perf vendor events: Fix DRAM_BW_Use 0 issue for CLX/SKX
PATCH 1/1 perf vendor events: Fix DRAM_BW_Use 0 issue for CLX/SKX
Ian reports an issue that the metric DRAM_BW_Use often remains 0.
The metric expression for DRAM_BW_Use on CLX/SKX:
"( 64 * ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) / 1000000000 ) / duration_time"
The counts of uncore_imc/cas_count_read/ and uncore_imc/cas_count_write/ are scaled up by 64, that is to turn a count of cache lines into bytes, the count is then divided by 1000000000 to give GB.
However, the counts of uncore_imc/cas_count_read/ and uncore_imc/cas_count_write/ have been scaled yet.
The scale values are from sysfs, such as /sys/devices/uncore_imc_0/events/cas_count_read.scale. It's 6.103515625e-5 (64 / 1024.0 / 1024.0).
So if we use original metric expression, the result is not correct.
But the difficulty is, for SKL client, the counts are not scaled.
The metric expression for DRAM_BW_Use on SKL:
"64 * ( arb@event\\=0x81\\,umask\\=0x1@ + arb@event\\=0x84\\,umask\\=0x1@ ) / 1000000 / duration_time / 1000"
root@kbl-ppc:~# perf stat -M DRAM_BW_Use -a -- sleep 1
Performance counter stats for 'system wide':
190 arb/event=0x84,umask=0x1/ # 1.86 DRAM_BW_Use
29,093,178 arb/event=0x81,umask=0x1/
1,000,703,287 ns duration_time1.000703287 seconds time elapsed
The result is expected.
So the easy way is just change the metric expression for CLX/SKX. This patch changes the metric expression to:
"( ( ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) * 1048576 ) / 1000000000 ) / duration_time"
1048576 = 1024 * 1024.
Before (tested on CLX):
root@lkp-csl-2sp5 ~# perf stat -M DRAM_BW_Use -a -- sleep 1
Performance counter stats for 'system wide':
765.35 MiB uncore_imc/cas_count_read/ # 0.00 DRAM_BW_Use
5.42 MiB uncore_imc/cas_count_write/
1001515088 ns duration_time1.001515088 seconds time elapsed
After:
root@lkp-csl-2sp5 ~# perf stat -M DRAM_BW_Use -a -- sleep 1
Performance counter stats for 'system wide':
767.95 MiB uncore_imc/cas_count_read/ # 0.80 DRAM_BW_Use
5.02 MiB uncore_imc/cas_count_write/
1001900010 ns duration_time1.001900010 seconds time elapsed
--- tools/perf/pmu-events/arch/x86/cascadelakex/clx-metrics.json | 2 +- tools/perf/pmu-events/arch/x86/skylakex/skx-metrics.json | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
tools/perf/pmu-events/arch/x86/cascadelakex/clx-metrics.json +1 -1
@@ -329,7 +329,7 @@ }, { "BriefDescription": "Average external Memory Bandwidth Use for reads and writes [GB / sec]", - "MetricExpr": "( 64 * ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) / 1000000000 ) / duration_time", + "MetricExpr": "( ( ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) * 1048576 ) / 1000000000 ) / duration_time", "MetricGroup": "Memory_BW;SoC", "MetricName": "DRAM_BW_Use" },
tools/perf/pmu-events/arch/x86/skylakex/skx-metrics.json +1 -1
@@ -323,7 +323,7 @@ }, { "BriefDescription": "Average external Memory Bandwidth Use for reads and writes [GB / sec]", - "MetricExpr": "( 64 * ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) / 1000000000 ) / duration_time", + "MetricExpr": "( ( ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) * 1048576 ) / 1000000000 ) / duration_time", "MetricGroup": "Memory_BW;SoC", "MetricName": "DRAM_BW_Use" },
Signature · 1 line
2.17.1
Quoted 57 lines · Jin Yao
Ian reports an issue that the metric
DRAM_BW_Useoften remains 0.The metric expression for
DRAM_BW_Useon CLX/SKX:"( 64 * ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) / 1000000000 ) /
duration_time"The counts of uncore_imc/cas_count_read/ and uncore_imc/cas_count_write/ are scaled up by 64, that is to turn a count of cache lines into bytes, the count is then divided by 1000000000 to give GB.
