event synthesization multithreading for perf record
v2 cover event synthesization multithreading for perf record
From: Kan Liang <[email protected]>
The event synthesization multithreading is introduced in ("perf top optimization") https://lkml.org/lkml/2017/9/29/269
But it was not enabled for perf record. Because the process function process_synthesized_event was not multithreading friendly.
The patch series temporarily stores the process result in per-thread file, which make the processing in parallel. Then it dumps the file one by one to the perf.data at the end of event synthesization.
The source code is also available at
https://github.com/kliang2/perf.git perf_record_opt
Usually, the event synthesization only happens once on either start or end. With the snapshotting code, we synthesize events multiple times, once per each new perf.data file. Both of the cases are verified.
Here are the latency test result on Knights Mill and Skylake server
The workload is to compile Linux kernel as below "sudo nice make -j$(grep -c '^processor' /proc/cpuinfo)" Then, "sudo perf record -e cycles -a -- sleep 1"
The latency is the time cost of __machine__synthesize_threads or its multithreading replacement, record__multithread_synthesize.
- Latency on Knights Mill (272 CPUs)
Original(s) With patch(s) Speedup 12.74 5.54 2.3X
- Latency on Skylake server (192 CPUs)
Original(s) With patch(s) Speedup 0.36 0.25 1.47X
Changes since V1:
- Dump the synthesized result to per-thread file and merge them to perf.data at the end. (Arnaldo)
Kan Liang (5):
perf tools: pass thread info to process function perf tools: pass thread info in event synthesization perf tools: expose copyfile_offset()
perf record: synthesize event multithreading support perf record: add option to set the number of thread for event
synthesize
tools/perf/Documentation/perf-record.txt | 4 + tools/perf/arch/x86/util/tsc.c | 2 +- tools/perf/builtin-inject.c | 12 ++- tools/perf/builtin-record.c | 129 ++++++++++++++++++++++++++++--- tools/perf/builtin-sched.c | 12 ++- tools/perf/builtin-stat.c | 3 +- tools/perf/builtin-trace.c | 3 +- tools/perf/tests/cpumap.c | 6 +- tools/perf/tests/dwarf-unwind.c | 6 +- tools/perf/tests/event_update.c | 12 ++- tools/perf/tests/stat.c | 9 ++- tools/perf/tests/thread-map.c | 3 +- tools/perf/util/auxtrace.c | 2 +- tools/perf/util/event.c | 111 ++++++++++++++++---------- tools/perf/util/event.h | 19 +++-- tools/perf/util/header.c | 16 ++-- tools/perf/util/intel-bts.c | 3 +- tools/perf/util/intel-pt.c | 3 +- tools/perf/util/session.c | 4 +- tools/perf/util/util.c | 2 +- tools/perf/util/util.h | 2 + 21 files changed, 271 insertions(+), 92 deletions(-)
Signature · 1 line
2.7.4
PATCH v2 1/5 perf tools: pass thread info to process function
From: Kan Liang <[email protected]>
For multithreading, the process function needs to know the thread related information. E.g. saving the process result to the buffer or file which belongs to specific thread.
Add struct thread_info parameter for process function. Currently, it only includes thread index.
perf_event__repipe is shared by process function and event_op of perf_tool in builtin-inject.c. Add dedicated process function perf_event__repipe_threads.
No functional change.
--- tools/perf/arch/x86/util/tsc.c | 2 +- tools/perf/builtin-inject.c | 12 +++++++++++- tools/perf/builtin-record.c | 3 ++- tools/perf/builtin-sched.c | 12 ++++++++---- tools/perf/builtin-stat.c | 3 ++- tools/perf/builtin-trace.c | 3 ++- tools/perf/tests/cpumap.c | 6 ++++-- tools/perf/tests/dwarf-unwind.c | 3 ++- tools/perf/tests/event_update.c | 12 ++++++++---- tools/perf/tests/stat.c | 9 ++++++--- tools/perf/tests/thread-map.c | 3 ++- tools/perf/util/auxtrace.c | 2 +- tools/perf/util/event.c | 15 ++++++++------- tools/perf/util/event.h | 10 ++++++++-- tools/perf/util/header.c | 16 ++++++++-------- tools/perf/util/intel-bts.c | 3 ++- tools/perf/util/intel-pt.c | 3 ++- tools/perf/util/session.c | 4 ++-- 18 files changed, 79 insertions(+), 42 deletions(-)
tools/perf/arch/x86/util/tsc.c +1 -1
@@ -76,5 +76,5 @@ int perf_event__synth_time_conv(const struct perf_event_mmap_page *pc, event.time_conv.time_shift = tc.time_shift; event.time_conv.time_zero = tc.time_zero; -return process (tool, &event,NULL , machine); +return process (tool, &event,NULL , machine,NULL ); }
tools/perf/builtin-inject.c +11 -1
@@ -191,6 +191,15 @@ static int perf_event__repipe(struct perf_tool *tool,return perf_event__repipe_synth (tool, event); } +static int perf_event__repipe_threads (struct perf_tool *tool, +union perf_event *event, +struct perf_sample *sample, +struct machine *machine, +struct thread_info *thread__maybe_unused ) +{ +return perf_event__repipe (tool, event, sample, machine); +} + static int perf_event__drop(struct perf_tool *tool __maybe_unused, union perf_event *event __maybe_unused, struct perf_sample *sample __maybe_unused,
@@ -413,7 +422,8 @@ static int dso__inject_build_id(struct dso *dso, struct perf_tool *tool,if (dso->kernel) misc =PERF_RECORD_MISC_KERNEL ; - err =perf_event__synthesize_build_id (tool, dso, misc, perf_event__repipe, + err =perf_event__synthesize_build_id (tool, dso, misc, + perf_event__repipe_threads, machine);if (err) {pr_err ("Can't synthesize build_id event for %s\n" , dso->long_name);
tools/perf/builtin-record.c +2 -1
