- // We decided this is almost good enough. Stamp it as done.
p->bind_time_stamp = get_time_stamp();
return NULL;
}
}
- if ((pid <= 0) && (num_target_nodes <= 0)) {
- // Always provide at least one node for pre-placement advice
+ // Must always provide at least one node for pre-placement advice
+ // FIXME: verify this can happen only if no resources requested...
+ if ((pid <= 0) && (NUM_IDS_IN_LIST(target_node_list_p) <= 0)) {
ADD_ID_TO_LIST(node[0].node_id, target_node_list_p);
}
-try_memory_move_again:
- str_from_id_list(buf, BUF_SIZE, existing_mems_list_p);
+ // Log advice, and return target node list
+ if ((pid > 0) && (p->bind_time_stamp)) {
+ str_from_id_list(buf, BUF_SIZE, p->node_list_p);
+ } else {
+ str_from_id_list(buf, BUF_SIZE, all_nodes_list_p);
+ }
+ char buf2[BUF_SIZE];
str_from_id_list(buf2, BUF_SIZE, target_node_list_p);
char *cmd_name = "(unknown)";
if ((p) && (p->comm)) {
cmd_name = p->comm;
}
numad_log(LOG_NOTICE, "Advising pid %d %s move from nodes (%s) to nodes (%s)\n", pid, cmd_name, buf, buf2);
+ if (pid > 0) {
+ COPY_LIST(target_node_list_p, p->node_list_p);
+ }
return target_node_list_p;
}
-
-void show_processes(process_data_p *ptr, int nprocs) {
- time_t ts = time(NULL);
- fprintf(log_fs, "%s", ctime(&ts));
- fprintf(log_fs, "Candidates: %d\n", nprocs);
- for (int ix = 0; (ix < nprocs); ix++) {
- process_data_p p = ptr[ix];
- char buf[BUF_SIZE];
- snprintf(buf, BUF_SIZE, "%s%s/cpuset.mems", cpuset_dir, p->cpuset_name);
- FILE *fs = fopen(buf, "r");
- buf[0] = '\0';
- if (fs) {
- if (fgets(buf, BUF_SIZE, fs)) {
- ELIM_NEW_LINE(buf);
- }
- fclose(fs);
- }
- fprintf(log_fs, "%ld: PID %d: %s, Threads %2ld, MBs_used %6ld, CPUs_used %4ld, Magnitude %6ld, Nodes: %s\n",
- p->data_time_stamp, p->pid, p->comm, p->num_threads, p->MBs_used, p->CPUs_used, p->MBs_used * p->CPUs_used, buf);
- }
- fprintf(log_fs, "\n");
- fflush(log_fs);
-}
-
-
-
int manage_loads() {
+ uint64_t time_stamp = get_time_stamp();
// Use temporary index to access and sort hash table entries
- static process_data_p *pindex;
static int pindex_size;
+ static process_data_p *pindex;
if (pindex_size < process_hash_table_size) {
pindex_size = process_hash_table_size;
pindex = realloc(pindex, pindex_size * sizeof(process_data_p));
@@ -1923,19 +2049,54 @@ int manage_loads() {
return min_interval / 2;
}
memset(pindex, 0, pindex_size * sizeof(process_data_p));
- // Copy live candidate pointers to the index for sorting, etc
+ // Copy live candidate pointers to the index for sorting
+ // if they meet the threshold for memory usage and CPU usage.
int nprocs = 0;
+ long sum_CPUs_used = 0;
for (int ix = 0; (ix < process_hash_table_size); ix++) {
process_data_p p = &process_hash_table[ix];
- if (p->pid) {
+ if ((p->pid) && (p->CPUs_used > CPU_THRESHOLD) && (p->MBs_used > MEMORY_THRESHOLD)) {
pindex[nprocs++] = p;
+ sum_CPUs_used += p->CPUs_used;
+ // Initialize node list, if not already done for this process.
+ if (p->node_list_p == NULL) {
+ initialize_mem_node_list(p);
+ }
}
}
- // Sort index by amount of CPU used * amount of memory used. Not expecting
- // a long list here. Use a simple sort -- however, sort into bins,
- // treating values within 10% as aquivalent. Within bins, order by
- // bind_time_stamp so oldest bound will be higher priority to evaluate.
+ // Order candidate considerations using timestamps and magnitude: amount of
+ // CPU used * amount of memory used. Not expecting a long list here. Use
+ // a simplistic sort -- however move all not yet bound to front of list and
+ // order by decreasing magnitude. Previously bound processes follow in
+ // bins of increasing magnitude treating values within 20% as aquivalent.
