lkml.org 
[lkml]   [2009]   [Jun]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patches in this message
/
SubjectRe: [PATCH -tip] perf_counter tools: add support to set of multiple events in one short
From
Date
On Fri, 2009-06-26 at 14:25 +0200, Ingo Molnar wrote:
> * Jaswinder Singh Rajput <jaswinder@kernel.org> wrote:
>
> > On Fri, 2009-06-26 at 03:58 +0530, Jaswinder Singh Rajput wrote:
> > > On Fri, 2009-06-26 at 02:32 +0530, Jaswinder Singh Rajput wrote:
> > > > Add support for HARDWARE and SOFTWARE events :
> > > > perf stat -e all-sw-events
> > > > perf stat -e sw-events
> > > > perf stat -e all-hw-events
> > > > perf stat -e hw-events
> > > >
> > > > On AMD box :
> > > >
> > > > ./perf stat -e hw-events -e all-sw-events -- ls -lR > /dev/null
> > > >
> > > > Performance counter stats for 'ls -lR':
> > > >
> > > > 9977353 cycles # 557.193 M/sec (scaled from 21.81%)
> > > > 4244800 instructions # 0.425 IPC (scaled from 27.51%)
> > > > 2953188 cache-references # 164.923 M/sec (scaled from 89.10%)
> > > > 72469 cache-misses # 4.047 M/sec (scaled from 89.13%)
> > > > 775760 branches # 43.323 M/sec (scaled from 89.10%)
> > > > 57814 branch-misses # 3.229 M/sec (scaled from 83.34%)
> > > > <not counted> bus-cycles
> > > > 17.970985 cpu-clock-msecs
> > > > 17.906460 task-clock-msecs # 0.955 CPUs
> > > > 386 page-faults # 0.022 M/sec
> > > > 386 minor-faults # 0.022 M/sec
> > > > 0 major-faults # 0.000 M/sec
> > > > 4 context-switches # 0.000 M/sec
> > > > 1 CPU-migrations # 0.000 M/sec
> > > >
> > > > 0.018750671 seconds time elapsed.
> > > >
> > > > Reported-by : Ingo Molnar <mingo@elte.hu>
> > > > Signed-off-by: Jaswinder Singh Rajput <jaswinderrajput@gmail.com>
> > > > ---
> > > > tools/perf/util/parse-events.c | 66 ++++++++++++++++++++++++++++++++++++++-
> > > > 1 files changed, 64 insertions(+), 2 deletions(-)
> > >
> > > Please treat :
> > > [PATCH -tip] perf_counter tools: add support to set of multiple events in one short
> > > as
> > > [PATCH 1/2-tip] perf_counter tools: add support to set of multiple events in one short
> > >
> > > And here is 2/2 :
> > >
> > > [PATCH 2/2 -tip] perf_counter tools: Add support for all CACHE events
> > >
> > > Add support for all CACHE events :
> > > perf stat -e all-cache-events
> > > perf stat -e cache-events
> > >
> > > On AMD box (<not-counted> events are not available for AMD):
> > >
> > > ./perf stat -e all-cache-events -- ls -lR /usr/include/ > /dev/null
> > >
> > > Performance counter stats for 'ls -lR /usr/include/':
> > >
> > > 246370884 L1-d$-loads (scaled from 23.55%)
> > > 1074018 L1-d$-load-misses (scaled from 23.38%)
> > > 150708 L1-d$-stores (scaled from 23.57%)
> > > <not counted> L1-d$-store-misses
> > > 428804 L1-d$-prefetches (scaled from 23.47%)
> > > 314446 L1-d$-prefetch-misses (scaled from 23.42%)
> > > 252626137 L1-i$-loads (scaled from 23.24%)
> > > 3985110 L1-i$-load-misses (scaled from 23.24%)
> > > 93754 L1-i$-prefetches (scaled from 23.34%)
> > > <not counted> L1-i$-prefetch-misses
> > > 5202314 LLC-loads (scaled from 23.34%)
> > > 525467 LLC-load-misses (scaled from 23.25%)
> > > 5220558 LLC-stores (scaled from 23.21%)
> > > <not counted> LLC-store-misses
> > > <not counted> LLC-prefetches
> > > <not counted> LLC-prefetch-misses
> > > 251954203 dTLB-loads (scaled from 23.70%)
> > > 5297550 dTLB-load-misses (scaled from 23.96%)
> > > <not counted> dTLB-stores
> > > <not counted> dTLB-store-misses
> > > <not counted> dTLB-prefetches
> > > <not counted> dTLB-prefetch-misses
> > > 248561524 iTLB-loads (scaled from 24.15%)
> > > 4693 iTLB-load-misses (scaled from 24.18%)
> > > 106992392 branch-loads (scaled from 23.67%)
> > > 5239561 branch-load-misses (scaled from 23.43%)
> > >
> > > 0.395946903 seconds time elapsed.
> > >
> > > Reported-by: Ingo Molnar <mingo@elte.hu>
> > > Signed-off-by: Jaswinder Singh Rajput <jaswinderrajput@gmail.com>
> > > ---
> > > tools/perf/util/parse-events.c | 70 +++++++++++++++++++++++++++++++++++++---
> > > 1 files changed, 65 insertions(+), 5 deletions(-)
> > >
> >
> >
> > If this looks OK then can I send following patches.
>
> Would be nice to do the 'scaled' cleanup too that i suggested in the
> other thread, plus size things so that there's no such lines:
>
> 428804 L1-d$-prefetches (scaled from 23.47%)
> 314446 L1-d$-prefetch-misses (scaled from 23.42%)
>
> if that's done then it would be nice to have a series submitted to
> lkml with numbered patches and a 0/3 (or so) mail summarizing the
> changes, and with each patch having code and commit log quality that
> you can stand behind and which needs no modification from the
> maintainers.
>