@@ -123,7 +123,8 @@ static int record__write(struct record *rec, void *bf, size_t size)static int process_synthesized_event (struct perf_tool *tool,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) { struct record *rec = container_of(tool, struct record, tool);return record__write (rec, event, event->header.size);
tools/perf/builtin-sched.c +8 -4
@@ -1432,7 +1432,8 @@ static void perf_sched__sort_lat(struct perf_sched *sched)static int process_sched_wakeup_event (struct perf_tool *tool,struct perf_evsel *evsel,struct perf_sample *sample, -struct machine *machine) +struct machine *machine, +struct thread_info *thread__maybe_unused ) { struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
@@ -1603,7 +1604,8 @@ static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,static int process_sched_switch_event (struct perf_tool *tool,struct perf_evsel *evsel,struct perf_sample *sample, -struct machine *machine) +struct machine *machine, +struct thread_info *thread__maybe_unused ) { struct perf_sched *sched = container_of(tool, struct perf_sched, tool);int this_cpu = sample->cpu, err =0 ;
@@ -1629,7 +1631,8 @@ static int process_sched_switch_event(struct perf_tool *tool,static int process_sched_runtime_event (struct perf_tool *tool,struct perf_evsel *evsel,struct perf_sample *sample, -struct machine *machine) +struct machine *machine, +struct thread_info *thread__maybe_unused ) { struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
@@ -1659,7 +1662,8 @@ static int perf_sched__process_fork_event(struct perf_tool *tool,static int process_sched_migrate_task_event (struct perf_tool *tool,struct perf_evsel *evsel,struct perf_sample *sample, -struct machine *machine) +struct machine *machine, +struct thread_info *thread__maybe_unused ) { struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
tools/perf/builtin-stat.c +2 -1
@@ -293,7 +293,8 @@ static inline int nsec_counter(struct perf_evsel *evsel)static int process_synthesized_event (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {if (perf_data_file__write (&perf_stat.file, event, event->header.size) <0 ) {pr_err ("failed to write perf data, error: %m\n" );
tools/perf/builtin-trace.c +2 -1
@@ -1091,7 +1091,8 @@ static int trace__process_event(struct trace *trace, struct machine *machine,static int trace__tool_process (struct perf_tool *tool,union perf_event *event,struct perf_sample *sample, -struct machine *machine) +struct machine *machine, +struct thread_info *thread__maybe_unused ) { struct trace *trace = container_of(tool, struct trace, tool);return trace__process_event (trace, machine, event, sample);
tools/perf/tests/cpumap.c +4 -2
@@ -11,7 +11,8 @@ struct machine;static int process_event_mask (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct cpu_map_event *map_event = &event->cpu_map;struct cpu_map_mask *mask;
@@ -45,7 +46,8 @@ static int process_event_mask(struct perf_tool *tool __maybe_unused,static int process_event_cpus (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct cpu_map_event *map_event = &event->cpu_map;struct cpu_map_entries *cpus;
tools/perf/tests/dwarf-unwind.c +2 -1
@@ -22,7 +22,8 @@static int mmap_handler (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample, -struct machine *machine) +struct machine *machine, +struct thread_info *thread__maybe_unused ) {return machine__process_mmap2_event (machine, event, sample); }
tools/perf/tests/event_update.c +8 -4
@@ -8,7 +8,8 @@static int process_event_unit (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct event_update_event *ev = (struct event_update_event *) event;
@@ -21,7 +22,8 @@ static int process_event_unit(struct perf_tool *tool __maybe_unused,static int process_event_scale (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct event_update_event *ev = (struct event_update_event *) event;struct event_update_event_scale *ev_data;
@@ -42,7 +44,8 @@ struct event_name {static int process_event_name (struct perf_tool *tool,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) { struct event_name *tmp = container_of(tool, struct event_name, tool);struct event_update_event *ev = (struct event_update_event *) event;
@@ -56,7 +59,8 @@ static int process_event_name(struct perf_tool *tool,static int process_event_cpus (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct event_update_event *ev = (struct event_update_event *) event;struct event_update_event_cpus *ev_data;
tools/perf/tests/stat.c +6 -3
@@ -22,7 +22,8 @@ static bool has_term(struct stat_config_event *config,static int process_stat_config_event (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct stat_config_event *config = &event->stat_config;struct perf_stat_config stat_config;
@@ -62,7 +63,8 @@ int test__synthesize_stat_config(struct test *test __maybe_unused, int subtest _static int process_stat_event (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct stat_event *st = &event->stat;
@@ -92,7 +94,8 @@ int test__synthesize_stat(struct test *test __maybe_unused, int subtest __maybe_static int process_stat_round_event (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct stat_round_event *stat_round = &event->stat_round;
tools/perf/tests/thread-map.c +2 -1
@@ -52,7 +52,8 @@ int test__thread_map(struct test *test __maybe_unused, int subtest __maybe_unusestatic int process_event (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct thread_map_event *map = &event->thread_map;struct thread_map *threads;