+ // Within bins, order by bind_time_stamp so oldest bound will be higher
+ // priority to evaluate. Start by moving all unbound to beginning.
+ int num_unbound = 0;
for (int ij = 0; (ij < nprocs); ij++) {
+ if (pindex[ij]->bind_time_stamp == 0) {
+ process_data_p tmp = pindex[num_unbound];
+ pindex[num_unbound++] = pindex[ij];
+ pindex[ij] = tmp;
+ }
+ }
+ // Sort all unbound so biggest magnitude comes first
+ for (int ij = 0; (ij < num_unbound); ij++) {
+ int best = ij;
+ for (int ik = ij + 1; (ik < num_unbound); ik++) {
+ uint64_t ik_mag = (pindex[ ik]->CPUs_used * pindex[ ik]->MBs_used);
+ uint64_t best_mag = (pindex[best]->CPUs_used * pindex[best]->MBs_used);
+ if (ik_mag <= best_mag) continue;
+ best = ik;
+ }
+ if (best != ij) {
+ process_data_p tmp = pindex[ij];
+ pindex[ij] = pindex[best];
+ pindex[best] = tmp;
+ }
+ }
+ // Sort the remaining candidates into bins of increasting magnitude, and by
+ // timestamp within bins.
+ for (int ij = num_unbound; (ij < nprocs); ij++) {
int best = ij;
for (int ik = ij + 1; (ik < nprocs); ik++) {
uint64_t ik_mag = (pindex[ ik]->CPUs_used * pindex[ ik]->MBs_used);
@@ -1946,11 +2107,11 @@ int manage_loads() {
diff_mag = -(diff_mag);
min_mag = best_mag;
}
- if ((diff_mag > 0) && (min_mag / diff_mag < 10)) {
- // difference > 10 percent. Use strict ordering
+ if ((diff_mag > 0) && (min_mag / diff_mag < 5)) {
+ // difference > 20 percent. Use magnitude ordering
if (ik_mag <= best_mag) continue;
} else {
- // difference within 10 percent. Sort these by bind_time_stamp.
+ // difference within 20 percent. Sort these by bind_time_stamp.
if (pindex[ik]->bind_time_stamp > pindex[best]->bind_time_stamp) continue;
}
best = ik;
@@ -1961,23 +2122,57 @@ int manage_loads() {
pindex[best] = tmp;
}
}
+ // Show the candidate processes in the log file
if ((log_level >= LOG_INFO) && (nprocs > 0)) {
- show_processes(pindex, nprocs);
+ numad_log(LOG_INFO, "Candidates: %d\n", nprocs);
+ for (int ix = 0; (ix < nprocs); ix++) {
+ process_data_p p = pindex[ix];
+ char buf[BUF_SIZE];
+ str_from_id_list(buf, BUF_SIZE, p->node_list_p);
+ fprintf(log_fs, "%ld: PID %d: %s, Threads %2ld, MBs_size %6ld, MBs_used %6ld, CPUs_used %4ld, Magnitude %6ld, Nodes: %s\n",
+ p->data_time_stamp, p->pid, p->comm, p->num_threads, p->MBs_size, p->MBs_used, p->CPUs_used, p->MBs_used * p->CPUs_used, buf);
+ }
+ fflush(log_fs);
}
- // Estimate desired size and make resource requests for each significant process
+ // Estimate desired size (+ margin capacity) and
+ // make resource requests for each candidate process
for (int ix = 0; (ix < nprocs); ix++) {
process_data_p p = pindex[ix];
- if (p->CPUs_used * p->MBs_used < CPU_THRESHOLD * MEMORY_THRESHOLD) {
- break; // No more significant processes worth worrying about...
+ // If this process has interleaved memory, recheck it only every 30 minutes...
+#define MIN_DELAY_FOR_INTERLEAVE (1800 * ONE_HUNDRED)
+ if (((p->flags & PROCESS_FLAG_INTERLEAVED) > 0)
+ && (p->bind_time_stamp + MIN_DELAY_FOR_INTERLEAVE > time_stamp)) {
+ if (log_level >= LOG_DEBUG) {
+ numad_log(LOG_DEBUG, "Skipping evaluation of PID %d because of interleaved memory.\n", p->pid);
+ }
+ continue;
+ }
+ // Expand resources needed estimate using target_utilization factor.