In the mean time I also wrote another patch.

Please let me know which option is better then I will make it 4/4 :

Subject: [PATCH] perf stat: use set_multiple_events() to select default
events

Select SOFTWARE and HARDWARE events, if no event is selected.
this avoids replicating same arrays and reduce book-keeping

OR

[PATCH] perf stat: fix default attrs and nr_counters

memcpy(attrs, default_attrs, sizeof(attrs)) is only required
if no event is selected and only need to copy sizeof(default_attrs)

and set nr_counters as ARRAY_SIZE(default_attrs) in place of hardcoded value

Also make default_attrs table small and simple

Complete patches :

Subject: [PATCH] perf stat: use set_multiple_events() to select default events

Select SOFTWARE and HARDWARE events, if no event is selected.
this avoids replicating same arrays and reduce book-keeping

Signed-off-by: Jaswinder Singh Rajput <jaswinderrajput@gmail.com>
---
tools/perf/builtin-stat.c | 58 ++++++++++++++++++---------------------
tools/perf/util/parse-events.c | 2 +-
tools/perf/util/parse-events.h | 2 +
3 files changed, 30 insertions(+), 32 deletions(-)

diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index 8420ec5..ca68bb5 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -4,23 +4,28 @@
* Builtin stat command: Give a precise performance counters summary
* overview about any workload, CPU or specific PID.
*
- * Sample output:
+ * Sample output on AMD box (bus-cycles event is not available for AMD)

- $ perf stat ~/hackbench 10
- Time: 0.104
+ #./perf stat -- ls -lR /usr/include/ > /dev/null

- Performance counter stats for '/home/mingo/hackbench':
+ Performance counter stats for 'ls -lR /usr/include/':

- 1255.538611 task clock ticks # 10.143 CPU utilization factor
- 54011 context switches # 0.043 M/sec
- 385 CPU migrations # 0.000 M/sec
- 17755 pagefaults # 0.014 M/sec
- 3808323185 CPU cycles # 3033.219 M/sec
- 1575111190 instructions # 1254.530 M/sec
- 17367895 cache references # 13.833 M/sec
- 7674421 cache misses # 6.112 M/sec
+ 1912.810168 cpu-clock-msecs
+ 1903.386989 task-clock-msecs # 0.362 CPUs
+ 440 page-faults # 0.000 M/sec
+ 440 minor-faults # 0.000 M/sec
+ 0 major-faults # 0.000 M/sec
+ 1876 context-switches # 0.001 M/sec
+ 1 CPU-migrations # 0.000 M/sec
+ 972932473 cycles # 511.159 M/sec (scaled from 31.42%)
+ 588142134 instructions # 0.605 IPC (scaled from 30.98%)
+ 287837533 cache-references # 151.224 M/sec (scaled from 83.54%)
+ 7667661 cache-misses # 4.028 M/sec (scaled from 84.13%)
+ 75792456 branches # 39.820 M/sec (scaled from 85.04%)
+ 4457813 branch-misses # 2.342 M/sec (scaled from 84.89%)
+ <not counted> bus-cycles

- Wall-clock time elapsed: 123.786620 msecs
+ 5.257401849 seconds time elapsed.

