Skip to contentSkip to conversation

perf, tools, stat: Initialize statistics correctly

PATCH In review Andi Kleen 4 participants 8 messages Mar 25, 2014 – Apr 14, 2014 perf events

lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

PATCH 1/2 perf, tools, stat: Initialize statistics correctly

From: Andi Kleen <[email protected]>

perf stat did initialize the stats structure used to compute stddev etc. incorrectly. It merely zeroes it. But one member (min) needs to be set to a non zero value. This causes min to be not computed at all. Call init_stats() correctly.

It doesn't matter for stat currently because it doesn't use min, but it's still better to do it correctly.

The other users of statistics are already correct.

---
 tools/perf/builtin-stat.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
tools/perf/builtin-stat.c +9 -2
index 8b0e1c9..65a151e 100644
@@ -174,13 +174,20 @@ static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
 
 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
 {
-	memset(evsel->priv, 0, sizeof(struct perf_stat));
+	int i;
+	struct perf_stat *ps = evsel->priv;
+
+	for (i = 0; i < 3; i++)
+		init_stats(&ps->res_stats[i]);
 }
 
 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 {
 	evsel->priv = zalloc(sizeof(struct perf_stat));
-	return evsel->priv == NULL ? -ENOMEM : 0;
+	if (evsel == NULL)
+		return -ENOMEM;
+	perf_evsel__reset_stat_priv(evsel);
+	return 0;
 }
 
 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
Signature · 1 line
1.8.5.3
author
#1 lore raw

PATCH 2/2 perf, tools: Support spark lines in perf stat

From: Andi Kleen <[email protected]>

perf stat -rX prints the stddev for multiple measurements. Just looking at the stddev for judging the quality of the data is a bit dangerous The simplest sanity check is to just look at a simple plot. This patchs add a sparkline to the end of the measurements to make it simple to judge the data.

The sparkline only uses UTF-8, so should be readable in all modern tools and terminals.

The sparkline is between the minimum and maximum of the data, so it's mainly a indicator of variance. To keep the code simple and make the output not too wide only the first 8 values are printed. If more values are there it adds '..'

The code is inspired by Zach Holman's spark shell script.

Example output (view in non-proportial font):

Performance counter stats for 'true' (10 runs):

          0.175672      task-clock (msec)         #    0.555 CPUs utilized            ( +-  1.77% ) █▄▁▁▁▁▁▁..
                 0      context-switches          #    0.000 K/sec
                 0      cpu-migrations            #    0.000 K/sec
               114      page-faults               #    0.647 M/sec                    ( +-  0.14% ) ▁█▁▁████..
           520,798      cycles                    #    2.965 GHz                      ( +-  1.75% ) █▄▁▁▁▁▁▁..
           433,525      instructions              #    0.83  insns per cycle          ( +-  0.28% ) ▅▇▅▄▇█▁▆..
            83,012      branches                  #  472.537 M/sec                    ( +-  0.31% ) ▅▇▆▄▇█▁▆..
             3,157      branch-misses             #    3.80% of all branches          ( +-  2.55% ) ▇█▃▅▁▃▁▂..

0.000316660 seconds time elapsed ( +- 1.78% ) █▅▁▁▁▁▁▁..

As you can see even in the most simple run there are quite interesting patterns. The time sparkline suggests it would be also useful to have an option to throw the first measurement away.

Known issues:
- Makes the perf stat output wider. Could be adjust by shrinking some white space. Not done so far.
- No output for -A/--per-socket/--per-core with -rX. This code is missing the basic noise detection code. Once it's added there sparklines could be shown too.