+ // Start with the CPUs actually used (capped by number of threads) for
+ // CPUs required, and the RSS MBs actually used for the MBs
+ // requirement,
+ int mem_target_utilization = target_utilization;
+ int cpu_target_utilization = target_utilization;
+ // Cap memory utilization at 100 percent (but allow CPUs to oversubscribe)
+ if (mem_target_utilization > 100) {
+ mem_target_utilization = 100;
+ }
+ // If the process virtual memory size is bigger than one node, and it
+ // is already using more than 80 percent of a node, then request MBs
+ // based on the virtual size rather than on the current amount in use.
+ int mb_request;
+ if ((p->MBs_size > node[0].MBs_total) && ((p->MBs_used * 5 / 4) > node[0].MBs_total)) {
+ mb_request = (p->MBs_size * 100) / mem_target_utilization;
+ } else {
+ mb_request = (p->MBs_used * 100) / mem_target_utilization;
}
- int mb_request = (p->MBs_used * 100) / target_utilization;
- int cpu_request = (p->CPUs_used * 100) / target_utilization;
- // Do not give a process more CPUs than it has threads!
- // FIXME: For guest VMs, should limit max to VCPU threads. Will
- // need to do something more intelligent with guest IO threads
- // when eventually considering devices and IRQs.
+ int cpu_request = (p->CPUs_used * 100) / cpu_target_utilization;
+ // But do not give a process more CPUs than it has threads!
int thread_limit = p->num_threads;
- // If process looks like a KVM guest, try to limit to number of vCPU threads
+ // If process looks like a KVM guest, try to limit thread count to the
+ // number of vCPU threads. FIXME: Will need to do something more
+ // intelligent than this with guest IO threads when eventually
+ // considering devices and IRQs.
if ((p->comm) && (p->comm[0] == '(') && (p->comm[1] == 'q') && (strcmp(p->comm, "(qemu-kvm)") == 0)) {
int kvm_vcpu_threads = get_num_kvm_vcpu_threads(p->pid);
if (thread_limit > kvm_vcpu_threads) {
@@ -1988,23 +2183,51 @@ int manage_loads() {
if (cpu_request > thread_limit) {
cpu_request = thread_limit;
}
+ // If this process was recently bound, enforce a five-minute minimum
+ // delay between repeated attempts to potentially move the process.
+#define MIN_DELAY_FOR_REEVALUATION (300 * ONE_HUNDRED)
+ if (p->bind_time_stamp + MIN_DELAY_FOR_REEVALUATION > time_stamp) {
+ // Skip re-evaluation because we just did it recently, but check
+ // first for node utilization balance to see if we should
+ // re-evaluate this particular process right now. If this process
+ // is running on one of the busiest nodes, go ahead and re-evaluate
+ // it if it looks like it should have a better place with
+ // sufficient resources. FIXME: this is currently implemented for
+ // only smallish processes that will fit in a single node.
+ if ( ( ID_IS_IN_LIST(min_node_CPUs_free_ix, p->node_list_p) || ID_IS_IN_LIST(min_node_MBs_free_ix, p->node_list_p))
+ && (cpu_request < node[0].CPUs_total) && (mb_request < node[0].MBs_total)
+ && (abs(min_node_CPUs_free + p->CPUs_used - avg_node_CPUs_free)
+ + abs((max_node_CPUs_free - p->CPUs_used) - avg_node_CPUs_free)
+ < (max_node_CPUs_free - min_node_CPUs_free) - CPU_THRESHOLD) // CPU slop
+ && (abs(min_node_MBs_free + p->MBs_used - avg_node_MBs_free)
+ + abs((max_node_MBs_free - p->MBs_used) - avg_node_MBs_free)
+ < (max_node_MBs_free - min_node_MBs_free)) ) {
+ if (log_level >= LOG_DEBUG) {
+ numad_log(LOG_DEBUG, "Bypassing delay for %d because it looks like it can do better.\n", p->pid);
+ }
+ } else {
+ if (log_level >= LOG_DEBUG) {
+ numad_log(LOG_DEBUG, "Skipping evaluation of PID %d because done too recently.\n", p->pid);
+ }
+ continue;
+ }
+ }
+ // OK, now pick NUMA nodes for this process and bind it!
pthread_mutex_lock(&node_info_mutex);
- id_list_p node_list_p = pick_numa_nodes(p->pid, cpu_request, mb_request);
- // FIXME: ?? copy node_list_p to shorten mutex region?