tools/perf/util/auxtrace.c +1 -1
@@ -887,7 +887,7 @@ int perf_event__synthesize_auxtrace_info(struct auxtrace_record *itr,if (err) goto out_free; - err =process (tool, ev,NULL ,NULL ); + err =process (tool, ev,NULL ,NULL ,NULL ); out_free:free (ev);return err;
tools/perf/util/event.c +8 -7
@@ -102,7 +102,7 @@ static int perf_tool__process_synth_event(struct perf_tool *tool, .cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK, }; -return process (tool, event, &synth_sample, machine); +return process (tool, event, &synth_sample, machine,NULL ); };/*
@@ -976,7 +976,7 @@ int perf_event__synthesize_thread_map2(struct perf_tool *tool,strncpy ((char *) &entry->comm, comm,sizeof (entry->comm)); } - err =process (tool, event,NULL , machine); + err =process (tool, event,NULL , machine,NULL );free (event);return err;
@@ -1107,7 +1107,7 @@ int perf_event__synthesize_cpu_map(struct perf_tool *tool,if (!event)return -ENOMEM ; - err =process (tool, (union perf_event *) event,NULL , machine); + err =process (tool, (union perf_event *) event,NULL , machine,NULL );free (event);return err;
@@ -1145,7 +1145,7 @@ int perf_event__synthesize_stat_config(struct perf_tool *tool, "stat config terms unbalanced\n");#undef ADD - err =process (tool, (union perf_event *) event,NULL , machine); + err =process (tool, (union perf_event *) event,NULL , machine,NULL );free (event);return err;
@@ -1170,7 +1170,7 @@ int perf_event__synthesize_stat(struct perf_tool *tool, event.ena = count->ena; event.run = count->run; -return process (tool, (union perf_event *) &event,NULL , machine); +return process (tool, (union perf_event *) &event,NULL , machine,NULL ); } int perf_event__synthesize_stat_round(struct perf_tool *tool,
@@ -1187,7 +1187,7 @@ int perf_event__synthesize_stat_round(struct perf_tool *tool, event.time = evtime; event.type = type; -return process (tool, (union perf_event *) &event,NULL , machine); +return process (tool, (union perf_event *) &event,NULL , machine,NULL ); } void perf_event__read_stat_config(struct perf_stat_config *config,
@@ -1476,7 +1476,8 @@ size_t perf_event__fprintf(union perf_event *event, FILE *fp)int perf_event__process (struct perf_tool *tool__maybe_unused ,union perf_event *event,struct perf_sample *sample, -struct machine *machine) +struct machine *machine, +struct thread_info *thread__maybe_unused ) {return machine__process_event (machine, event, sample); }
tools/perf/util/event.h +8 -2
@@ -659,10 +659,15 @@ struct cpu_map;struct perf_stat_config ;struct perf_counts_values ; +struct thread_info { +int idx; +}; +typedef int (*perf_event__handler_t )(struct perf_tool *tool,union perf_event *event,struct perf_sample *sample, -struct machine *machine); +struct machine *machine, +struct thread_info *thread); int perf_event__synthesize_thread_map(struct perf_tool *tool, struct thread_map *threads,
@@ -751,7 +756,8 @@ int perf_event__process_exit(struct perf_tool *tool,int perf_event__process (struct perf_tool *tool,union perf_event *event,struct perf_sample *sample, -struct machine *machine); +struct machine *machine, +struct thread_info *thread__maybe_unused );struct addr_location ;
tools/perf/util/header.c +8 -8
@@ -2986,7 +2986,7 @@ int perf_event__synthesize_attr(struct perf_tool *tool, ev->attr.header.size = (u16 )size;if (ev->attr.header.size == size) - err =process (tool, ev,NULL ,NULL ); + err =process (tool, ev,NULL ,NULL ,NULL );else err = -E2BIG ;
@@ -3040,7 +3040,7 @@ int perf_event__synthesize_features(struct perf_tool *tool, fe->header.type =PERF_RECORD_HEADER_FEATURE ; fe->header.size = ff.offset; - ret =process (tool, ff.buf,NULL ,NULL ); + ret =process (tool, ff.buf,NULL ,NULL ,NULL );if (ret) {free (ff.buf);return ret;
@@ -3124,7 +3124,7 @@ perf_event__synthesize_event_update_unit(struct perf_tool *tool,return -ENOMEM ;strncpy (ev->data, evsel->unit, size); - err =process (tool, (union perf_event *)ev,NULL ,NULL ); + err =process (tool, (union perf_event *)ev,NULL ,NULL ,NULL );free (ev);return err; }
@@ -3144,7 +3144,7 @@ perf_event__synthesize_event_update_scale(struct perf_tool *tool, ev_data = (struct event_update_event_scale *) ev->data; ev_data->scale = evsel->scale; - err =process (tool, (union perf_event *) ev,NULL ,NULL ); + err =process (tool, (union perf_event *) ev,NULL ,NULL ,NULL );free (ev);return err; }
@@ -3163,7 +3163,7 @@ perf_event__synthesize_event_update_name(struct perf_tool *tool,return -ENOMEM ;strncpy (ev->data, evsel->name, len); - err =process (tool, (union perf_event *) ev,NULL ,NULL ); + err =process (tool, (union perf_event *) ev,NULL ,NULL ,NULL );free (ev);return err; }
@@ -3194,7 +3194,7 @@ perf_event__synthesize_event_update_cpus(struct perf_tool *tool, evsel->own_cpus, type, max); - err =process (tool, (union perf_event *) ev,NULL ,NULL ); + err =process (tool, (union perf_event *) ev,NULL ,NULL ,NULL );free (ev);return err; }
@@ -3377,7 +3377,7 @@ int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd, ev.tracing_data.header.size =sizeof (ev.tracing_data); ev.tracing_data.size = aligned_size; -process (tool, &ev,NULL ,NULL ); +process (tool, &ev,NULL ,NULL ,NULL );/* * The put function will copy all the tracing data
@@ -3455,7 +3455,7 @@ int perf_event__synthesize_build_id(struct perf_tool *tool, ev.build_id.header.size =sizeof (ev.build_id) + len;memcpy (&ev.build_id.filename, pos->long_name, pos->long_name_len); - err =process (tool, &ev,NULL , machine); + err =process (tool, &ev,NULL , machine,NULL );return err; }
