perf report: Add latency and parallelism profiling
v7 cover perf report: Add latency and parallelism profiling
There are two notions of time: wall-clock time and CPU time. For a single-threaded program, or a program running on a single-core machine, these notions are the same. However, for a multi-threaded/ multi-process program running on a multi-core machine, these notions are significantly different. Each second of wall-clock time we have number-of-cores seconds of CPU time.
Currently perf only allows to profile CPU time. Perf (and all other existing profilers to the be best of my knowledge) does not allow to profile wall-clock time.
Optimizing CPU overhead is useful to improve 'throughput', while optimizing wall-clock overhead is useful to improve 'latency'. These profiles are complementary and are not interchangeable. Examples of where latency profile is needed:
- optimzing build latency
- optimizing server request latency
- optimizing ML training/inference latency
- optimizing running time of any command line program
CPU profile is useless for these use cases at best (if a user understands the difference), or misleading at worst (if a user tries to use a wrong profile for a job).
This series add latency and parallelization profiling. See the added documentation and flags descriptions for details.
Brief outline of the implementation:
- add context switch collection during record
- calculate number of threads running on CPUs (parallelism level) during report
- divide each sample weight by the parallelism level
This effectively models that we were taking 1 sample per unit of wall-clock time.
We still default to the CPU profile, so it's up to users to learn about the second profiling mode and use it when appropriate.
Changes in v7:
- add comment perf_hpp__init() re was_taken logic
- add tips.txt example with time/time-quantum
- rebased to perf-tools-next HEAD
Changes in v6:
- remove latency column in perf_hpp__cancel_latency if sort order is specified, but does not include latency
- add tests
- rebased to perf-tools-next HEAD
Changes in v5:
- fixed formatting of latency field in --stdout mode
- added description of --latency flag in Documentation flags
Changes in v4:
- added "Shrink struct hist_entry size" commit
- rebased to perf-tools-next HEAD
Changes in v3:
- rebase and split into patches
- rename 'wallclock' to 'latency' everywhere
- don't enable latency profiling by default,
instead add record/report --latency flag
Dmitry Vyukov (9):
perf report: Add machine parallelism
perf report: Add parallelism sort key
perf report: Switch filtered from u8 to u16
perf report: Add parallelism filter
perf report: Add latency output field
perf report: Add --latency flag
perf report: Add latency and parallelism profiling documentation perf test: Add tests for latency and parallelism profiling perf hist: Shrink struct hist_entry size
.../callchain-overhead-calculation.txt | 5 +- .../cpu-and-latency-overheads.txt | 85 +++++++++++++ tools/perf/Documentation/perf-record.txt | 4 + tools/perf/Documentation/perf-report.txt | 54 +++++--- tools/perf/Documentation/tips.txt | 4 + tools/perf/builtin-record.c | 20 +++ tools/perf/builtin-report.c | 39 ++++++ tools/perf/tests/shell/base_report/setup.sh | 18 ++- .../tests/shell/base_report/test_basic.sh | 52 ++++++++ tools/perf/ui/browsers/hists.c | 27 ++-- tools/perf/ui/hist.c | 117 ++++++++++++------ tools/perf/util/addr_location.c | 1 + tools/perf/util/addr_location.h | 7 +- tools/perf/util/event.c | 11 ++ tools/perf/util/events_stats.h | 2 + tools/perf/util/hist.c | 90 +++++++++++--- tools/perf/util/hist.h | 32 ++++- tools/perf/util/machine.c | 7 ++ tools/perf/util/machine.h | 6 + tools/perf/util/sample.h | 2 +- tools/perf/util/session.c | 12 ++ tools/perf/util/session.h | 1 + tools/perf/util/sort.c | 69 ++++++++++- tools/perf/util/sort.h | 3 +- tools/perf/util/symbol.c | 34 +++++ tools/perf/util/symbol_conf.h | 8 +- 26 files changed, 614 insertions(+), 96 deletions(-) create mode 100644 tools/perf/Documentation/cpu-and-latency-overheads.txt
base-commit: ee8aef2d232142e5fdfed9c16132815969a0bf81
Signature · 1 line
2.48.1.502.g6dc24dfdaf-goog
PATCH v7 1/9 perf report: Add machine parallelism
Add calculation of the current parallelism level (number of threads actively running on CPUs). The parallelism level can be shown in reports on its own, and to calculate latency overheads.
--- tools/perf/builtin-report.c | 1 + tools/perf/util/addr_location.c | 1 + tools/perf/util/addr_location.h | 2 ++ tools/perf/util/event.c | 3 +++ tools/perf/util/machine.c | 7 +++++++ tools/perf/util/machine.h | 6 ++++++ 6 files changed, 20 insertions(+)
tools/perf/builtin-report.c +1 -0
@@ -1568,6 +1568,7 @@ int cmd_report(int argc, const char **argv) report.tool.cgroup = perf_event__process_cgroup; report.tool.exit = perf_event__process_exit; report.tool.fork = perf_event__process_fork; + report.tool.context_switch = perf_event__process_switch; report.tool.lost = perf_event__process_lost; report.tool.read = process_read_event; report.tool.attr = process_attr;
tools/perf/util/addr_location.c +1 -0
@@ -17,6 +17,7 @@ void addr_location__init(struct addr_location *al) al->cpumode =0 ; al->cpu =0 ; al->socket =0 ; + al->parallelism =1 ; }/*
tools/perf/util/addr_location.h +2 -0
@@ -21,6 +21,8 @@ struct addr_location {u8 cpumode;s32 cpu;s32 socket; +/* Same as machine.parallelism but within [1, nr_cpus]. */ +int parallelism; };void addr_location__init (struct addr_location *al);
tools/perf/util/event.c +3 -0
@@ -767,6 +767,9 @@ int machine__resolve(struct machine *machine, struct addr_location *al, al->socket = env->cpu[al->cpu].socket_id; } +/* Account for possible out-of-order switch events. */ + al->parallelism =max (1 ,min (machine->parallelism,machine__nr_cpus_avail (machine))); +if (al->map) { if (symbol_conf.dso_list && (!dso || !(strlist__has_entry(symbol_conf.dso_list,
tools/perf/util/machine.c +7 -0
@@ -94,6 +94,8 @@ int machine__init(struct machine *machine, const char *root_dir, pid_t pid) machine->comm_exec = false; machine->kernel_start =0 ; machine->vmlinux_map =NULL ; +/* There is no initial context switch in, so we start at 1. */ + machine->parallelism =1 ; machine->root_dir =strdup (root_dir);if (machine->root_dir ==NULL )
@@ -677,8 +679,11 @@ int machine__process_aux_output_hw_id_event(struct machine *machine __maybe_unusint machine__process_switch_event (struct machine *machine__maybe_unused ,union perf_event *event) { +bool out = event->header.misc &PERF_RECORD_MISC_SWITCH_OUT ; +if (dump_trace)perf_event__fprintf_switch (event, stdout); + machine->parallelism += out ?-1 :1 ;return 0 ; }
@@ -1880,6 +1885,8 @@ int machine__process_exit_event(struct machine *machine, union perf_event *eventif (dump_trace)perf_event__fprintf_task (event, stdout); +/* There is no context switch out before exit, so we decrement here. */ + machine->parallelism--;if (thread !=NULL ) {if (symbol_conf.keep_exited_threads)thread__set_exited (thread,/*exited=*/ true);
tools/perf/util/machine.h +6 -0
@@ -50,6 +50,12 @@ struct machine {u64 text_start;u64 text_end; } sched, lock, traceiter, trace; +/* +* The current parallelism level (number of threads that run on CPUs). +* This value can be less than 1, or larger than the total number +* of CPUs, if events are poorly ordered. +*/ +int parallelism;pid_t *current_tid;size_t current_tid_sz;union {/* Tool specific area */
Signature · 1 line
2.48.1.502.g6dc24dfdaf-goog
PATCH v7 2/9 perf report: Add parallelism sort key
In reply to Dmitry VyukovShow parallelism level in profiles if requested by user.