- if ((node_list_p != NULL) && (bind_process_and_migrate_memory(p->pid, p->cpuset_name, node_list_p, NULL))) {
- // Shorten interval if actively moving processes
+ int assume_enough_cpus = (sum_CPUs_used <= sum_CPUs_total);
+ id_list_p node_list_p = pick_numa_nodes(p->pid, cpu_request, mb_request, assume_enough_cpus);
+ if ((node_list_p != NULL) && (bind_process_and_migrate_memory(p))) {
pthread_mutex_unlock(&node_info_mutex);
- p->bind_time_stamp = get_time_stamp();
+ // Return minimum interval when actively moving processes
return min_interval;
}
pthread_mutex_unlock(&node_info_mutex);
}
- // Return maximum interval if no process movement
+ // Return maximum interval when no process movement
return max_interval;
}
-
void *set_dynamic_options(void *arg) {
// int arg_value = *(int *)arg;
char buf[BUF_SIZE];
@@ -2013,6 +2236,18 @@ void *set_dynamic_options(void *arg) {
msg_t msg;
recv_msg(&msg);
switch (msg.body.cmd) {
+ case 'C':
+ use_inactive_file_cache = (msg.body.arg1 != 0);
+ if (use_inactive_file_cache) {
+ numad_log(LOG_NOTICE, "Counting inactive file cache as available\n");
+ } else {
+ numad_log(LOG_NOTICE, "Counting inactive file cache as unavailable\n");
+ }
+ break;
+ case 'H':
+ thp_scan_sleep_ms = msg.body.arg1;
+ set_thp_scan_sleep_ms(thp_scan_sleep_ms);
+ break;
case 'i':
min_interval = msg.body.arg1;
max_interval = msg.body.arg2;
@@ -2033,6 +2268,10 @@ void *set_dynamic_options(void *arg) {
numad_log(LOG_NOTICE, "Changing log level to %d\n", msg.body.arg1);
log_level = msg.body.arg1;
break;
+ case 'm':
+ numad_log(LOG_NOTICE, "Changing target memory locality to %d\n", msg.body.arg1);
+ target_memlocality = msg.body.arg1;
+ break;
case 'p':
numad_log(LOG_NOTICE, "Adding PID %d to inclusion PID list\n", msg.body.arg1);
pthread_mutex_lock(&pid_list_mutex);
@@ -2055,6 +2294,11 @@ void *set_dynamic_options(void *arg) {
numad_log(LOG_NOTICE, "Scanning only explicit PID list processes\n");
}
break;
+ case 't':
+ numad_log(LOG_NOTICE, "Changing logical CPU thread percent to %d\n", msg.body.arg1);
+ htt_percent = msg.body.arg1;
+ node_info_time_stamp = 0; // to force rescan of nodes/cpus soon
+ break;
case 'u':
numad_log(LOG_NOTICE, "Changing target utilization to %d\n", msg.body.arg1);
target_utilization = msg.body.arg1;
@@ -2064,7 +2308,7 @@ void *set_dynamic_options(void *arg) {
msg.body.arg1, msg.body.arg2);
pthread_mutex_lock(&node_info_mutex);
update_nodes();
- id_list_p node_list_p = pick_numa_nodes(-1, msg.body.arg1, msg.body.arg2);
+ id_list_p node_list_p = pick_numa_nodes(-1, msg.body.arg1, msg.body.arg2, 0);
str_from_id_list(buf, BUF_SIZE, node_list_p);
pthread_mutex_unlock(&node_info_mutex);
send_msg(msg.body.src_pid, 'w', 0, 0, buf);
@@ -2134,30 +2378,50 @@ void parse_two_arg_values(char *p, int *
int main(int argc, char *argv[]) {
int opt;
+ int C_flag = 0;
int d_flag = 0;
+ int H_flag = 0;
int i_flag = 0;
int K_flag = 0;
int l_flag = 0;
+ int m_flag = 0;
int p_flag = 0;
int r_flag = 0;
int S_flag = 0;
+ int t_flag = 0;
int u_flag = 0;
int v_flag = 0;
int w_flag = 0;
int x_flag = 0;
+ int tmp_int = 0;
long list_pid = 0;