tools/perf/util/intel-bts.c +2 -1
@@ -755,7 +755,8 @@ struct intel_bts_synth {static int intel_bts_event_synth (struct perf_tool *tool,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct intel_bts_synth *intel_bts_synth = container_of(tool, struct intel_bts_synth, dummy_tool);
tools/perf/util/intel-pt.c +2 -1
@@ -2121,7 +2121,8 @@ struct intel_pt_synth {static int intel_pt_event_synth (struct perf_tool *tool,union perf_event *event,struct perf_sample *sample__maybe_unused , -struct machine *machine__maybe_unused ) +struct machine *machine__maybe_unused , +struct thread_info *thread__maybe_unused ) {struct intel_pt_synth *intel_pt_synth = container_of(tool, struct intel_pt_synth, dummy_tool);
tools/perf/util/session.c +2 -2
@@ -2211,7 +2211,7 @@ int perf_event__synthesize_id_index(struct perf_tool *tool,struct perf_sample_id *sid;if (i >= n) { - err =process (tool, ev,NULL , machine); + err =process (tool, ev,NULL , machine,NULL );if (err) goto out_err; nr -= n;
@@ -2238,7 +2238,7 @@ int perf_event__synthesize_id_index(struct perf_tool *tool, ev->id_index.header.size = sz; ev->id_index.nr = nr; - err =process (tool, ev,NULL , machine); + err =process (tool, ev,NULL , machine,NULL ); out_err:free (ev);
Signature · 1 line
2.7.4
PATCH v2 2/5 perf tools: pass thread info in event synthesization
In reply to [email protected]From: Kan Liang <[email protected]>
Pass the thread idx to process function, which is used by the following patch.
--- tools/perf/builtin-record.c | 4 +- tools/perf/tests/dwarf-unwind.c | 3 +- tools/perf/util/event.c | 98 ++++++++++++++++++++++++++--------------- tools/perf/util/event.h | 9 ++-- 4 files changed, 72 insertions(+), 42 deletions(-)
tools/perf/builtin-record.c +2 -2
@@ -922,7 +922,7 @@ static int __cmd_record(struct record *rec, int argc, const char **argv) tgid =perf_event__synthesize_comm (tool, event, rec->evlist->workload.pid, process_synthesized_event, - machine); + machine,NULL );free (event);if (tgid ==-1 )
@@ -942,7 +942,7 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)perf_event__synthesize_namespaces (tool, event, rec->evlist->workload.pid, tgid, process_synthesized_event, - machine); + machine,NULL );free (event);perf_evlist__start_workload (rec->evlist);
tools/perf/tests/dwarf-unwind.c +2 -1
@@ -34,7 +34,8 @@ static int init_live_machine(struct machine *machine)pid_t pid =getpid ();return perf_event__synthesize_mmap_events (NULL , &event, pid, pid, - mmap_handler, machine, true,500 ); + mmap_handler, machine, + true,500 ,NULL ); }#define MAX_STACK 8
tools/perf/util/event.c +62 -36
257 lines not on this page · raw message
tools/perf/util/event.h +6 -3
@@ -785,13 +785,15 @@ int perf_event__synthesize_sample(union perf_event *event, u64 type,pid_t perf_event__synthesize_comm (struct perf_tool *tool,union perf_event *event,pid_t pid,perf_event__handler_t process, -struct machine *machine); +struct machine *machine, +struct thread_info *thread);int perf_event__synthesize_namespaces (struct perf_tool *tool,union perf_event *event,pid_t pid,pid_t tgid,perf_event__handler_t process, -struct machine *machine); +struct machine *machine, +struct thread_info *thread); int perf_event__synthesize_mmap_events(struct perf_tool *tool, union perf_event *event,
@@ -799,7 +801,8 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,perf_event__handler_t process, struct machine *machine, bool mmap_data, - unsigned int proc_map_timeout); + unsigned int proc_map_timeout, + struct thread_info *thread);size_t perf_event__fprintf_comm (union perf_event *event,FILE *fp);size_t perf_event__fprintf_mmap (union perf_event *event,FILE *fp);
Signature · 1 line
2.7.4
PATCH v2 3/5 perf tools: expose copyfile_offset()
In reply to [email protected]From: Kan Liang <[email protected]>
copyfile_offset could be used to merge per thread file to perf.data in the following patch.
--- tools/perf/util/util.c | 2 +- tools/perf/util/util.h | 2 ++ 2 files changed, 3 insertions(+), 1 deletion(-)
tools/perf/util/util.c +1 -1
@@ -188,7 +188,7 @@ static int slow_copyfile(const char *from, const char *to, struct nsinfo *nsi)return err; } -static int copyfile_offset (int ifd,loff_t off_in,int ofd,loff_t off_out,u64 size) +int copyfile_offset (int ifd,loff_t off_in,int ofd,loff_t off_out,u64 size) {void *ptr;loff_t pgoff;
tools/perf/util/util.h +2 -0
@@ -5,6 +5,7 @@/* glibc 2.20 deprecates _BSD_SOURCE in favour of _DEFAULT_SOURCE */ #define _DEFAULT_SOURCE 1 +#include <fcntl.h> #include <stdbool.h> #include <stddef.h> #include <stdlib.h>
@@ -34,6 +35,7 @@ bool lsdir_no_dot_filter(const char *name, struct dirent *d);int copyfile (const char *from,const char *to);int copyfile_mode (const char *from,const char *to,mode_t mode);int copyfile_ns (const char *from,const char *to,struct nsinfo *nsi); +int copyfile_offset (int fromfd,loff_t from_ofs,int tofd,loff_t to_ofs,u64 size);ssize_t readn (int fd,void *buf,size_t n);ssize_t writen (int fd,const void *buf,size_t n);
Signature · 1 line
2.7.4
PATCH v2 4/5 perf record: synthesize event multithreading support
In reply to [email protected]From: Kan Liang <[email protected]>
The process function process_synthesized_event writes the process result to perf.data, which is not multithreading friendly.