*
* Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
@@ -32,6 +37,7 @@
* Wu Fengguang <fengguang.wu@intel.com>
* Mike Galbraith <efault@gmx.de>
* Paul Mackerras <paulus@samba.org>
+ * Jaswinder Singh Rajput <jaswinder@kernel.org>
*
* Released under the GPL v2. (and only v2, not any later version)
*/
@@ -45,20 +51,6 @@
#include <sys/prctl.h>
#include <math.h>

-static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {
-
- { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
- { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
- { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
- { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
-
- { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
- { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
- { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
- { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
-
-};
-
#define MAX_RUN 100

static int system_wide = 0;
@@ -468,16 +460,20 @@ int cmd_stat(int argc, const char **argv, const char *prefix)
{
int status;

- memcpy(attrs, default_attrs, sizeof(attrs));
-
argc = parse_options(argc, argv, options, stat_usage, 0);
if (!argc)
usage_with_options(stat_usage, options);
if (run_count <= 0 || run_count > MAX_RUN)
usage_with_options(stat_usage, options);

- if (!nr_counters)
- nr_counters = 8;
+ /*
+ * By default select SOFTWARE and HARDWARE events,
+ * if no event is selected
+ */
+ if (!nr_counters) {
+ set_multiple_events(PERF_TYPE_SOFTWARE);
+ set_multiple_events(PERF_TYPE_HARDWARE);
+ }

nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
assert(nr_cpus <= MAX_NR_CPUS);
diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
index c1cd93e..eea71c5 100644
--- a/tools/perf/util/parse-events.c
+++ b/tools/perf/util/parse-events.c
@@ -296,7 +296,7 @@ parse_generic_hw_symbols(const char *str, struct perf_counter_attr *attr)
return 0;
}

-static int set_multiple_events(unsigned int type)
+int set_multiple_events(unsigned int type)
{
struct perf_counter_attr attr;
int i;
diff --git a/tools/perf/util/parse-events.h b/tools/perf/util/parse-events.h
index e3d5529..ca44465 100644
--- a/tools/perf/util/parse-events.h
+++ b/tools/perf/util/parse-events.h
@@ -9,6 +9,8 @@ extern struct perf_counter_attr attrs[MAX_COUNTERS];

extern char *event_name(int ctr);

+extern int set_multiple_events(unsigned int type);
+
extern int parse_events(const struct option *opt, const char *str, int unset);

#define EVENTS_HELP_MAX (128*1024)
--
1.6.0.6
OR

Subject: [PATCH] perf stat: fix default attrs and nr_counters

memcpy(attrs, default_attrs, sizeof(attrs)) is only required
if no event is selected and only need to copy sizeof(default_attrs)

and set nr_counters as ARRAY_SIZE(default_attrs) in place of hardcoded value

Also make default_attrs table small and simple

Signed-off-by: Jaswinder Singh Rajput <jaswinderrajput@gmail.com>
---
tools/perf/builtin-stat.c | 31 ++++++++++++++++++-------------
1 files changed, 18 insertions(+), 13 deletions(-)
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index 8420ec5..e2b24f4 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -32,6 +32,7 @@
* Wu Fengguang <fengguang.wu@intel.com>
* Mike Galbraith <efault@gmx.de>
* Paul Mackerras <paulus@samba.org>
+ * Jaswinder Singh Rajput <jaswinder@kernel.org>
*
* Released under the GPL v2. (and only v2, not any later version)
*/
@@ -45,17 +46,20 @@
#include <sys/prctl.h>
#include <math.h>

-static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {
+#define CHW(x) .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_##x
+#define CSW(x) .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_##x

- { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
- { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
- { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
- { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
+static struct perf_counter_attr default_attrs[] = {

- { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
- { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
- { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
- { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
+ { CSW(TASK_CLOCK), },
+ { CSW(CONTEXT_SWITCHES), },
+ { CSW(CPU_MIGRATIONS), },
+ { CSW(PAGE_FAULTS), },
+
+ { CHW(CPU_CYCLES), },
+ { CHW(INSTRUCTIONS), },
+ { CHW(CACHE_REFERENCES), },
+ { CHW(CACHE_MISSES), },

};

@@ -468,16 +472,17 @@ int cmd_stat(int argc, const char **argv, const char *prefix)
{
int status;

- memcpy(attrs, default_attrs, sizeof(attrs));
-
argc = parse_options(argc, argv, options, stat_usage, 0);
if (!argc)
usage_with_options(stat_usage, options);
if (run_count <= 0 || run_count > MAX_RUN)
usage_with_options(stat_usage, options);

- if (!nr_counters)
- nr_counters = 8;
+ /* Set default attrs if no event is selected */
+ if (!nr_counters) {
+ memcpy(attrs, default_attrs, sizeof(default_attrs));
+ nr_counters = ARRAY_SIZE(default_attrs);
+ }

nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
assert(nr_cpus <= MAX_NR_CPUS);
--
1.6.0.6




\
 
 \ /
  Last update: 2009-06-26 14:43    [W:0.064 / U:1.956 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site