--- tools/perf/builtin-report.c | 11 +++++++++++ tools/perf/util/hist.c | 2 ++ tools/perf/util/hist.h | 3 +++ tools/perf/util/session.c | 12 ++++++++++++ tools/perf/util/session.h | 1 + tools/perf/util/sort.c | 23 +++++++++++++++++++++++ tools/perf/util/sort.h | 1 + 7 files changed, 53 insertions(+)
tools/perf/builtin-report.c +11 -0
@@ -1720,6 +1720,17 @@ int cmd_report(int argc, const char **argv) symbol_conf.annotate_data_sample = true; } +if (report.disable_order || !perf_session__has_switch_events (session)) { +if ((sort_order &&strstr (sort_order,"parallelism" )) || + (field_order &&strstr (field_order,"parallelism" ))) { +if (report.disable_order) +ui__error ("Use of parallelism is incompatible with --disable-order.\n" ); +else +ui__error ("Use of parallelism requires --switch-events during record.\n" ); +return -1 ; + } + } +if (sort_order &&strstr (sort_order,"ipc" )) {parse_options_usage (report_usage, options,"s" ,1 ); goto error;
tools/perf/util/hist.c +2 -0
@@ -207,6 +207,7 @@ void hists__calc_col_len(struct hists *hists, struct hist_entry *h)hists__new_col_len (hists,HISTC_CGROUP ,6 );hists__new_col_len (hists,HISTC_CGROUP_ID ,20 ); +hists__new_col_len (hists,HISTC_PARALLELISM ,11 );hists__new_col_len (hists,HISTC_CPU ,3 );hists__new_col_len (hists,HISTC_SOCKET ,6 );hists__new_col_len (hists,HISTC_MEM_LOCKED ,6 );
@@ -741,6 +742,7 @@ __hists__add_entry(struct hists *hists, .ip = al->addr, .level = al->level, .code_page_size = sample->code_page_size, + .parallelism = al->parallelism, .stat = { .nr_events = 1, .period = sample->period,
tools/perf/util/hist.h +3 -0
@@ -42,6 +42,7 @@ enum hist_column {HISTC_CGROUP_ID ,HISTC_CGROUP ,HISTC_PARENT , +HISTC_PARALLELISM ,HISTC_CPU ,HISTC_SOCKET , HISTC_SRCLINE,
@@ -228,6 +229,7 @@ struct hist_entry {u64 transaction;s32 socket;s32 cpu; +int parallelism;u64 code_page_size;u64 weight;u64 ins_lat;
@@ -580,6 +582,7 @@ bool perf_hpp__is_thread_entry(struct perf_hpp_fmt *fmt);bool perf_hpp__is_comm_entry (struct perf_hpp_fmt *fmt);bool perf_hpp__is_dso_entry (struct perf_hpp_fmt *fmt);bool perf_hpp__is_sym_entry (struct perf_hpp_fmt *fmt); +bool perf_hpp__is_parallelism_entry (struct perf_hpp_fmt *fmt);struct perf_hpp_fmt *perf_hpp_fmt__dup (struct perf_hpp_fmt *fmt);
tools/perf/util/session.c +12 -0
@@ -2403,6 +2403,18 @@ bool perf_session__has_traces(struct perf_session *session, const char *msg)return false; } +bool perf_session__has_switch_events (struct perf_session *session) +{ +struct evsel *evsel; + +evlist__for_each_entry (session->evlist, evsel) { +if (evsel->core.attr.context_switch) +return true; + } + +return false; +} +int map__set_kallsyms_ref_reloc_sym (struct map *map,const char *symbol_name,u64 addr) {char *bracket;
tools/perf/util/session.h +1 -0
@@ -141,6 +141,7 @@ int perf_session__resolve_callchain(struct perf_session *session, struct symbol **parent);bool perf_session__has_traces (struct perf_session *session,const char *msg); +bool perf_session__has_switch_events (struct perf_session *session);void perf_event__attr_swap (struct perf_event_attr *attr);
tools/perf/util/sort.c +23 -0
@@ -892,6 +892,27 @@ struct sort_entry sort_cpu = { .se_width_idx = HISTC_CPU, }; +/* --sort parallelism */ + +static int64_t +sort__parallelism_cmp (struct hist_entry *left,struct hist_entry *right) +{ +return right->parallelism - left->parallelism; +} + +static int hist_entry__parallelism_snprintf (struct hist_entry *he,char *bf, +size_t size,unsigned int width) +{ +return repsep_snprintf (bf, size,"%*d" , width, he->parallelism); +} + +struct sort_entry sort_parallelism = { + .se_header ="Parallelism" , + .se_cmp = sort__parallelism_cmp, + .se_snprintf = hist_entry__parallelism_snprintf, + .se_width_idx =HISTC_PARALLELISM , +}; +/* --sort cgroup_id */ static int64_t _sort__cgroup_dev_cmp (u64 left_dev,u64 right_dev)
@@ -2534,6 +2555,7 @@ static struct sort_dimension common_sort_dimensions[] = {DIM (SORT_ANNOTATE_DATA_TYPE_OFFSET ,"typeoff" , sort_type_offset),DIM (SORT_SYM_OFFSET ,"symoff" , sort_sym_offset),DIM (SORT_ANNOTATE_DATA_TYPE_CACHELINE ,"typecln" , sort_type_cacheline), + DIM(SORT_PARALLELISM, "parallelism", sort_parallelism), };#undef DIM
@@ -2735,6 +2757,7 @@ MK_SORT_ENTRY_CHK(thread) MK_SORT_ENTRY_CHK(comm) MK_SORT_ENTRY_CHK(dso) MK_SORT_ENTRY_CHK(sym) +MK_SORT_ENTRY_CHK(parallelism) static bool __sort__hpp_equal(struct perf_hpp_fmt *a, struct perf_hpp_fmt *b)
tools/perf/util/sort.h +1 -0
@@ -72,6 +72,7 @@ enum sort_type {SORT_ANNOTATE_DATA_TYPE_OFFSET ,SORT_SYM_OFFSET ,SORT_ANNOTATE_DATA_TYPE_CACHELINE , +SORT_PARALLELISM ,/* branch stack specific sort keys */ __SORT_BRANCH_STACK,
Signature · 1 line
2.48.1.502.g6dc24dfdaf-goog
PATCH v7 3/9 perf report: Switch filtered from u8 to u16
In reply to Dmitry VyukovWe already have all u8 bits taken, adding one more filter leads to unpleasant failure mode, where code compiles w/o warnings, but the last filters silently don't work. Add a typedef and switch to u16.
--- tools/perf/util/addr_location.h | 3 ++- tools/perf/util/hist.c | 2 +- tools/perf/util/hist.h | 4 +++- 3 files changed, 6 insertions(+), 3 deletions(-)
tools/perf/util/addr_location.h +2 -1
@@ -3,6 +3,7 @@#define __PERF_ADDR_LOCATION 1#include <linux/types.h> +#include "hist.h" struct thread ;struct maps ;
@@ -17,7 +18,7 @@ struct addr_location {const char *srcline;u64 addr;char level; -u8 filtered; +filter_mask_t filtered;u8 cpumode;s32 cpu;s32 socket;
tools/perf/util/hist.c +1 -1
@@ -585,7 +585,7 @@ static struct hist_entry *hist_entry__new(struct hist_entry *template,return he; } -static u8 symbol__parent_filter (const struct symbol *parent) +static filter_mask_t symbol__parent_filter (const struct symbol *parent) {if (symbol_conf.exclude_other && parent ==NULL )return 1 <<HIST_FILTER__PARENT ;
tools/perf/util/hist.h +3 -1
@@ -33,6 +33,8 @@ enum hist_filter { HIST_FILTER__C2C, }; +typedef u16 filter_mask_t ; +enum hist_column {HISTC_SYMBOL ,HISTC_TIME ,
@@ -244,7 +246,7 @@ struct hist_entry {bool leaf;char level; -u8 filtered; +filter_mask_t filtered;u16 callchain_size;union {
Signature · 1 line
2.48.1.502.g6dc24dfdaf-goog
Namhyung Kim
maintainerRe: [PATCH v7 3/9] perf report: Switch filtered from u8 to u16
Quoted 29 lines · Dmitry Vyukov
We already have all u8 bits taken, adding one more filter leads to unpleasant failure mode, where code compiles w/o warnings, but the last filters silently don't work. Add a typedef and switch to u16.