Create per thread file to temporarily keep the processing result. Write them to the perf.data at the end of event synthesization. The new method doesn't impact the final result, because the order of the synthesized event is not important.
The threads number hard code to online CPU number. The following patch will introduce an option to set it.
The multithreading synthesize is only available for per cpu monitoring.
--- tools/perf/builtin-record.c | 115 ++++++++++++++++++++++++++++++++++++++++---- 1 file changed, 106 insertions(+), 9 deletions(-)
tools/perf/builtin-record.c +106 -9
@@ -50,9 +50,12 @@#include <signal.h> #include <sys/mman.h> #include <sys/wait.h> +#include <sys/stat.h> #include <asm/bug.h> #include <linux/time64.h> +#define SYNTHESIZED_PATH "perf.synthesized" +struct switch_output {bool enabled;bool signal;
@@ -80,6 +83,7 @@ struct record {bool timestamp_filename;struct switch_output switch_output;unsigned long long samples; +struct perf_data_file *synthesized_file; };static volatile int auxtrace_record__snapshot_started;
@@ -105,6 +109,14 @@ static bool switch_output_time(struct record *rec)trigger_is_ready (&switch_output_trigger); } +static void update_bytes_written (struct record *rec,size_t size) +{ + rec->bytes_written += size; + +if (switch_output_size (rec)) +trigger_hit (&switch_output_trigger); +} +static int record__write (struct record *rec,void *bf,size_t size) {if (perf_data_file__write (rec->session->file, bf, size) <0 ) {
@@ -112,10 +124,7 @@ static int record__write(struct record *rec, void *bf, size_t size)return -1 ; } - rec->bytes_written += size; - -if (switch_output_size (rec)) -trigger_hit (&switch_output_trigger); +update_bytes_written (rec, size);return 0 ; }
@@ -124,10 +133,15 @@ static int process_synthesized_event(struct perf_tool *tool,union perf_event *event, struct perf_sample *sample __maybe_unused, struct machine *machine __maybe_unused, - struct thread_info *thread __maybe_unused) + struct thread_info *thread) { struct record *rec = container_of(tool, struct record, tool); -return record__write (rec, event, event->header.size); + +if (!perf_singlethreaded && thread) +return (perf_data_file__write (&rec->synthesized_file[thread->idx], + event, event->header.size) <0 ) ?-1 :0 ; +else +return record__write (rec, event, event->header.size); }static int record__pushfn (void *to,void *bf,size_t size)
@@ -690,6 +704,82 @@ static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)return NULL ; } +static int record__multithread_synthesize (struct record *rec, +struct machine *machine, +struct perf_tool *tool, +struct record_opts *opts) +{ +int i, err, nr_thread =sysconf (_SC_NPROCESSORS_ONLN); +char name[PATH_MAX ]; +struct stat st; + +if (nr_thread <=1 ) +return __machine__synthesize_threads (machine, tool, + &opts->target, + rec->evlist->threads, + process_synthesized_event, + opts->sample_address, + opts->proc_map_timeout, +1 ); + + rec->synthesized_file =calloc (nr_thread,sizeof (struct perf_data_file )); +if (rec->synthesized_file ==NULL ) { +pr_debug ("Could not do multithread synthesize." +"Roll back to single thread\n" ); + nr_thread =1 ; + }else { +perf_set_multithreaded (); +for (i =0 ; i < nr_thread; i++) { +snprintf (name,sizeof (name),"%s.%d" , +SYNTHESIZED_PATH , i); + rec->synthesized_file[i].path = name; + err =perf_data_file__open (&rec->synthesized_file[i]); +if (err) { +pr_err ("Failed to open file %s\n" , + rec->synthesized_file[i].path); + goto free; + } + } + } + + err =__machine__synthesize_threads (machine, tool, &opts->target, + rec->evlist->threads, + process_synthesized_event, + opts->sample_address, + opts->proc_map_timeout, nr_thread); +if (err <0 ) + goto free; + +if (nr_thread >1 ) { +int fd_from, fd_to; + + fd_to = rec->session->file->fd; +for (i =0 ; i < nr_thread; i++) { + fd_from = rec->synthesized_file[i].fd; + +fstat (fd_from, &st); +if (st.st_size ==0 ) +continue ; + err =copyfile_offset (fd_from,0 , fd_to, +lseek (fd_to,0 ,SEEK_END ), + st.st_size); +update_bytes_written (rec, st.st_size); + } + } + +free: +if (nr_thread >1 ) { +for (i =0 ; i < nr_thread; i++) { +if (rec->synthesized_file[i].fd >0 ) +perf_data_file__close (&rec->synthesized_file[i]); + } +free (rec->synthesized_file); +perf_set_singlethreaded (); + } + +return err; +} +static int record__synthesize (struct record *rec,bool tail) {struct perf_session *session = rec->session;
@@ -766,9 +856,16 @@ static int record__synthesize(struct record *rec, bool tail) perf_event__synthesize_guest_os, tool); } - err =__machine__synthesize_threads (machine, tool, &opts->target, rec->evlist->threads, - process_synthesized_event, opts->sample_address, - opts->proc_map_timeout,1 ); +/* multithreading synthesize is only available for cpu monitoring */ +if (target__has_cpu (&opts->target)) + err =record__multithread_synthesize (rec, machine, tool, opts); +else + err =__machine__synthesize_threads (machine, tool, + &opts->target, + rec->evlist->threads, + process_synthesized_event, + opts->sample_address, + opts->proc_map_timeout,1 ); out:return err; }