---
 tools/perf/Makefile.perf  |  1 +
 tools/perf/builtin-stat.c | 12 ++++++++++++
 tools/perf/util/spark.c   | 45 +++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/spark.h   |  3 +++
 tools/perf/util/stat.c    | 33 +++++++++++++++++++++++++++++++++
 tools/perf/util/stat.h    | 10 ++++++++++
 6 files changed, 104 insertions(+)
 create mode 100644 tools/perf/util/spark.c
 create mode 100644 tools/perf/util/spark.h
tools/perf/Makefile.perf +1 -0
index 7257e7e..432d099 100644
@@ -359,6 +359,7 @@ LIB_OBJS += $(OUTPUT)util/trace-event-scripting.o
 LIB_OBJS += $(OUTPUT)util/trace-event.o
 LIB_OBJS += $(OUTPUT)util/svghelper.o
 LIB_OBJS += $(OUTPUT)util/sort.o
+LIB_OBJS += $(OUTPUT)util/spark.o
 LIB_OBJS += $(OUTPUT)util/hist.o
 LIB_OBJS += $(OUTPUT)util/probe-event.o
 LIB_OBJS += $(OUTPUT)util/util.o
tools/perf/builtin-stat.c +12 -0
index 65a151e..6bf7b6c 100644
@@ -1176,6 +1176,9 @@ static void print_aggr(char *prefix)
 				if (run != ena)
 					fprintf(output, "  (%.2f%%)",
 						100.0 * run / ena);
+
+				fputc(' ', output);
+				print_stat_spark(output, counter->priv);
 			}
 			fputc('\n', output);
 		}
@@ -1224,6 +1227,9 @@ static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
 
 	print_noise(counter, avg);
 
+	fputc(' ', output);
+	print_stat_spark(output, counter->priv);
+
 	if (csv_output) {
 		fputc('\n', output);
 		return;
@@ -1295,6 +1301,9 @@ static void print_counter(struct perf_evsel *counter, char *prefix)
 			if (run != ena)
 				fprintf(output, "  (%.2f%%)",
 					100.0 * run / ena);
+
+			fputc(' ', output);
+			print_stat_spark(output, counter->priv);
 		}
 		fputc('\n', output);
 	}
@@ -1355,6 +1364,9 @@ static void print_stat(int argc, const char **argv)
 			fprintf(output, "                                        ");
 			print_noise_pct(stddev_stats(&walltime_nsecs_stats),
 					avg_stats(&walltime_nsecs_stats));
+
+			fputc(' ', output);
+			print_stat_spark(output, &walltime_nsecs_stats);
 		}
 		fprintf(output, "\n\n");
 	}
tools/perf/util/spark.c new file +45 -0

45 lines not on this page · raw message

tools/perf/util/spark.h new file +3 -0
new file mode 100644
index 0000000..f2d5ac5
@@ -0,0 +1,3 @@
+#pragma once
+void print_spark(FILE *outf, unsigned long *val, int numval);
+
tools/perf/util/stat.c +33 -0
index 6506b3d..2b26d74 100644
@@ -1,10 +1,16 @@
 #include <math.h>
+#include <stdio.h>
 
 #include "stat.h"
+#include "spark.h"
 
 void update_stats(struct stats *stats, u64 val)
 {
 	double delta;
+	int n = stats->n;
+
+	if (n < NUM_SPARK_VALS)
+		stats->svals[n] = val;
 
 	stats->n++;
 	delta = val - stats->mean;
@@ -61,3 +67,30 @@ double rel_stddev_stats(double stddev, double avg)
 
 	return pct;
 }
+
+static int all_the_same(unsigned long *vals, int len)
+{
+	int i;
+	unsigned long v0 = vals[0];
+
+	for (i = 1; i < len; i++)
+		if (vals[i] != v0)
+			return 0;
+	return 1;
+}
+
+void print_stat_spark(FILE *f, struct stats *stat)
+{
+	int n = stat->n, len;
+
+	if (n <= 1)
+		return;
+	len = n;
+	if (len > NUM_SPARK_VALS)
+		len = NUM_SPARK_VALS;
+	if (all_the_same(stat->svals, len))
+		return;
+	print_spark(f, stat->svals, len);
+	if (stat->n > NUM_SPARK_VALS)
+		fputs("..", f);
+}
tools/perf/util/stat.h +10 -0
index ae8ccd7..1b4dc71 100644
@@ -1,12 +1,16 @@
 #ifndef __PERF_STATS_H
 #define __PERF_STATS_H
 
+#include <stdio.h>
 #include "types.h"
 
+#define NUM_SPARK_VALS 8 /* support spark line on first N items */
+
 struct stats
 {
 	double n, mean, M2;
 	u64 max, min;
+	unsigned long svals[NUM_SPARK_VALS];
 };
 
 void update_stats(struct stats *stats, u64 val);
@@ -14,12 +18,18 @@ double avg_stats(struct stats *stats);
 double stddev_stats(struct stats *stats);
 double rel_stddev_stats(double stddev, double avg);
 