Signed-off-by Dmitry Vyukov <[email protected]>Cc Namhyung Kim <[email protected]>Cc Arnaldo Carvalho de Melo <[email protected]>Cc Ian Rogers <[email protected]>--- tools/perf/util/addr_location.h | 3 ++- tools/perf/util/hist.c | 2 +- tools/perf/util/hist.h | 4 +++- 3 files changed, 6 insertions(+), 3 deletions(-)tools/perf/util/addr_location.h +1
-0@@ -3,6 +3,7 @@#define __PERF_ADDR_LOCATION 1#include <linux/types.h> +#include "hist.h" struct thread ;struct maps ;
tools/perf/util/addr_location.h struct addr_location { · line 24
Quoted 8 lines
tools/perf/util/addr_location.h +1
-1@@ -17,7 +18,7 @@ struct addr_location {const char *srcline;u64 addr;char level; -u8 filtered; +filter_mask_t filtered;u8 cpumode;s32 cpu;s32 socket;
This change introduced a build failure on Fedora 40 with libcapstone like below.
In file included from /usr/include/capstone/capstone.h:325,
from util/disasm.c:1333:
/usr/include/capstone/bpf.h:94:14: error: 'bpf_insn' defined as wrong kind of tag
94 | typedef enum bpf_insn {
| ^~~~~~~~ make[4]: *** [/linux/tools/build/Makefile.build:86: /build/util/disasm.o] Error 1 make[4]: *** Waiting for unfinished jobs....
make[3]: *** [/linux/tools/build/Makefile.build:138: util] Error 2 make[2]: *** [Makefile.perf:822: /build/perf-util-in.o] Error 2 make[2]: *** Waiting for unfinished jobs....
make[1]: *** [Makefile.perf:321: sub-make] Error 2
make: *** [Makefile:76: all] Error 2
I think we need the below change. I'll fold it.
Thanks,
Namhyung
---8<---
tools/perf/util/addr_location.h +1 -2
@@ -3,7 +3,6 @@#define __PERF_ADDR_LOCATION 1#include <linux/types.h> -#include "hist.h" struct thread ;struct maps ;
@@ -18,8 +17,8 @@ struct addr_location {const char *srcline;u64 addr;char level; -filter_mask_t filtered;u8 cpumode; +u16 filtered;s32 cpu;s32 socket;/* Same as machine.parallelism but within [1, nr_cpus]. */
PATCH v7 4/9 perf report: Add parallelism filter
In reply to Dmitry VyukovAdd parallelism filter that can be used to look at specific parallelism levels only. The format is the same as cpu lists. For example:
Only single-threaded samples: --parallelism=1 Low parallelism only: --parallelism=1-4 High parallelism only: --parallelism=64-128
--- tools/perf/builtin-report.c | 5 ++++- tools/perf/util/event.c | 2 ++ tools/perf/util/hist.c | 31 +++++++++++++++++++++++++++++++ tools/perf/util/hist.h | 6 +++++- tools/perf/util/sort.c | 11 +++++++++++ tools/perf/util/symbol.c | 34 ++++++++++++++++++++++++++++++++++ tools/perf/util/symbol_conf.h | 4 ++++ 7 files changed, 91 insertions(+), 2 deletions(-)
tools/perf/builtin-report.c +4 -1
@@ -1390,6 +1390,8 @@ int cmd_report(int argc, const char **argv) symbol__config_symfs),OPT_STRING ('C' ,"cpu" , &report.cpu_list,"cpu" ,"list of cpus to profile" ), +OPT_STRING (0 ,"parallelism" , &symbol_conf.parallelism_list_str,"parallelism" , +"only consider these parallelism levels (cpu set format)" ),OPT_BOOLEAN ('I' ,"show-info" , &report.show_full_info,"Display extended information about perf.data file" ), OPT_BOOLEAN(0, "source", &annotate_opts.annotate_src,
@@ -1721,7 +1723,8 @@ int cmd_report(int argc, const char **argv) }if (report.disable_order || !perf_session__has_switch_events (session)) { -if ((sort_order &&strstr (sort_order,"parallelism" )) || +if (symbol_conf.parallelism_list_str || + (sort_order &&strstr (sort_order,"parallelism" )) || (field_order &&strstr (field_order,"parallelism" ))) {if (report.disable_order)ui__error ("Use of parallelism is incompatible with --disable-order.\n" );
tools/perf/util/event.c +2 -0
@@ -769,6 +769,8 @@ int machine__resolve(struct machine *machine, struct addr_location *al,/* Account for possible out-of-order switch events. */ al->parallelism =max (1 ,min (machine->parallelism,machine__nr_cpus_avail (machine))); +if (test_bit (al->parallelism, symbol_conf.parallelism_filter)) + al->filtered |= (1 <<HIST_FILTER__PARALLELISM );if (al->map) { if (symbol_conf.dso_list &&
tools/perf/util/hist.c +31 -0
@@ -43,6 +43,8 @@ static bool hists__filter_entry_by_symbol(struct hists *hists, struct hist_entry *he);static bool hists__filter_entry_by_socket (struct hists *hists,struct hist_entry *he); +static bool hists__filter_entry_by_parallelism (struct hists *hists, +struct hist_entry *he);u16 hists__col_len (struct hists *hists,enum hist_column col) {
@@ -1457,6 +1459,10 @@ static void hist_entry__check_and_remove_filter(struct hist_entry *he,if (symbol_conf.sym_list ==NULL )return ;break ; +case HIST_FILTER__PARALLELISM : +if (__bitmap_weight (symbol_conf.parallelism_filter,MAX_NR_CPUS +1 ) ==0 ) +return ; +break ;case HIST_FILTER__PARENT :case HIST_FILTER__GUEST :case HIST_FILTER__HOST :
@@ -1515,6 +1521,9 @@ static void hist_entry__apply_hierarchy_filters(struct hist_entry *he)hist_entry__check_and_remove_filter (he,HIST_FILTER__SYMBOL , perf_hpp__is_sym_entry); +hist_entry__check_and_remove_filter (he,HIST_FILTER__PARALLELISM , + perf_hpp__is_parallelism_entry); +hists__apply_filters (he->hists, he); }
@@ -1711,6 +1720,7 @@ static void hists__apply_filters(struct hists *hists, struct hist_entry *he)hists__filter_entry_by_thread (hists, he);hists__filter_entry_by_symbol (hists, he);hists__filter_entry_by_socket (hists, he); +hists__filter_entry_by_parallelism (hists, he); }int hists__collapse_resort (struct hists *hists,struct ui_progress *prog)
@@ -2197,6 +2207,16 @@ static bool hists__filter_entry_by_socket(struct hists *hists,return false; } +static bool hists__filter_entry_by_parallelism (struct hists *hists, +struct hist_entry *he) +{ +if (test_bit (he->parallelism, hists->parallelism_filter)) { + he->filtered |= (1 <<HIST_FILTER__PARALLELISM ); +return true; + } +return false; +} +typedef bool (*filter_fn_t )(struct hists *hists,struct hist_entry *he);static void hists__filter_by_type (struct hists *hists,int type,filter_fn_t filter)