Signature · 1 line
2.7.4
Re: [PATCH V2 4/5] perf record: synthesize event multithreading support
On Wed, Oct 18, 2017 at 07:29:32AM -0700, [email protected] wrote:
SNIP
Quoted 13 lines
tools/perf/builtin-record.c +13
-0+static int record__multithread_synthesize (struct record *rec, +struct machine *machine, +struct perf_tool *tool, +struct record_opts *opts) +{ +int i, err, nr_thread =sysconf (_SC_NPROCESSORS_ONLN); +char name[PATH_MAX ]; +struct stat st; + +if (nr_thread <=1 ) + return __machine__synthesize_threads(machine, tool, + &opts->target, + rec->evlist->threads,
tools/perf/builtin-record.c static const struct perf_event_mmap_page *record__pick_pc(struct record *rec) · line 729
Quoted 10 lines
tools/perf/builtin-record.c +10
-0+ process_synthesized_event, + opts->sample_address, + opts->proc_map_timeout, + 1); + + rec->synthesized_file =calloc (nr_thread,sizeof (struct perf_data_file )); +if (rec->synthesized_file ==NULL ) { +pr_debug ("Could not do multithread synthesize." +"Roll back to single thread\n" ); + nr_thread =1 ;
I guess we could fail here anyway.. it'd simplify the code below ;-)
jirka
Quoted 52 lines
tools/perf/builtin-record.c +52
-0+ }else { +perf_set_multithreaded (); +for (i =0 ; i < nr_thread; i++) { +snprintf (name,sizeof (name),"%s.%d" , +SYNTHESIZED_PATH , i); + rec->synthesized_file[i].path = name; + err =perf_data_file__open (&rec->synthesized_file[i]); +if (err) { +pr_err ("Failed to open file %s\n" , + rec->synthesized_file[i].path); + goto free; + } + } + } + + err =__machine__synthesize_threads (machine, tool, &opts->target, + rec->evlist->threads, + process_synthesized_event, + opts->sample_address, + opts->proc_map_timeout, nr_thread); +if (err <0 ) + goto free; + +if (nr_thread >1 ) { +int fd_from, fd_to; + + fd_to = rec->session->file->fd; +for (i =0 ; i < nr_thread; i++) { + fd_from = rec->synthesized_file[i].fd; + +fstat (fd_from, &st); +if (st.st_size ==0 ) +continue ; + err =copyfile_offset (fd_from,0 , fd_to, +lseek (fd_to,0 ,SEEK_END ), + st.st_size); +update_bytes_written (rec, st.st_size); + } + } + +free: +if (nr_thread >1 ) { +for (i =0 ; i < nr_thread; i++) { +if (rec->synthesized_file[i].fd >0 ) +perf_data_file__close (&rec->synthesized_file[i]); + } +free (rec->synthesized_file); +perf_set_singlethreaded (); + } + +return err; +}
SNIP
Re: [PATCH V2 4/5] perf record: synthesize event multithreading support
In reply to [email protected]On Wed, Oct 18, 2017 at 07:29:32AM -0700, [email protected] wrote:
SNIP
Quoted 7 lines
tools/perf/builtin-record.c +7
-0+static int record__multithread_synthesize (struct record *rec, +struct machine *machine, +struct perf_tool *tool, +struct record_opts *opts) +{ +int i, err, nr_thread =sysconf (_SC_NPROCESSORS_ONLN); +char name[PATH_MAX ];
tools/perf/builtin-record.c static const struct perf_event_mmap_page *record__pick_pc(struct record *rec) · line 723
Quoted 10 lines
tools/perf/builtin-record.c +10
-0+struct stat st; + +if (nr_thread <=1 ) +return __machine__synthesize_threads (machine, tool, + &opts->target, + rec->evlist->threads, + process_synthesized_event, + opts->sample_address, + opts->proc_map_timeout, +1 );
^^^^^^^^^^^^^^^^^^
Quoted 11 lines
tools/perf/builtin-record.c +1
-3+static int record__synthesize (struct record *rec,bool tail) {struct perf_session *session = rec->session;@@ -766,9 +856,16 @@ static int record__synthesize(struct record *rec, bool tail) perf_event__synthesize_guest_os, tool); } - err =__machine__synthesize_threads (machine, tool, &opts->target, rec->evlist->threads, - process_synthesized_event, opts->sample_address, - opts->proc_map_timeout,1 );
tools/perf/builtin-record.c static int record__synthesize(struct record *rec, bool tail) · line 868
Quoted 10 lines
tools/perf/builtin-record.c +10
-0+/* multithreading synthesize is only available for cpu monitoring */ +if (target__has_cpu (&opts->target)) + err =record__multithread_synthesize (rec, machine, tool, opts); +else + err =__machine__synthesize_threads (machine, tool, + &opts->target, + rec->evlist->threads, + process_synthesized_event, + opts->sample_address, + opts->proc_map_timeout,1 );
hum, this could be checked together with nr_thread in record__multithread_synthesize
jirka
Re: [PATCH V2 4/5] perf record: synthesize event multithreading support
In reply to [email protected]On Wed, Oct 18, 2017 at 07:29:32AM -0700, [email protected] wrote:
SNIP
tools/perf/builtin-record.c static const struct perf_event_mmap_page *record__pick_pc(struct record *rec) · line 734
Quoted 10 lines
tools/perf/builtin-record.c +10
-0+ rec->synthesized_file =calloc (nr_thread,sizeof (struct perf_data_file )); +if (rec->synthesized_file ==NULL ) { +pr_debug ("Could not do multithread synthesize." +"Roll back to single thread\n" ); + nr_thread =1 ; + }else { +perf_set_multithreaded (); +for (i =0 ; i < nr_thread; i++) { +snprintf (name,sizeof (name),"%s.%d" , +SYNTHESIZED_PATH , i);
hum, I think we want some uniq temp names in here..