+void print_stat_spark(FILE *f, struct stats *stat);
+
 static inline void init_stats(struct stats *stats)
 {
+	int i;
+
 	stats->n    = 0.0;
 	stats->mean = 0.0;
 	stats->M2   = 0.0;
 	stats->min  = (u64) -1;
 	stats->max  = 0;
+	for (i = 0; i < NUM_SPARK_VALS; i++)
+		stats->svals[i] = 0;
 }
 #endif
Signature · 1 line
1.8.5.3
#2 lore raw

Re: [PATCH 2/2] perf, tools: Support spark lines in perf stat

Quoted 2 lines · Andi Kleen

From: Andi Kleen <[email protected]>

SNIP

tools/perf/builtin-stat.c static void print_aggr(char *prefix) · line 1181

Quoted 10 lines

index 65a151e..6bf7b6c 100644

tools/perf/builtin-stat.c +3 -0
@@ -1176,6 +1176,9 @@ static void print_aggr(char *prefix)
 				if (run != ena)
 					fprintf(output, "  (%.2f%%)",
 						100.0 * run / ena);
+
+				fputc(' ', output);
+				print_stat_spark(output, counter->priv);

I dont think we want it in csv output

[jolsa@krava perf]$ ./perf stat -r 20 -x , true
0.285896,,task-clock,2.39% █▃▃▂▂▁▁▁..
0,,context-switches,0.00%
0,,cpu-migrations,100.00%
91,,page-faults,0.08% █▁██████..
165335,,cycles,30.69% █▇▇▁▇▁▁▁..

jirka

#3 lore raw

Re: [PATCH 2/2] perf, tools: Support spark lines in perf stat

In reply to Andi Kleen
Quoted 2 lines · Andi Kleen

From: Andi Kleen <[email protected]>

SNIP

Quoted 31 lines
tools/perf/builtin-stat.c +9 -0
@@ -1176,6 +1176,9 @@ static void print_aggr(char *prefix)
 				if (run != ena)
 					fprintf(output, "  (%.2f%%)",
 						100.0 * run / ena);
+
+				fputc(' ', output);
+				print_stat_spark(output, counter->priv);
 			}
 			fputc('\n', output);
 		}
@@ -1224,6 +1227,9 @@ static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
 
 	print_noise(counter, avg);
 
+	fputc(' ', output);
+	print_stat_spark(output, counter->priv);
+
 	if (csv_output) {
 		fputc('\n', output);
 		return;
@@ -1295,6 +1301,9 @@ static void print_counter(struct perf_evsel *counter, char *prefix)
 			if (run != ena)
 				fprintf(output, "  (%.2f%%)",
 					100.0 * run / ena);
+
+			fputc(' ', output);
+			print_stat_spark(output, counter->priv);
 		}
 		fputc('\n', output);
 	}

tools/perf/builtin-stat.c static void print_stat(int argc, const char **argv) · line 1369

Quoted 6 lines
tools/perf/builtin-stat.c +3 -0
@@ -1355,6 +1364,9 @@ static void print_stat(int argc, const char **argv)
 			fprintf(output, "                                        ");
 			print_noise_pct(stddev_stats(&walltime_nsecs_stats),
 					avg_stats(&walltime_nsecs_stats));
+
+			fputc(' ', output);
+			print_stat_spark(output, &walltime_nsecs_stats);

Do we want to be that explicit about printing ' ' here? I dont see it elsewhere..

fput could go into print_stat_spark

SNIP

Quoted 5 lines
tools/perf/util/stat.c +5 -0
+void print_stat_spark(FILE *f, struct stats *stat)
+{
+	int n = stat->n, len;
+
+	if (n <= 1)

tools/perf/util/stat.c double rel_stddev_stats(double stddev, double avg) · line 96

Quoted 10 lines
tools/perf/util/stat.c +10 -0
+		return;
+	len = n;
+	if (len > NUM_SPARK_VALS)
+		len = NUM_SPARK_VALS;
+	if (all_the_same(stat->svals, len))
+		return;
+	print_spark(f, stat->svals, len);
+	if (stat->n > NUM_SPARK_VALS)
+		fputs("..", f);
+}

whats the reason for 'n' in here? looks like you could do only with 'len'

thanks,
jirka

author
#4 lore raw

Re: [PATCH 2/2] perf, tools: Support spark lines in perf stat

Quoted 1 line

fput could go into print_stat_spark

I wanted to keep the function generic, so that it can be easily used in other places. Adding the space is somewhat specific to stat.