@@ -2366,6 +2386,16 @@ void hists__filter_by_socket(struct hists *hists) hists__filter_entry_by_socket); } +void hists__filter_by_parallelism (struct hists *hists) +{ +if (symbol_conf.report_hierarchy) +hists__filter_hierarchy (hists,HIST_FILTER__PARALLELISM , + hists->parallelism_filter); +else +hists__filter_by_type (hists,HIST_FILTER__PARALLELISM , + hists__filter_entry_by_parallelism); +} +void events_stats__inc (struct events_stats *stats,u32 type) { ++stats->nr_events[0 ];
@@ -2872,6 +2902,7 @@ int __hists__init(struct hists *hists, struct perf_hpp_list *hpp_list) hists->entries =RB_ROOT_CACHED ;mutex_init (&hists->lock); hists->socket_filter =-1 ; + hists->parallelism_filter = symbol_conf.parallelism_filter; hists->hpp_list = hpp_list;INIT_LIST_HEAD (&hists->hpp_formats);return 0 ;
tools/perf/util/hist.h +5 -1
@@ -31,6 +31,7 @@ enum hist_filter {HIST_FILTER__HOST ,HIST_FILTER__SOCKET ,HIST_FILTER__C2C , + HIST_FILTER__PARALLELISM, };typedef u16 filter_mask_t ;
@@ -112,6 +113,7 @@ struct hists {const struct dso *dso_filter;const char *uid_filter_str;const char *symbol_filter_str; +unsigned long *parallelism_filter;struct mutex lock;struct hists_stats stats;u64 event_stream;
@@ -388,11 +390,13 @@ void hists__filter_by_dso(struct hists *hists);void hists__filter_by_thread (struct hists *hists);void hists__filter_by_symbol (struct hists *hists);void hists__filter_by_socket (struct hists *hists); +void hists__filter_by_parallelism (struct hists *hists);static inline bool hists__has_filter (struct hists *hists) {return hists->thread_filter || hists->dso_filter || - hists->symbol_filter_str || (hists->socket_filter >-1 ); + hists->symbol_filter_str || (hists->socket_filter >-1 ) || + hists->parallelism_filter; }u16 hists__col_len (struct hists *hists,enum hist_column col);
tools/perf/util/sort.c +11 -0
@@ -900,6 +900,16 @@ sort__parallelism_cmp(struct hist_entry *left, struct hist_entry *right)return right->parallelism - left->parallelism; } +static int hist_entry__parallelism_filter (struct hist_entry *he,int type,const void *arg) +{ +const unsigned long *parallelism_filter = arg; + +if (type !=HIST_FILTER__PARALLELISM ) +return -1 ; + +return test_bit (he->parallelism, parallelism_filter); +} +static int hist_entry__parallelism_snprintf (struct hist_entry *he,char *bf,size_t size,unsigned int width) {
@@ -909,6 +919,7 @@ static int hist_entry__parallelism_snprintf(struct hist_entry *he, char *bf,struct sort_entry sort_parallelism = { .se_header ="Parallelism" , .se_cmp = sort__parallelism_cmp, + .se_filter = hist_entry__parallelism_filter, .se_snprintf = hist_entry__parallelism_snprintf, .se_width_idx =HISTC_PARALLELISM , };
tools/perf/util/symbol.c +34 -0
@@ -18,6 +18,7 @@#include "annotate.h" #include "build-id.h" #include "cap.h" +#include "cpumap.h" #include "dso.h" #include "util.h" // lsdir() #include "debug.h"
@@ -2471,6 +2472,36 @@ int symbol__annotation_init(void)return 0 ; } +static int setup_parallelism_bitmap (void ) +{ +struct perf_cpu_map *map; +struct perf_cpu cpu; +int i, err =-1 ; + +if (symbol_conf.parallelism_list_str ==NULL ) +return 0 ; + + map =perf_cpu_map__new (symbol_conf.parallelism_list_str); +if (map ==NULL ) { +pr_err ("failed to parse parallelism filter list\n" ); +return -1 ; + } + +bitmap_fill (symbol_conf.parallelism_filter,MAX_NR_CPUS +1 ); +perf_cpu_map__for_each_cpu (cpu, i, map) { +if (cpu.cpu <=0 || cpu.cpu >MAX_NR_CPUS ) { +pr_err ("Requested parallelism level %d is invalid.\n" , cpu.cpu); + goto out_delete_map; + } +__clear_bit (cpu.cpu, symbol_conf.parallelism_filter); + } + + err =0 ; +out_delete_map: +perf_cpu_map__put (map); +return err; +} +int symbol__init (struct perf_env *env) {const char *symfs;
@@ -2490,6 +2521,9 @@ int symbol__init(struct perf_env *env)return -1 ; } +if (setup_parallelism_bitmap ()) +return -1 ; +if (setup_list (&symbol_conf.dso_list, symbol_conf.dso_list_str,"dso" ) <0 )return -1 ;
tools/perf/util/symbol_conf.h +4 -0
@@ -3,6 +3,8 @@#define __PERF_SYMBOL_CONF 1#include <stdbool.h> +#include <linux/bitmap.h> +#include "perf.h" struct strlist ;struct intlist ;
@@ -62,6 +64,7 @@ struct symbol_conf { *pid_list_str, *tid_list_str, *sym_list_str, + *parallelism_list_str, *col_width_list_str, *bt_stop_list_str;const char *addr2line_path;
@@ -82,6 +85,7 @@ struct symbol_conf {int pad_output_len_dso;int group_sort_idx;int addr_range; +DECLARE_BITMAP (parallelism_filter,MAX_NR_CPUS +1 ); };extern struct symbol_conf symbol_conf;
Signature · 1 line
2.48.1.502.g6dc24dfdaf-goog
PATCH v7 5/9 perf report: Add latency output field
In reply to Dmitry VyukovLatency output field is similar to overhead, but represents overhead for latency rather than CPU consumption. It's re-scaled from overhead by dividing weight by the current parallelism level at the time of the sample. It effectively models profiling with 1 sample taken per unit of wall-clock time rather than unit of CPU time.
---
Changes in v5:
- fixed formatting of latency field in --stdout mode
--- tools/perf/ui/browsers/hists.c | 27 ++++++++----- tools/perf/ui/hist.c | 69 ++++++++++++++++++--------------- tools/perf/util/addr_location.h | 2 + tools/perf/util/event.c | 6 +++ tools/perf/util/events_stats.h | 2 + tools/perf/util/hist.c | 55 +++++++++++++++++++------- tools/perf/util/hist.h | 12 ++++++ tools/perf/util/sort.c | 2 + 8 files changed, 120 insertions(+), 55 deletions(-)
tools/perf/ui/browsers/hists.c +17 -10
@@ -1226,7 +1226,7 @@ int __hpp__slsmg_color_printf(struct perf_hpp *hpp, const char *fmt, ...)return ret; } -#define __HPP_COLOR_PERCENT_FN (_type, _field) \ +#define __HPP_COLOR_PERCENT_FN (_type, _field, _fmttype) \ static u64 __hpp_get_##_field(struct hist_entry *he) \ { \ return he->stat._field; \
@@ -1238,10 +1238,10 @@ hist_browser__hpp_color_##_type(struct perf_hpp_fmt *fmt, \ struct hist_entry *he) \ { \ return hpp__fmt(fmt, hpp, he, __hpp_get_##_field, " %*.2f%%", \ - __hpp__slsmg_color_printf, true); \ + __hpp__slsmg_color_printf, _fmttype); \ } -#define __HPP_COLOR_ACC_PERCENT_FN (_type, _field) \ +#define __HPP_COLOR_ACC_PERCENT_FN (_type, _field, _fmttype) \ static u64 __hpp_get_acc_##_field(struct hist_entry *he) \ { \ return he->stat_acc->_field; \