- while ((opt = getopt(argc, argv, "dD:hi:K:l:p:r:S:u:vVw:x:")) != -1) {
+ while ((opt = getopt(argc, argv, "C:dD:hH:i:K:l:p:r:R:S:t:u:vVw:x:")) != -1) {
switch (opt) {
+ case 'C':
+ C_flag = 1;
+ use_inactive_file_cache = (atoi(optarg) != 0);
+ break;
case 'd':
d_flag = 1;
log_level = LOG_DEBUG;
break;
case 'D':
- cpuset_dir_list[0] = strdup(optarg);
+ // obsoleted
break;
case 'h':
print_usage_and_exit(argv[0]);
break;
+ case 'H':
+ tmp_int = atoi(optarg);
+ if ((tmp_int == 0) || ((tmp_int > 9) && (tmp_int < 1000001))) {
+ // 0 means do not change the system default value
+ H_flag = 1;
+ thp_scan_sleep_ms = tmp_int;
+ } else {
+ fprintf(stderr, "THP scan_sleep_ms must be > 9 and < 1000001\n");
+ exit(EXIT_FAILURE);
+ }
+ break;
case 'i':
i_flag = 1;
parse_two_arg_values(optarg, &min_interval, &max_interval, 1, 0);
@@ -2170,6 +2434,13 @@ int main(int argc, char *argv[]) {
l_flag = 1;
log_level = atoi(optarg);
break;
+ case 'm':
+ tmp_int = atoi(optarg);
+ if ((tmp_int >= 50) && (tmp_int <= 100)) {
+ m_flag = 1;
+ target_memlocality = tmp_int;
+ }
+ break;
case 'p':
p_flag = 1;
list_pid = atol(optarg);
@@ -2183,13 +2454,26 @@ int main(int argc, char *argv[]) {
include_pid_list = remove_pid_from_pid_list(include_pid_list, list_pid);
exclude_pid_list = remove_pid_from_pid_list(exclude_pid_list, list_pid);
break;
+ case 'R':
+ reserved_cpu_str = strdup(optarg);
+ break;
case 'S':
S_flag = 1;
scan_all_processes = (atoi(optarg) != 0);
break;
+ case 't':
+ tmp_int = atoi(optarg);
+ if ((tmp_int >= 0) && (tmp_int <= 100)) {
+ t_flag = 1;
+ htt_percent = tmp_int;
+ }
+ break;
case 'u':
- u_flag = 1;
- target_utilization = atoi(optarg);
+ tmp_int = atoi(optarg);
+ if ((tmp_int >= 10) && (tmp_int <= 130)) {
+ u_flag = 1;
+ target_utilization = tmp_int;
+ }
break;
case 'v':
v_flag = 1;
@@ -2234,6 +2518,12 @@ int main(int argc, char *argv[]) {
// Daemon is already running. So send dynamic options to persistant
// thread to handle requests, get the response (if any), and finish.
msg_t msg;
+ if (C_flag) {
+ send_msg(daemon_pid, 'C', use_inactive_file_cache, 0, "");
+ }
+ if (H_flag) {
+ send_msg(daemon_pid, 'H', thp_scan_sleep_ms, 0, "");
+ }
if (i_flag) {
send_msg(daemon_pid, 'i', min_interval, max_interval, "");
}
@@ -2243,6 +2533,9 @@ int main(int argc, char *argv[]) {
if (d_flag || l_flag || v_flag) {
send_msg(daemon_pid, 'l', log_level, 0, "");
}
+ if (m_flag) {
+ send_msg(daemon_pid, 'm', target_memlocality, 0, "");
+ }
if (p_flag) {
send_msg(daemon_pid, 'p', list_pid, 0, "");
}
@@ -2252,6 +2545,9 @@ int main(int argc, char *argv[]) {
if (S_flag) {
send_msg(daemon_pid, 'S', scan_all_processes, 0, "");
}
+ if (t_flag) {
+ send_msg(daemon_pid, 't', htt_percent, 0, "");
+ }
if (u_flag) {
send_msg(daemon_pid, 'u', target_utilization, 0, "");
}
@@ -2263,14 +2559,30 @@ int main(int argc, char *argv[]) {
if (x_flag) {
send_msg(daemon_pid, 'x', list_pid, 0, "");
}
- } else if (w_flag) {
- // Get pre-placement NUMA advice without starting daemon
+ close_log_file();
+ exit(EXIT_SUCCESS);
+ }
+ // No numad daemon running yet.
+ // First, make note of any reserved CPUs....