Quoted 30 lines
tools/perf/builtin-record.c +30
-0+ rec->synthesized_file[i].path = name; + err =perf_data_file__open (&rec->synthesized_file[i]); +if (err) { +pr_err ("Failed to open file %s\n" , + rec->synthesized_file[i].path); + goto free; + } + } + } + + err =__machine__synthesize_threads (machine, tool, &opts->target, + rec->evlist->threads, + process_synthesized_event, + opts->sample_address, + opts->proc_map_timeout, nr_thread); +if (err <0 ) + goto free; + +if (nr_thread >1 ) { +int fd_from, fd_to; + + fd_to = rec->session->file->fd; +for (i =0 ; i < nr_thread; i++) { + fd_from = rec->synthesized_file[i].fd; + +fstat (fd_from, &st); +if (st.st_size ==0 ) +continue ; + err = copyfile_offset(fd_from, 0, fd_to, + lseek(fd_to, 0, SEEK_END),
tools/perf/builtin-record.c static const struct perf_event_mmap_page *record__pick_pc(struct record *rec) · line 774
Quoted 10 lines
tools/perf/builtin-record.c +10
-0+ st.st_size); +update_bytes_written (rec, st.st_size); + } + } + +free: +if (nr_thread >1 ) { +for (i =0 ; i < nr_thread; i++) { +if (rec->synthesized_file[i].fd >0 ) +perf_data_file__close (&rec->synthesized_file[i]);
also those files should be removed
thanks,
jirka
Quoted 8 lines
tools/perf/builtin-record.c +8
-0+ } +free (rec->synthesized_file); +perf_set_singlethreaded (); + } + +return err; +} +
SNIP
RE: [PATCH V2 4/5] perf record: synthesize event multithreading support
Quoted 6 lines · On Wed, Oct 18, 2017 at 07:29:32AM -0700, [email protected] wrote:
SNIP
Review
Quoted 1 line
(quoted diff) +1
-0+ rec->synthesized_file = calloc(nr_thread, sizeof(struct
perf_data_file));
Quoted 11 lines
tools/perf/builtin-record.c static const struct perf_event_mmap_page *record__pick_pc(struct record *rec) · line 734
Quoted 9 lines
tools/perf/builtin-record.c +9
-0+if (rec->synthesized_file ==NULL ) { +pr_debug ("Could not do multithread synthesize." +"Roll back to single thread\n" ); + nr_thread =1 ; + }else { +perf_set_multithreaded (); +for (i =0 ; i < nr_thread; i++) { +snprintf (name,sizeof (name),"%s.%d" , +SYNTHESIZED_PATH , i);hum, I think we want some uniq temp names in here..
Any suggestions for the name?
perf.tmp.X?
Quoted 37 lines
Quoted 6 lines
tools/perf/builtin-record.c +5
-0+ rec->synthesized_file[i].path = name; + err =perf_data_file__open (&rec->synthesized_file[i]); +if (err) { +pr_err ("Failed to open file %s\n" , + rec->synthesized_file[i].path);tools/perf/builtin-record.c static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
Quoted 10 lines
tools/perf/builtin-record.c +10
-0+ goto free; + } + } + } + + err = __machine__synthesize_threads(machine, tool, &opts->target, + rec->evlist->threads, + process_synthesized_event, + opts->sample_address, + opts->proc_map_timeout,
nr_thread);Quoted 20 lines
tools/perf/builtin-record.c +20
-0+if (err <0 ) + goto free; + +if (nr_thread >1 ) { +int fd_from, fd_to; + + fd_to = rec->session->file->fd; +for (i =0 ; i < nr_thread; i++) { + fd_from = rec->synthesized_file[i].fd; + +fstat (fd_from, &st); +if (st.st_size ==0 ) +continue ; + err =copyfile_offset (fd_from,0 , fd_to, +lseek (fd_to,0 ,SEEK_END ), + st.st_size); +update_bytes_written (rec, st.st_size); + } + } +
Quoted 8 lines
Quoted 6 lines
tools/perf/builtin-record.c +5
-0+free: +if (nr_thread >1 ) { +for (i =0 ; i < nr_thread; i++) { +if (rec->synthesized_file[i].fd >0 ) + perf_data_file__close(&rec-synthesized_file[i]);
also those files should be removed
Sure.
I will modify the perf_data_file__close(file, bool remove) to do it.
Thanks,
Kan
PATCH v2 5/5 perf record: add option to set the number of thread for event synthesize
In reply to [email protected]From: Kan Liang <[email protected]>
Using UINT_MAX to indicate the default thread#, which is the number of online CPU.