@@ -1262,15 +1262,18 @@ hist_browser__hpp_color_##_type(struct perf_hpp_fmt *fmt, \return ret; \ } \ return hpp__fmt(fmt, hpp, he, __hpp_get_acc_##_field, \ -" %*.2f%%" , __hpp__slsmg_color_printf, true); \ +" %*.2f%%" , __hpp__slsmg_color_printf, \ + _fmttype); \ } -__HPP_COLOR_PERCENT_FN(overhead, period) -__HPP_COLOR_PERCENT_FN(overhead_sys, period_sys) -__HPP_COLOR_PERCENT_FN(overhead_us, period_us) -__HPP_COLOR_PERCENT_FN(overhead_guest_sys, period_guest_sys) -__HPP_COLOR_PERCENT_FN(overhead_guest_us, period_guest_us) -__HPP_COLOR_ACC_PERCENT_FN(overhead_acc, period) +__HPP_COLOR_PERCENT_FN(overhead, period, PERF_HPP_FMT_TYPE__PERCENT) +__HPP_COLOR_PERCENT_FN(latency, latency, PERF_HPP_FMT_TYPE__LATENCY) +__HPP_COLOR_PERCENT_FN(overhead_sys, period_sys, PERF_HPP_FMT_TYPE__PERCENT) +__HPP_COLOR_PERCENT_FN(overhead_us, period_us, PERF_HPP_FMT_TYPE__PERCENT) +__HPP_COLOR_PERCENT_FN(overhead_guest_sys, period_guest_sys, PERF_HPP_FMT_TYPE__PERCENT) +__HPP_COLOR_PERCENT_FN(overhead_guest_us, period_guest_us, PERF_HPP_FMT_TYPE__PERCENT) +__HPP_COLOR_ACC_PERCENT_FN(overhead_acc, period, PERF_HPP_FMT_TYPE__PERCENT) +__HPP_COLOR_ACC_PERCENT_FN(latency_acc, latency, PERF_HPP_FMT_TYPE__LATENCY)#undef __HPP_COLOR_PERCENT_FN#undef __HPP_COLOR_ACC_PERCENT_FN
@@ -1279,6 +1282,8 @@ void hist_browser__init_hpp(void) { perf_hpp__format[PERF_HPP__OVERHEAD ].color = hist_browser__hpp_color_overhead; + perf_hpp__format[PERF_HPP__LATENCY ].color = + hist_browser__hpp_color_latency; perf_hpp__format[PERF_HPP__OVERHEAD_SYS ].color = hist_browser__hpp_color_overhead_sys; perf_hpp__format[PERF_HPP__OVERHEAD_US].color =
@@ -1289,6 +1294,8 @@ void hist_browser__init_hpp(void) hist_browser__hpp_color_overhead_guest_us; perf_hpp__format[PERF_HPP__OVERHEAD_ACC ].color = hist_browser__hpp_color_overhead_acc; + perf_hpp__format[PERF_HPP__LATENCY_ACC ].color = + hist_browser__hpp_color_latency_acc;res_sample_init (); }
tools/perf/ui/hist.c +38 -31
177 lines not on this page · raw message
tools/perf/util/addr_location.h +2 -0
@@ -24,6 +24,8 @@ struct addr_location {s32 socket;/* Same as machine.parallelism but within [1, nr_cpus]. */ int parallelism; +/* See he_stat.latency. */ +u64 latency; };void addr_location__init (struct addr_location *al);
tools/perf/util/event.c +6 -0
@@ -771,6 +771,12 @@ int machine__resolve(struct machine *machine, struct addr_location *al, al->parallelism =max (1 ,min (machine->parallelism,machine__nr_cpus_avail (machine)));if (test_bit (al->parallelism, symbol_conf.parallelism_filter)) al->filtered |= (1 <<HIST_FILTER__PARALLELISM ); +/* +* Multiply it by some const to avoid precision loss or dealing +* with floats. The multiplier does not matter otherwise since +* we only print it as percents. +*/ + al->latency = sample->period *1000 / al->parallelism;if (al->map) { if (symbol_conf.dso_list &&
tools/perf/util/events_stats.h +2 -0
@@ -57,6 +57,8 @@ struct events_stats {struct hists_stats {u64 total_period;u64 total_non_filtered_period; +u64 total_latency; +u64 total_non_filtered_latency;u32 nr_samples;u32 nr_non_filtered_samples;u32 nr_lost_samples;
tools/perf/util/hist.c +41 -14
175 lines not on this page · raw message
tools/perf/util/hist.h +12 -0
@@ -109,6 +109,8 @@ struct hists {u64 nr_non_filtered_entries;u64 callchain_period;u64 callchain_non_filtered_period; +u64 callchain_latency; +u64 callchain_non_filtered_latency;struct thread *thread_filter;const struct dso *dso_filter;const char *uid_filter_str;
@@ -170,6 +172,12 @@ struct res_sample {struct he_stat {u64 period; +/* +* Period re-scaled from CPU time to wall-clock time (divided by the +* parallelism at the time of the sample). This represents effect of +* the event on latency rather than CPU consumption. +*/ +u64 latency;u64 period_sys;u64 period_us;u64 period_guest_sys;
@@ -374,6 +382,7 @@ void hists__output_recalc_col_len(struct hists *hists, int max_rows);struct hist_entry *hists__get_entry (struct hists *hists,int idx);u64 hists__total_period (struct hists *hists); +u64 hists__total_latency (struct hists *hists);void hists__reset_stats (struct hists *hists);void hists__inc_stats (struct hists *hists,struct hist_entry *h);void hists__inc_nr_events (struct hists *hists);
@@ -555,11 +564,13 @@ extern struct perf_hpp_fmt perf_hpp__format[];enum {/* Matches perf_hpp__format array. */ PERF_HPP__OVERHEAD , +PERF_HPP__LATENCY ,PERF_HPP__OVERHEAD_SYS ,PERF_HPP__OVERHEAD_US ,PERF_HPP__OVERHEAD_GUEST_SYS ,PERF_HPP__OVERHEAD_GUEST_US ,PERF_HPP__OVERHEAD_ACC , +PERF_HPP__LATENCY_ACC ,PERF_HPP__SAMPLES ,PERF_HPP__PERIOD ,PERF_HPP__WEIGHT1 ,
@@ -615,6 +626,7 @@ void hists__reset_column_width(struct hists *hists);enum perf_hpp_fmt_type {PERF_HPP_FMT_TYPE__RAW ,PERF_HPP_FMT_TYPE__PERCENT , +PERF_HPP_FMT_TYPE__LATENCY ,PERF_HPP_FMT_TYPE__AVERAGE , };
tools/perf/util/sort.c +2 -0
@@ -2628,11 +2628,13 @@ struct hpp_dimension {static struct hpp_dimension hpp_sort_dimensions[] = {DIM (PERF_HPP__OVERHEAD ,"overhead" ), +DIM (PERF_HPP__LATENCY ,"latency" ),DIM (PERF_HPP__OVERHEAD_SYS ,"overhead_sys" ),DIM (PERF_HPP__OVERHEAD_US ,"overhead_us" ),DIM (PERF_HPP__OVERHEAD_GUEST_SYS ,"overhead_guest_sys" ),DIM (PERF_HPP__OVERHEAD_GUEST_US ,"overhead_guest_us" ),DIM (PERF_HPP__OVERHEAD_ACC ,"overhead_children" ), +DIM (PERF_HPP__LATENCY_ACC ,"latency_children" ),DIM (PERF_HPP__SAMPLES ,"sample" ),DIM (PERF_HPP__PERIOD ,"period" ),DIM (PERF_HPP__WEIGHT1 ,"weight1" ),
Signature · 1 line
2.48.1.502.g6dc24dfdaf-goog
PATCH v7 6/9 perf report: Add --latency flag
In reply to Dmitry VyukovAdd record/report --latency flag that allows to capture and show latency-centric profiles rather than the default CPU-consumption-centric profiles. For latency profiles record captures context switch events, and report shows Latency as the first column.
