PAPI overhead on various kernel interfaces
Note these results are *extremely* old. For more recent results please
see here.
Older results
These are the results of running the "papi_cost" test program
with PAPI 4.1.2.1 on a Core2 machine running four different kernels.
The reported value is "average cycles", a smaller value is better.
Kernel | Start/Stop | Read | Read TS | Accum | Reset |
2.6.32-pe | 17749 | 8982 | 9019 | 11519 | 2550 |
2.6.35-pe | 8683 | 2089 | 2109 | 4196 | 2084 |
2.6.30-perfmon | 9644 | 1200 | 1202 | 3623 | 2406 |
2.6.32-perfctr | 5702 | 195 | 203 | 3772 | 3556 |
The poor behaviour on 2.6.32/perf_events is due to various workarounds
that need to be done to properly set up the counters and read out the
data. These were fixed by 2.6.35.
Back to my projects page