--- tools/perf/Documentation/perf-record.txt | 4 ++++ tools/perf/builtin-record.c | 13 +++++++++++-- 2 files changed, 15 insertions(+), 2 deletions(-)
tools/perf/Documentation/perf-record.txt +4 -0
@@ -483,6 +483,10 @@ config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'. Implies --tail-synthesize. +--num-thread-synthesize:: +The number of threads to run event synthesize. +By default, the number of threads equals to the online CPU number. + SEE ALSO -------- linkperf:perf-stat[1], linkperf:perf-list[1]
tools/perf/builtin-record.c +11 -2
@@ -84,6 +84,7 @@ struct record {struct switch_output switch_output;unsigned long long samples;struct perf_data_file *synthesized_file; +unsigned int nr_threads_synthesize; };static volatile int auxtrace_record__snapshot_started;
@@ -709,10 +710,15 @@ static int record__multithread_synthesize(struct record *rec,struct perf_tool *tool, struct record_opts *opts) { - int i, err, nr_thread = sysconf(_SC_NPROCESSORS_ONLN); + int i, err, nr_thread;char name[PATH_MAX ];struct stat st; +if (rec->nr_threads_synthesize ==UINT_MAX ) + nr_thread =sysconf (_SC_NPROCESSORS_ONLN); +else + nr_thread = rec->nr_threads_synthesize; + if (nr_thread <= 1) return __machine__synthesize_threads(machine, tool, &opts->target,
@@ -857,7 +863,7 @@ static int record__synthesize(struct record *rec, bool tail) }/* multithreading synthesize is only available for cpu monitoring */ -if (target__has_cpu (&opts->target)) +if (target__has_cpu (&opts->target) && (rec->nr_threads_synthesize >1 )) err =record__multithread_synthesize (rec, machine, tool, opts);else err = __machine__synthesize_threads(machine, tool,
@@ -1542,6 +1548,7 @@ static struct record record = { .mmap2 = perf_event__process_mmap2, .ordered_events = true, }, + .nr_threads_synthesize = UINT_MAX, };const char record_callchain_help[] =CALLCHAIN_RECORD_HELP
@@ -1683,6 +1690,8 @@ static struct option __record_options[] = { "signal"),OPT_BOOLEAN (0 ,"dry-run" , &dry_run,"Parse options then exit" ), +OPT_UINTEGER (0 ,"num-thread-synthesize" , &record.nr_threads_synthesize, +"number of thread to run event synthesize" ), OPT_END() };
Signature · 1 line
2.7.4
Quoted 29 lines · * [email protected] <[email protected]> wrote:
From: Kan Liang <[email protected]>
The event synthesization multithreading is introduced in ("perf top optimization") https://lkml.org/lkml/2017/9/29/269
But it was not enabled for perf record. Because the process functionprocess_synthesized_eventwas not multithreading friendly.The patch series temporarily stores the process result in per-thread file, which make the processing in parallel. Then it dumps the file one by one to the perf.data at the end of event synthesization.
The source code is also available at
https://github.com/kliang2/perf.gitperf_record_optUsually, the event synthesization only happens once on either start or end. With the snapshotting code, we synthesize events multiple times, once per each new perf.data file. Both of the cases are verified.
Here are the latency test result on Knights Mill and Skylake server
The workload is to compile Linux kernel as below "sudo nice make -j$(grep -c '^processor'
/proc/cpuinfo)" Then, "sudo perf record -e cycles -a -- sleep 1"The latency is the time cost of
__machine__synthesize_threadsor its multithreading replacement,record__multithread_synthesize.- Latency on Knights Mill (272 CPUs)
Quoted 7 lines
Original(s) With patch(s) Speedup 12.74 5.54 2.3X- Latency on Skylake server (192 CPUs)
Original(s) With patch(s) Speedup 0.36 0.25 1.47X
Btw., just as an interesting experiment, could you try to measure how it performs to create just the per-CPU files, and *not* dump them into a single file?
I.e. how much faster will it get if the serialization at the end is avoided?
Of course nothing can read such per-CPU files yet, so this is just for scalability measurement.
Thanks,
Ingo
Quoted 35 lines · * [email protected] <[email protected]> wrote:
Quoted 33 lines
From: Kan Liang <[email protected]>
The event synthesization multithreading is introduced in ("perf top optimization") https://lkml.org/lkml/2017/9/29/269
But it was not enabled for perf record. Because the process functionprocess_synthesized_eventwas not multithreading friendly.The patch series temporarily stores the process result in per-thread file, which make the processing in parallel. Then it dumps the file one by one to the perf.data at the end of event synthesization.
The source code is also available at
https://github.com/kliang2/perf.gitperf_record_optUsually, the event synthesization only happens once on either start or end. With the snapshotting code, we synthesize events multiple times, once per each new perf.data file. Both of the cases are verified.
Here are the latency test result on Knights Mill and Skylake server
The workload is to compile Linux kernel as below "sudo nice make -j$(grep -c '^processor'
/proc/cpuinfo)"
Then, "sudo perf record -e cycles -a -- sleep 1"The latency is the time cost of
__machine__synthesize_threadsor its multithreading replacement,record__multithread_synthesize.- Latency on Knights Mill (272 CPUs)
Original(s) With patch(s) Speedup 12.74 5.54 2.3X
Quoted 9 lines
Quoted 4 lines
- Latency on Skylake server (192 CPUs)
Original(s) With patch(s) Speedup 0.36 0.25 1.47XBtw., just as an interesting experiment, could you try to measure how it performs to create just the per-CPU files, and *not* dump them into a single file?
Sure, please find the experiment result in the cover letter of V3 patch series.
Thanks,
Kan
Quoted 8 lines
I.e. how much faster will it get if the serialization at the end is avoided?
Of course nothing can read such per-CPU files yet, so this is just for scalability measurement.
Thanks,
Ingo