---
Changes in v7:
- added comment perf_hpp__init() re was_taken logic
Changes in v6:
- remove latency column in perf_hpp__cancel_latency if sort order is specified, but does not include latency
Changes in v5:
- added description of --latency flag in Documentation flags
--- tools/perf/Documentation/perf-record.txt | 4 ++ tools/perf/Documentation/perf-report.txt | 5 +++ tools/perf/builtin-record.c | 20 +++++++++ tools/perf/builtin-report.c | 32 ++++++++++++-- tools/perf/ui/hist.c | 54 ++++++++++++++++++++---- tools/perf/util/hist.h | 1 + tools/perf/util/sort.c | 33 ++++++++++++--- tools/perf/util/sort.h | 2 +- tools/perf/util/symbol_conf.h | 4 +- 9 files changed, 135 insertions(+), 20 deletions(-)
tools/perf/Documentation/perf-record.txt +4 -0
@@ -227,6 +227,10 @@ OPTIONS '--filter' exists, the new filter expression will be combined with them by '&&'. +--latency:: + Enable data collection for latency profiling. + Use perf report --latency for latency-centric profile. + -a:: --all-cpus:: System-wide collection from all CPUs (default if no target is specified).
tools/perf/Documentation/perf-report.txt +5 -0
@@ -68,6 +68,11 @@ OPTIONS --hide-unresolved:: Only display entries resolved to a symbol. +--latency:: + Show latency-centric profile rather than the default + CPU-consumption-centric profile + (requires perf record --latency flag). + -s:: --sort=:: Sort histogram entries by given key(s) - multiple keys can be specified
tools/perf/builtin-record.c +20 -0
@@ -161,6 +161,7 @@ struct record {struct evlist *sb_evlist;pthread_t thread_id;int realtime_prio; +bool latency;bool switch_output_event_set;bool no_buildid;bool no_buildid_set;
@@ -3371,6 +3372,9 @@ static struct option __record_options[] = { parse_events_option),OPT_CALLBACK (0 ,"filter" , &record.evlist,"filter" ,"event filter" , parse_filter), +OPT_BOOLEAN (0 ,"latency" , &record.latency, +"Enable data collection for latency profiling.\n" +"\t\t\t Use perf report --latency for latency-centric profile." ),OPT_CALLBACK_NOOPT (0 ,"exclude-perf" , &record.evlist,NULL ,"don't record events from perf itself" , exclude_perf),
@@ -4017,6 +4021,22 @@ int cmd_record(int argc, const char **argv) } +if (record.latency) { +/* +* There is no fundamental reason why latency profiling +* can't work for system-wide mode, but exact semantics +* and details are to be defined. +* See the following thread for details: +* https://lore.kernel.org/all/[email protected]/ +*/ +if (record.opts.target.system_wide) { +pr_err ("Failed: latency profiling is not supported with system-wide collection.\n" ); + err = -EINVAL ; + goto out_opts; + } + record.opts.record_switch_events = true; + } +if (rec->buildid_mmap) {if (!perf_can_record_build_id ()) {pr_err ("Failed: no support to record build id in mmap events, update your kernel.\n" );
tools/perf/builtin-report.c +28 -4
@@ -112,6 +112,8 @@ struct report {u64 nr_entries;u64 queue_size;u64 total_cycles; +u64 total_samples; +u64 singlethreaded_samples;int socket_filter;DECLARE_BITMAP (cpu_bitmap,MAX_NR_CPUS );struct branch_type_stat brtype_stat;
@@ -331,6 +333,10 @@ static int process_sample_event(const struct perf_tool *tool, &rep->total_cycles, evsel); } + rep->total_samples++; +if (al.parallelism ==1 ) + rep->singlethreaded_samples++; + ret =hist_entry_iter__add (&iter, &al, rep->max_stack, rep);if (ret <0 )pr_debug ("problem adding hist entry, skipping event\n" );
@@ -1079,6 +1085,11 @@ static int __cmd_report(struct report *rep)return ret; } +/* Don't show Latency column for non-parallel profiles by default. */ +if (!symbol_conf.prefer_latency && rep->total_samples && + rep->singlethreaded_samples *100 / rep->total_samples >=99 ) +perf_hpp__cancel_latency (); +evlist__check_mem_load_aux (session->evlist);if (rep->stats_mode)
@@ -1468,6 +1479,10 @@ int cmd_report(int argc, const char **argv) "Disable raw trace ordering"),OPT_BOOLEAN (0 ,"skip-empty" , &report.skip_empty,"Do not display empty (or dummy) events in the output" ), +OPT_BOOLEAN (0 ,"latency" , &symbol_conf.prefer_latency, +"Show latency-centric profile rather than the default\n" +"\t\t\t CPU-consumption-centric profile\n" +"\t\t\t (requires perf record --latency flag)." ),OPT_END () }; struct perf_data data = {
@@ -1722,16 +1737,25 @@ int cmd_report(int argc, const char **argv) symbol_conf.annotate_data_sample = true; } + symbol_conf.enable_latency = true;if (report.disable_order || !perf_session__has_switch_events (session)) {if (symbol_conf.parallelism_list_str || - (sort_order &&strstr (sort_order,"parallelism" )) || - (field_order &&strstr (field_order,"parallelism" ))) { + symbol_conf.prefer_latency || + (sort_order && (strstr (sort_order,"latency" ) || +strstr (sort_order,"parallelism" ))) || + (field_order && (strstr (field_order,"latency" ) || +strstr (field_order,"parallelism" )))) {if (report.disable_order) -ui__error ("Use of parallelism is incompatible with --disable-order.\n" ); +ui__error ("Use of latency profile or parallelism is incompatible with --disable-order.\n" );else -ui__error ("Use of parallelism requires --switch-events during record.\n" ); +ui__error ("Use of latency profile or parallelism requires --latency flag during record.\n" );return -1 ; } +/* +* If user did not ask for anything related to +* latency/parallelism explicitly, just don't show it. +*/ + symbol_conf.enable_latency = false; }if (sort_order &&strstr (sort_order,"ipc" )) {
tools/perf/ui/hist.c +46 -8
80 lines not on this page · raw message
tools/perf/util/hist.h +1 -0
@@ -582,6 +582,7 @@ enum {void perf_hpp__init (void );void perf_hpp__cancel_cumulate (void ); +void perf_hpp__cancel_latency (void );void perf_hpp__setup_output_field (struct perf_hpp_list *list);void perf_hpp__reset_output_field (struct perf_hpp_list *list);void perf_hpp__append_sort_keys (struct perf_hpp_list *list);
tools/perf/util/sort.c +27 -6
@@ -2622,6 +2622,7 @@ struct hpp_dimension {const char *name;struct perf_hpp_fmt *fmt;int taken; +int was_taken; };#define DIM (d, n) { .name = n, .fmt = &perf_hpp__format[d], }
@@ -3513,6 +3514,7 @@ static int __hpp_dimension__add(struct hpp_dimension *hd,return -1 ; hd->taken =1 ; + hd->was_taken =1 ;perf_hpp_list__register_sort_field (list, fmt);return 0 ; }
@@ -3547,10 +3549,15 @@ static int __hpp_dimension__add_output(struct perf_hpp_list *list,return 0 ; } -int hpp_dimension__add_output (unsigned col) +int hpp_dimension__add_output (unsigned col,bool implicit) { +struct hpp_dimension *hd; +BUG_ON (col >=PERF_HPP__MAX_INDEX ); -return __hpp_dimension__add_output (&perf_hpp_list, &hpp_sort_dimensions[col]); + hd = &hpp_sort_dimensions[col]; +if (implicit && !hd->was_taken) +return 0 ; +return __hpp_dimension__add_output (&perf_hpp_list, hd); } int sort_dimension__add(struct perf_hpp_list *list, const char *tok,
@@ -3809,10 +3816,24 @@ static char *setup_overhead(char *keys)if (sort__mode ==SORT_MODE__DIFF )return keys; - keys =prefix_if_not_in ("overhead" , keys); - -if (symbol_conf.cumulate_callchain) - keys =prefix_if_not_in ("overhead_children" , keys); +if (symbol_conf.prefer_latency) { + keys =prefix_if_not_in ("overhead" , keys); + keys =prefix_if_not_in ("latency" , keys); +if (symbol_conf.cumulate_callchain) { + keys =prefix_if_not_in ("overhead_children" , keys); + keys =prefix_if_not_in ("latency_children" , keys); + } + }else if (!keys || (!strstr (keys,"overhead" ) && + !strstr (keys,"latency" ))) { +if (symbol_conf.enable_latency) + keys =prefix_if_not_in ("latency" , keys); + keys =prefix_if_not_in ("overhead" , keys); +if (symbol_conf.cumulate_callchain) { +if (symbol_conf.enable_latency) + keys =prefix_if_not_in ("latency_children" , keys); + keys =prefix_if_not_in ("overhead_children" , keys); + } + }return keys; }
tools/perf/util/sort.h +1 -1
@@ -141,7 +141,7 @@ int report_parse_ignore_callees_opt(const struct option *opt, const char *arg, ibool is_strict_order (const char *order); -int hpp_dimension__add_output (unsigned col); +int hpp_dimension__add_output (unsigned col,bool implicit);void reset_dimensions (void ); int sort_dimension__add(struct perf_hpp_list *list, const char *tok, struct evlist *evlist,
tools/perf/util/symbol_conf.h +3 -1
@@ -49,7 +49,9 @@ struct symbol_conf { keep_exited_threads, annotate_data_member, annotate_data_sample, - skip_empty; + skip_empty, + enable_latency, + prefer_latency;const char *vmlinux_name, *kallsyms_name, *source_prefix,
Signature · 1 line
2.48.1.502.g6dc24dfdaf-goog
PATCH v7 7/9 perf report: Add latency and parallelism profiling documentation
In reply to Dmitry VyukovDescribe latency and parallelism profiling, related flags, and differences with the currently only supported CPU-consumption-centric profiling.