+ if (reserved_cpu_str != NULL) {
+ CLEAR_CPU_LIST(reserved_cpu_mask_list_p);
+ int n = add_ids_to_list_from_str(reserved_cpu_mask_list_p, reserved_cpu_str);
char buf[BUF_SIZE];
+ str_from_id_list(buf, BUF_SIZE, reserved_cpu_mask_list_p);
+ numad_log(LOG_NOTICE, "Reserving %d CPUs (%s) for non-numad use\n", n, buf);
+ // turn reserved list into a negated mask for later ANDing use...
+ negate_cpu_list(reserved_cpu_mask_list_p);
+ }
+ // If it is a "-w" pre-placement request, handle that without starting
+ // the daemon. Otherwise start the numad daemon.
+ if (w_flag) {
+ // Get pre-placement NUMA advice without starting daemon
update_nodes();
sleep(2);
update_nodes();
numad_log(LOG_NOTICE, "Getting NUMA pre-placement advice for %d CPUs and %d MBs\n", requested_cpus, requested_mbs);
- id_list_p node_list_p = pick_numa_nodes(-1, requested_cpus, requested_mbs);
+ id_list_p node_list_p = pick_numa_nodes(-1, requested_cpus, requested_mbs, 0);
+ char buf[BUF_SIZE];
str_from_id_list(buf, BUF_SIZE, node_list_p);
fprintf(stdout, "%s\n", buf);
close_log_file();
@@ -2278,6 +2590,7 @@ int main(int argc, char *argv[]) {
} else if (max_interval > 0) {
// Start the numad daemon...
check_prereqs(argv[0]);
+#if (!NO_DAEMON)
// Daemonize self...
daemon_pid = fork();
if (daemon_pid < 0) { numad_log(LOG_CRIT, "fork() failed\n"); exit(EXIT_FAILURE); }
@@ -2298,9 +2611,20 @@ int main(int argc, char *argv[]) {
if (log_fs != stderr) {
fclose(stderr);
}
+#endif
+ // Set up signal handlers
+ struct sigaction sa;
+ memset(&sa, 0, sizeof(sa));
+ sa.sa_handler = sig_handler;
+ if (sigaction(SIGHUP, &sa, NULL)
+ || sigaction(SIGTERM, &sa, NULL)
+ || sigaction(SIGQUIT, &sa, NULL)) {
+ numad_log(LOG_CRIT, "sigaction does not work?\n");
+ exit(EXIT_FAILURE);
+ }
// Allocate initial process hash table
process_hash_table_expand();
- // Spawn thread to handle messages from subsequent invocation requests
+ // Spawn a thread to handle messages from subsequent invocation requests
pthread_mutex_init(&pid_list_mutex, NULL);
pthread_mutex_init(&node_info_mutex, NULL);
pthread_attr_t attr;
@@ -2310,7 +2634,7 @@ int main(int argc, char *argv[]) {
}
pthread_t tid;
if (pthread_create(&tid, &attr, &set_dynamic_options, &tid) != 0) {
- numad_log(LOG_CRIT, "pthread_create failure\n");
+ numad_log(LOG_CRIT, "pthread_create failure: setting thread\n");
exit(EXIT_FAILURE);
}
// Loop here forwever...
@@ -2322,16 +2646,26 @@ int main(int argc, char *argv[]) {
if (nodes > 1) {
update_processes();
interval = manage_loads();
+ if (interval < max_interval) {
+ // Update node info since we moved something
+ nodes = update_nodes();
+ }
}
sleep(interval);
+ if (got_sigterm | got_sigquit) {
+ shut_down_numad();
+ }
+ if (got_sighup) {
+ got_sighup = 0;
+ close_log_file();
+ open_log_file();
+ }
}
if (pthread_attr_destroy(&attr) != 0) {
numad_log(LOG_WARNING, "pthread_attr_destroy failure\n");
}
pthread_mutex_destroy(&pid_list_mutex);
pthread_mutex_destroy(&node_info_mutex);
- } else {
- shut_down_numad();
}
exit(EXIT_SUCCESS);
}
diff -rup numad-0.5git/numad.init numad-0.5git-new/numad.init
--- numad-0.5git/numad.init 2012-12-03 15:40:40.000000000 +0100
+++ numad-0.5git-new/numad.init 2016-08-30 08:45:19.000000000 +0200
@@ -37,7 +37,7 @@ start() {
[ -f $config ] || exit 6
echo -n $"Starting $prog: "
. $config
- daemon "$exec -i $INTERVAL"
+ daemon $exec -i $INTERVAL
retval=$?
echo
[ $retval -eq 0 ] && touch $lockfile