Back in February, I could gather performance data and get results like this
with perf stat -T:
Performance counter stats for './ants-ocm':
1653.846545 task-clock (msec) # 3.318 CPUs utilized
4,791,092,035 instructions # 0.76 insns per cycle
6,293,499,230 cycles # 3.805 GHz
532,780,020 cpu/cycles-t/ # 8.47% transactional cycles
0 cpu/tx-start/ # 0.000 K/sec
1,601,050 cpu/el-start/ # 333 cycles / elision
532,427,121 cpu/cycles-ct/ # 0.01% aborted cycles
0.498513995 seconds time elapsed
but today when I try the same thing on the same machine (modulo time, reboots,
updates), I get:
Performance counter stats for './ants-ocm':
1594.655064 task-clock (msec) # 3.325 CPUs utilized
<not counted> instructions
<not counted> cycles
<not counted> cpu/cycles-t/
<not counted> cpu/tx-start/
<not counted> cpu/el-start/
<not counted> cpu/cycles-ct/
0.479534926 seconds time elapsed
If I select events by hand, with perf stat -e
'{task-clock,instructions,cycles,cpu/cycles-t/,cpu/tx-start/,cpu/el-start/,cpu/cycles-ct/}',
I get similar results (except that even task-clock can't be counted), but if I
switch to one fewer statistics (e.g., omitting tx-start), I get
Performance counter stats for './ants-ocm':
1316.494067 task-clock (msec) # 3.969 CPUs utilized
4,486,382,012 instructions # 0.92 insns per cycle
4,856,514,753 cycles # 3.689 GHz
371,558,738 cpu/cycles-t/ # 282.234 M/sec
1,600,709 cpu/el-start/ # 1.216 M/sec
371,279,122 cpu/cycles-ct/ # 282.021 M/sec
0.331677347 seconds time elapsed
It seems like I'm restricted to a fixed number of events (despite what it says
in the docs about the kernel multiplexing between different kinds of events).
If we drop cycles and add cpu/el-abort/ and cpu/el-capacity/ it then claims
that cpu/cycles-ct/ isn't supported.
Performance counter stats for './ants-ocm':
1316.258972 task-clock (msec) # 3.972 CPUs utilized
4,485,569,487 instructions
371,931,292 cpu/cycles-t/ # 282.567 M/sec
1,600,656 cpu/el-start/ # 1.216 M/sec
2,068 cpu/el-abort/ # 0.002 M/sec
1 cpu/el-capacity/ # 0.001 K/sec
<not supported> cpu/cycles-ct/
0.331412647 seconds time elapsed
What's odd to me is that this used to work. The machine is currently running
Linux common 3.13.0-55-generic #94-Ubuntu SMP Thu Jun 18 00:27:10 UTC
2015 x86_64 x86_64 x86_64 GNU/Linux
Any thoughts would be appreciated!
M.E.O.
--
To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in
