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 .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 #include #include +#include +#include +#include + #include #include #include @@ -49,26 +53,16 @@ Inc., 59 Temple Place, Suite 330, Boston #include #include #include -#include -#include -#include + +#include -#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) cpuset bitmasks, + // Use CPU_SET(3) 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 ...\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 to specify cgroup mount point\n"); fprintf(stderr, "-h to print this usage info\n"); + fprintf(stderr, "-H to set THP scan_sleep_ms (default %d)\n", DEFAULT_THP_SCAN_SLEEP_MS); fprintf(stderr, "-i [:] 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 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 to specify logging level (usually 5, 6, or 7 -- default 5)\n"); + fprintf(stderr, "-m to specify memory locality target percent (default %d)\n", DEFAULT_MEMLOCALITY_PERCENT); fprintf(stderr, "-p to add PID to inclusion pid list\n"); fprintf(stderr, "-r 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 to specify target utilization percent (default 85)\n"); + fprintf(stderr, "-R 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 to specify thread / logical CPU valuation percent (default %d)\n", DEFAULT_HTT_PERCENT); + fprintf(stderr, "-u 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 [:] 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 \n"); - fprintf(stderr, " # mount cgroup -t cgroup -o cpuset \n"); - fprintf(stderr, " where 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 ' 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 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 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" filename present, save 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/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