Also the rest of the stat code has the white space printing centralized in the caller too. I suppose makes it easier to change.

Quoted 11 lines

tools/perf/util/stat.c double rel_stddev_stats(double stddev, double avg) · line 96

Quoted 9 lines
tools/perf/util/stat.c +9 -0
+	len = n;
+	if (len > NUM_SPARK_VALS)
+		len = NUM_SPARK_VALS;
+	if (all_the_same(stat->svals, len))
+		return;
+	print_spark(f, stat->svals, len);
+	if (stat->n > NUM_SPARK_VALS)
+		fputs("..", f);
+}

whats the reason for 'n' in here? looks like you could do only with 'len'

len is capped to NUM_SPARK_VALS, but the test needs to be on the uncapped value. Otherwise exactly NUM_SPARK_VALS entries would get a .. incorrectly.

-Andi

Acked-by on v1 1/2
#5 lore raw
In reply to Andi Kleen
Quoted 11 lines · Andi Kleen

From: Andi Kleen <[email protected]>

perf stat did initialize the stats structure used to compute stddev etc. incorrectly. It merely zeroes it. But one member (min) needs to be set to a non zero value. This causes min to be not computed at all. Call init_stats() correctly.

It doesn't matter for stat currently because it doesn't use min, but it's still better to do it correctly.

The other users of statistics are already correct.

Thanks,
Namhyung

Acked-by on v1 1/2
#6 lore raw
In reply to Andi Kleen
Quoted 7 lines · Andi Kleen

From: Andi Kleen <[email protected]>

perf stat did initialize the stats structure used to compute stddev etc. incorrectly. It merely zeroes it. But one member (min) needs to be set to a non zero value. This causes min to be not computed at all. Call init_stats() correctly.

Quoted 6 lines

It doesn't matter for stat currently because it doesn't use min, but it's still better to do it correctly.

The other users of statistics are already correct.

thanks,
jirka

#7 lore raw

[tip:perf/urgent] perf stat: Initialize statistics correctly

In reply to Andi Kleen
perf stat: Initialize statistics correctly
Commit-ID:  90f6bb6c98ffef42125d7be6d4612505f561fbce
Gitweb:     http://git.kernel.org/tip/90f6bb6c98ffef42125d7be6d4612505f561fbce
Author:     Andi Kleen <[email protected]>
AuthorDate: Tue, 25 Mar 2014 10:31:38 -0700
Committer:  Jiri Olsa <[email protected]>
CommitDate: Mon, 14 Apr 2014 12:56:06 +0200

perf stat: Initialize statistics correctly

perf stat did initialize the stats structure used to compute stddev etc. incorrectly. It merely zeroes it. But one member (min) needs to be set to a non zero value. This causes min to be not computed at all. Call init_stats() correctly.

It doesn't matter for stat currently because it doesn't use min, but it's still better to do it correctly.

The other users of statistics are already correct.

---
 tools/perf/builtin-stat.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
tools/perf/builtin-stat.c +9 -2
index 8b0e1c9..65a151e 100644
@@ -174,13 +174,20 @@ static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
 
 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
 {
-	memset(evsel->priv, 0, sizeof(struct perf_stat));
+	int i;
+	struct perf_stat *ps = evsel->priv;
+
+	for (i = 0; i < 3; i++)
+		init_stats(&ps->res_stats[i]);
 }
 
 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 {
 	evsel->priv = zalloc(sizeof(struct perf_stat));
-	return evsel->priv == NULL ? -ENOMEM : 0;
+	if (evsel == NULL)
+		return -ENOMEM;
+	perf_evsel__reset_stat_priv(evsel);
+	return 0;
 }
 
 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)