However, the counts of uncore_imc/cas_count_read/ and uncore_imc/cas_count_write/ have been scaled yet.
The scale values are from sysfs, such as
/sys/devices/uncore_imc_0/events/cas_count_read.scale. It's 6.103515625e-5 (64 / 1024.0 / 1024.0).So if we use original metric expression, the result is not correct.
But the difficulty is, for SKL client, the counts are not scaled.
The metric expression for
DRAM_BW_Useon SKL:"64 * ( arb@event\\=0x81\\,umask\\=0x1@ + arb@event\\=0x84\\,umask\\=0x1@ ) / 1000000 /
duration_time/ 1000"root@kbl-ppc:~# perf stat -M DRAM_BW_Use -a -- sleep 1Performance counter stats for 'system wide':
190 arb/event=0x84,umask=0x1/ # 1.86 DRAM_BW_Use 29,093,178 arb/event=0x81,umask=0x1/ 1,000,703,287 ns duration_time1.000703287 seconds time elapsedThe result is expected.
So the easy way is just change the metric expression for CLX/SKX. This patch changes the metric expression to:
"( ( ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) * 1048576 ) / 1000000000 ) /
duration_time"1048576 = 1024 * 1024.
Before (tested on CLX):
root@lkp-csl-2sp5 ~# perf stat -M
DRAM_BW_Use-a -- sleep 1Performance counter stats for 'system wide':
765.35 MiB uncore_imc/cas_count_read/ # 0.00 DRAM_BW_Use 5.42 MiB uncore_imc/cas_count_write/ 1001515088 ns duration_time1.001515088 seconds time elapsed
Quoted 7 lines
After:
root@lkp-csl-2sp5 ~# perf stat -M
DRAM_BW_Use-a -- sleep 1Performance counter stats for 'system wide':
767.95 MiB uncore_imc/cas_count_read/ # 0.80 DRAM_BW_Use
Nit, using ScaleUnit would allow this to be 0.80GB/s.
Quoted 8 lines
5.02 MiB uncore_imc/cas_count_write/ 1001900010 ns duration_time 1.001900010 seconds time elapsedFixes 038d3b53c284 ("perf vendor events intel: Update CascadelakeX events to v1.08")Fixes b5ff7f2799a4 ("perf vendor events: Update SkylakeX events to v1.21")Signed-off-by Jin Yao <[email protected]>
Thanks,
Ian
Quoted 34 lines
--- tools/perf/pmu-events/arch/x86/cascadelakex/clx-metrics.json | 2 +- tools/perf/pmu-events/arch/x86/skylakex/skx-metrics.json | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)tools/perf/pmu-events/arch/x86/cascadelakex/clx-metrics.json +1
-1@@ -329,7 +329,7 @@ }, { "BriefDescription": "Average external Memory Bandwidth Use for reads and writes [GB / sec]", - "MetricExpr": "( 64 * ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) / 1000000000 ) / duration_time", + "MetricExpr": "( ( ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) * 1048576 ) / 1000000000 ) / duration_time", "MetricGroup": "Memory_BW;SoC", "MetricName": "DRAM_BW_Use" },tools/perf/pmu-events/arch/x86/skylakex/skx-metrics.json +1
-2@@ -323,7 +323,7 @@ }, { "BriefDescription": "Average external Memory Bandwidth Use for reads and writes [GB / sec]", - "MetricExpr": "( 64 * ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) / 1000000000 ) / duration_time", + "MetricExpr": "( ( ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) * 1048576 ) / 1000000000 ) / duration_time", "MetricGroup": "Memory_BW;SoC", "MetricName": "DRAM_BW_Use" }, --2.17.1
Em Thu, Oct 22, 2020 at 06:02:31PM -0700, Ian Rogers escreveu:
Quoted 71 lines · Jin Yao
Quoted 57 lines
Ian reports an issue that the metric
DRAM_BW_Useoften remains 0.The metric expression for
DRAM_BW_Useon CLX/SKX:"( 64 * ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) / 1000000000 ) /
duration_time"The counts of uncore_imc/cas_count_read/ and uncore_imc/cas_count_write/ are scaled up by 64, that is to turn a count of cache lines into bytes, the count is then divided by 1000000000 to give GB.