---
Changes in v7:
- add tips.txt example with time/time-quantum
--- .../callchain-overhead-calculation.txt | 5 +- .../cpu-and-latency-overheads.txt | 85 +++++++++++++++++++ tools/perf/Documentation/perf-report.txt | 49 +++++++---- tools/perf/Documentation/tips.txt | 4 + 4 files changed, 124 insertions(+), 19 deletions(-)
tools/perf/Documentation/callchain-overhead-calculation.txt +3 -2
@@ -1,7 +1,8 @@ Overhead calculation -------------------- -The overhead can be shown in two columns as 'Children' and 'Self' when -perf collects callchains. The 'self' overhead is simply calculated by +The CPU overhead can be shown in two columns as 'Children' and 'Self' +when perf collects callchains (and corresponding 'Wall' columns for +wall-clock overhead). The 'self' overhead is simply calculated by adding all period values of the entry - usually a function (symbol). This is the value that perf shows traditionally and sum of all the 'self' overhead values should be 100%.
tools/perf/Documentation/cpu-and-latency-overheads.txt new file +85 -0
85 lines not on this page · raw message
tools/perf/Documentation/perf-report.txt +32 -17
114 lines not on this page · raw message
tools/perf/Documentation/tips.txt +4 -0
@@ -62,3 +62,7 @@ To show context switches in perf report sample context add --switch-events to pe To show time in nanoseconds in record/report add --ns To compare hot regions in two workloads use perf record -b -o file ... ; perf diff --stream file1 file2 To compare scalability of two workload samples use perf diff -c ratio file1 file2 +For latency profiling, try: perf record/report --latency +For parallelism histogram, try: perf report --hierarchy --sort latency,parallelism,comm,symbol +To analyze particular parallelism levels, try: perf report --latency --parallelism=32-64 +To see how parallelism changes over time, try: perf report -F time,latency,parallelism --time-quantum=1s
Signature · 1 line
2.48.1.502.g6dc24dfdaf-goog
PATCH v7 8/9 perf test: Add tests for latency and parallelism profiling
In reply to Dmitry VyukovEnsure basic operation of latency/parallelism profiling and that main latency/parallelism record/report invocations don't fail/crash.
--- tools/perf/tests/shell/base_report/setup.sh | 18 ++++++- .../tests/shell/base_report/test_basic.sh | 52 +++++++++++++++++++ 2 files changed, 69 insertions(+), 1 deletion(-)
tools/perf/tests/shell/base_report/setup.sh +17 -1
@@ -15,6 +15,8 @@ # include working environment . ../common/init.sh +TEST_RESULT=0 + test -d "$HEADER_TAR_DIR" || mkdir -p "$HEADER_TAR_DIR" SW_EVENT="cpu-clock"
@@ -26,7 +28,21 @@ PERF_EXIT_CODE=$? CHECK_EXIT_CODE=$? print_results $PERF_EXIT_CODE $CHECK_EXIT_CODE "prepare the perf.data file" -TEST_RESULT=$? +(( TEST_RESULT += $? )) + +# Some minimal parallel workload. +$CMD_PERF record --latency -o $CURRENT_TEST_DIR/perf.data.1 bash -c "for i in {1..100} ; do cat /proc/cpuinfo 1> /dev/null & done; sleep 1" 2> $LOGS_DIR/setup-latency.log +PERF_EXIT_CODE=$? + +echo ================== +cat $LOGS_DIR/setup-latency.log +echo ================== + +../common/check_all_patterns_found.pl "$RE_LINE_RECORD1" "$RE_LINE_RECORD2" < $LOGS_DIR/setup-latency.log +CHECK_EXIT_CODE=$? + +print_results $PERF_EXIT_CODE $CHECK_EXIT_CODE "prepare the perf.data.1 file" +(( TEST_RESULT += $? )) print_overall_results $TEST_RESULT exit $?
tools/perf/tests/shell/base_report/test_basic.sh +52 -0
58 lines not on this page · raw message
Signature · 1 line
2.48.1.502.g6dc24dfdaf-goog
PATCH v7 9/9 perf hist: Shrink struct hist_entry size
In reply to Dmitry VyukovReorder the struct fields by size to reduce paddings and reduce struct simd_flags size from 8 to 1 byte.
This reduces struct hist_entry size by 8 bytes (592->584), and leaves a single more usable 6 byte padding hole.
---
Pahole output before:
struct hist_entry {
struct rb_node rb_node_in __attribute__((__aligned__(8))); /* 0 24 */
struct rb_node rb_node __attribute__((__aligned__(8))); /* 24 24 */
union {
struct list_head node; /* 48 16 */
struct list_head head; /* 48 16 */
} pairs; /* 48 16 */
/* --- cacheline 1 boundary (64 bytes) --- */
struct he_stat stat; /* 64 80 */
/* XXX last struct has 4 bytes of padding */
/* --- cacheline 2 boundary (128 bytes) was 16 bytes ago --- */
struct he_stat * stat_acc; /* 144 8 */
struct map_symbol ms; /* 152 24 */
struct thread * thread; /* 176 8 */
struct comm * comm; /* 184 8 */
/* --- cacheline 3 boundary (192 bytes) --- */
struct namespace_id cgroup_id; /* 192 16 */
u64 cgroup; /* 208 8 */
u64 ip; /* 216 8 */
u64 transaction; /* 224 8 */
s32 socket; /* 232 4 */
s32 cpu; /* 236 4 */
int parallelism; /* 240 4 */
/* XXX 4 bytes hole, try to pack */
u64 code_page_size; /* 248 8 */
/* --- cacheline 4 boundary (256 bytes) --- */
u64 weight; /* 256 8 */
u64 ins_lat; /* 264 8 */
u64 p_stage_cyc; /* 272 8 */
u8 cpumode; /* 280 1 */
u8 depth; /* 281 1 */
/* XXX 2 bytes hole, try to pack */
int mem_type_off; /* 284 4 */
struct simd_flags simd_flags; /* 288 8 */
_Bool dummy; /* 296 1 */
_Bool leaf; /* 297 1 */
char level; /* 298 1 */
/* XXX 1 byte hole, try to pack */
filter_mask_t filtered; /* 300 2 */
u16 callchain_size; /* 302 2 */
union {
struct hist_entry_diff diff; /* 304 120 */
struct {
u16 row_offset; /* 304 2 */
u16 nr_rows; /* 306 2 */
_Bool init_have_children; /* 308 1 */
_Bool unfolded; /* 309 1 */
_Bool has_children; /* 310 1 */
_Bool has_no_entry; /* 311 1 */
}; /* 304 8 */
}; /* 304 120 */
/* --- cacheline 6 boundary (384 bytes) was 40 bytes ago --- */
char * srcline; /* 424 8 */
char * srcfile; /* 432 8 */
struct symbol * parent; /* 440 8 */
/* --- cacheline 7 boundary (448 bytes) --- */
struct branch_info * branch_info; /* 448 8 */
long int time; /* 456 8 */
struct hists * hists; /* 464 8 */
struct mem_info * mem_info; /* 472 8 */
struct block_info * block_info; /* 480 8 */
struct kvm_info * kvm_info; /* 488 8 */
void * raw_data; /* 496 8 */
u32 raw_size; /* 504 4 */
int num_res; /* 508 4 */
/* --- cacheline 8 boundary (512 bytes) --- */
struct res_sample * res_samples; /* 512 8 */
