Am Donnerstag, den 16.04.2020, 07:16 +0200 schrieb Werner LEMBERG: > > I did some better structured measurements, with interleaved runs on > > MSDM: [...] > > Have you ever tried valgrind's `callgrind` tool for profiling (and > using `kcachegrind` for displaying the results)? While very slow it > would avoid temperature issues and the like – no need to call it > multiple times to get reliable values.
What magic does callgrind do to prevent influence of frequency changes from the processor? Or cache misses due to the OS migrating the process from one core to another? From my experience working in HPC: Repeating measurements is always a must, right after making sure there's as little as possible running on the system while benchmarking and taskset'ting the process to one particular core. Jonas
signature.asc
Description: This is a digitally signed message part