However, the counts of uncore_imc/cas_count_read/ and uncore_imc/cas_count_write/ have been scaled yet.
The scale values are from sysfs, such as
/sys/devices/uncore_imc_0/events/cas_count_read.scale. It's 6.103515625e-5 (64 / 1024.0 / 1024.0).So if we use original metric expression, the result is not correct.
But the difficulty is, for SKL client, the counts are not scaled.
The metric expression for
DRAM_BW_Useon SKL:"64 * ( arb@event\\=0x81\\,umask\\=0x1@ + arb@event\\=0x84\\,umask\\=0x1@ ) / 1000000 /
duration_time/ 1000"root@kbl-ppc:~# perf stat -M DRAM_BW_Use -a -- sleep 1Performance counter stats for 'system wide':
190 arb/event=0x84,umask=0x1/ # 1.86 DRAM_BW_Use 29,093,178 arb/event=0x81,umask=0x1/ 1,000,703,287 ns duration_time1.000703287 seconds time elapsedThe result is expected.
So the easy way is just change the metric expression for CLX/SKX. This patch changes the metric expression to:
"( ( ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) * 1048576 ) / 1000000000 ) /
duration_time"1048576 = 1024 * 1024.
Before (tested on CLX):
root@lkp-csl-2sp5 ~# perf stat -M
DRAM_BW_Use-a -- sleep 1Performance counter stats for 'system wide':
765.35 MiB uncore_imc/cas_count_read/ # 0.00 DRAM_BW_Use 5.42 MiB uncore_imc/cas_count_write/ 1001515088 ns duration_time1.001515088 seconds time elapsedQuoted 7 lines
After:
root@lkp-csl-2sp5 ~# perf stat -M
DRAM_BW_Use-a -- sleep 1Performance counter stats for 'system wide':
767.95 MiB uncore_imc/cas_count_read/ # 0.80 DRAM_BW_UseNit, using ScaleUnit would allow this to be 0.80GB/s.
Quoted 3 lines
5.02 MiB uncore_imc/cas_count_write/
1001900010 nsduration_time
Quoted 7 lines
Quoted 5 lines
1.001900010 seconds time elapsedFixes 038d3b53c284 ("perf vendor events intel: Update CascadelakeX events to v1.08")Fixes b5ff7f2799a4 ("perf vendor events: Update SkylakeX events to v1.21")Signed-off-by Jin Yao <[email protected]>Acked-by Ian Rogers <[email protected]>
Thanks, applied.
- Arnaldo
Quoted 37 lines
Thanks,
IanQuoted 34 lines
--- tools/perf/pmu-events/arch/x86/cascadelakex/clx-metrics.json | 2 +- tools/perf/pmu-events/arch/x86/skylakex/skx-metrics.json | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)tools/perf/pmu-events/arch/x86/cascadelakex/clx-metrics.json +1
-1@@ -329,7 +329,7 @@ }, { "BriefDescription": "Average external Memory Bandwidth Use for reads and writes [GB / sec]", - "MetricExpr": "( 64 * ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) / 1000000000 ) / duration_time", + "MetricExpr": "( ( ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) * 1048576 ) / 1000000000 ) / duration_time", "MetricGroup": "Memory_BW;SoC", "MetricName": "DRAM_BW_Use" },tools/perf/pmu-events/arch/x86/skylakex/skx-metrics.json +1
-2@@ -323,7 +323,7 @@ }, { "BriefDescription": "Average external Memory Bandwidth Use for reads and writes [GB / sec]", - "MetricExpr": "( 64 * ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) / 1000000000 ) / duration_time", + "MetricExpr": "( ( ( uncore_imc@cas_count_read@ + uncore_imc@cas_count_write@ ) * 1048576 ) / 1000000000 ) / duration_time", "MetricGroup": "Memory_BW;SoC", "MetricName": "DRAM_BW_Use" }, --2.17.1
Signature · 1 line
- Arnaldo