void * trace_output; /* 520 8 */
struct perf_hpp_list * hpp_list; /* 528 8 */
struct hist_entry * parent_he; /* 536 8 */
struct hist_entry_ops * ops; /* 544 8 */
struct annotated_data_type * mem_type; /* 552 8 */
union {
struct {
struct rb_root_cached hroot_in; /* 560 16 */
/* --- cacheline 9 boundary (576 bytes) --- */
struct rb_root_cached hroot_out; /* 576 16 */
}; /* 560 32 */
struct rb_root sorted_chain; /* 560 8 */
}; /* 560 32 */
/* --- cacheline 9 boundary (576 bytes) was 16 bytes ago --- */
struct callchain_root callchain[] __attribute__((__aligned__(8))); /* 592 0 */
/* size: 592, cachelines: 10, members: 49 */ /* sum members: 585, holes: 3, sum holes: 7 */ /* paddings: 1, sum paddings: 4 */
/* forced alignments: 3 */
/* last cacheline: 16 bytes */
} __attribute__((__aligned__(8)));
After:
struct hist_entry {
struct rb_node rb_node_in __attribute__((__aligned__(8))); /* 0 24 */
struct rb_node rb_node __attribute__((__aligned__(8))); /* 24 24 */
union {
struct list_head node; /* 48 16 */
struct list_head head; /* 48 16 */
} pairs; /* 48 16 */
/* --- cacheline 1 boundary (64 bytes) --- */
struct he_stat stat; /* 64 80 */
/* XXX last struct has 4 bytes of padding */
/* --- cacheline 2 boundary (128 bytes) was 16 bytes ago --- */
struct he_stat * stat_acc; /* 144 8 */
struct map_symbol ms; /* 152 24 */
struct thread * thread; /* 176 8 */
struct comm * comm; /* 184 8 */
/* --- cacheline 3 boundary (192 bytes) --- */
struct namespace_id cgroup_id; /* 192 16 */
u64 cgroup; /* 208 8 */
u64 ip; /* 216 8 */
u64 transaction; /* 224 8 */
u64 code_page_size; /* 232 8 */
u64 weight; /* 240 8 */
u64 ins_lat; /* 248 8 */
/* --- cacheline 4 boundary (256 bytes) --- */
u64 p_stage_cyc; /* 256 8 */
s32 socket; /* 264 4 */
s32 cpu; /* 268 4 */
int parallelism; /* 272 4 */
int mem_type_off; /* 276 4 */
u8 cpumode; /* 280 1 */
u8 depth; /* 281 1 */
struct simd_flags simd_flags; /* 282 1 */
_Bool dummy; /* 283 1 */
_Bool leaf; /* 284 1 */
char level; /* 285 1 */
filter_mask_t filtered; /* 286 2 */
u16 callchain_size; /* 288 2 */
/* XXX 6 bytes hole, try to pack */
union {
struct hist_entry_diff diff; /* 296 120 */
struct {
u16 row_offset; /* 296 2 */
u16 nr_rows; /* 298 2 */
_Bool init_have_children; /* 300 1 */
_Bool unfolded; /* 301 1 */
_Bool has_children; /* 302 1 */
_Bool has_no_entry; /* 303 1 */
}; /* 296 8 */
}; /* 296 120 */
/* --- cacheline 6 boundary (384 bytes) was 32 bytes ago --- */
char * srcline; /* 416 8 */
char * srcfile; /* 424 8 */
struct symbol * parent; /* 432 8 */
struct branch_info * branch_info; /* 440 8 */
/* --- cacheline 7 boundary (448 bytes) --- */
long int time; /* 448 8 */
struct hists * hists; /* 456 8 */
struct mem_info * mem_info; /* 464 8 */
struct block_info * block_info; /* 472 8 */
struct kvm_info * kvm_info; /* 480 8 */
void * raw_data; /* 488 8 */
u32 raw_size; /* 496 4 */
int num_res; /* 500 4 */
struct res_sample * res_samples; /* 504 8 */
/* --- cacheline 8 boundary (512 bytes) --- */
void * trace_output; /* 512 8 */
struct perf_hpp_list * hpp_list; /* 520 8 */
struct hist_entry * parent_he; /* 528 8 */
struct hist_entry_ops * ops; /* 536 8 */
struct annotated_data_type * mem_type; /* 544 8 */
union {
struct {
struct rb_root_cached hroot_in; /* 552 16 */
struct rb_root_cached hroot_out; /* 568 16 */
}; /* 552 32 */
struct rb_root sorted_chain; /* 552 8 */
}; /* 552 32 */
/* --- cacheline 9 boundary (576 bytes) was 8 bytes ago --- */
struct callchain_root callchain[] __attribute__((__aligned__(8))); /* 584 0 */
/* size: 584, cachelines: 10, members: 49 */ /* sum members: 578, holes: 1, sum holes: 6 */ /* paddings: 1, sum paddings: 4 */
/* forced alignments: 3 */
/* last cacheline: 8 bytes */
} __attribute__((__aligned__(8)));
--- tools/perf/util/hist.h | 8 ++++---- tools/perf/util/sample.h | 2 +- 2 files changed, 5 insertions(+), 5 deletions(-)
tools/perf/util/hist.h +4 -4
@@ -239,16 +239,16 @@ struct hist_entry {u64 cgroup;u64 ip;u64 transaction; -s32 socket; -s32 cpu; -int parallelism;u64 code_page_size;u64 weight;u64 ins_lat;u64 p_stage_cyc; +s32 socket; +s32 cpu; +int parallelism; +int mem_type_off;u8 cpumode;u8 depth; -int mem_type_off;struct simd_flags simd_flags;/* We are added by hists__add_dummy_entry. */
tools/perf/util/sample.h +1 -1
@@ -67,7 +67,7 @@ struct aux_sample { };struct simd_flags { -u64 arch:1 ,/* architecture (isa) */ +u8 arch:1 ,/* architecture (isa) */ pred:2 ;/* predication */ };
Signature · 1 line
2.48.1.502.g6dc24dfdaf-goog
Quoted 6 lines · Dmitry Vyukov
There are two notions of time: wall-clock time and CPU time. For a single-threaded program, or a program running on a single-core machine, these notions are the same. However, for a multi-threaded/ multi-process program running on a multi-core machine, these notions are significantly different. Each second of wall-clock time we have number-of-cores seconds of CPU time.
I read through it and it looks good to me.
For the patchkit.
Quoted 9 lines · Andi Kleen
Quoted 6 lines · Dmitry Vyukov
There are two notions of time: wall-clock time and CPU time. For a single-threaded program, or a program running on a single-core machine, these notions are the same. However, for a multi-threaded/ multi-process program running on a multi-core machine, these notions are significantly different. Each second of wall-clock time we have number-of-cores seconds of CPU time.
Quoted 5 lines
I read through it and it looks good to me.
For the patchkit.
Reviewed-by Andi Kleen <[email protected]>
Thanks for the review!
Namhyung, Arnaldo, what's left for this series to be merged?
Quoted 8 lines · Dmitry Vyukov
There are two notions of time: wall-clock time and CPU time. For a single-threaded program, or a program running on a single-core machine, these notions are the same. However, for a multi-threaded/ multi-process program running on a multi-core machine, these notions are significantly different. Each second of wall-clock time we have number-of-cores seconds of CPU time.
[...]
Applied to perf-tools-next, thanks!
Best regards,
Namhyung