Patch to enable build on s390x
krebbel at linux.vnet.ibm.com
Thu Sep 22 10:15:37 CEST 2011
On 09/20/2011 03:07 PM, Torbjorn Granlund wrote:
> Andreas Krebbel <krebbel at linux.vnet.ibm.com> writes:
> Yes that's intentional. clock_gettime should be the better choice for both platforms.
> Using STCK directly has a higher resolution but is not stopped when the CPU is taken away
> from your system so in general the results will not be reliable when trying to measure the
> CPU cycles spent for a certain workload.
> This is the case for many accurate measurement mechanisms under the
> kernel Linux. E.g., the x86 rdtsc is not managed properly by that
> Our experience is that it is nevertheless better to use rdtsc than
> low-accuracy measurement mechanisms. Most GMP threshold parameters are
> for computations that take just a few nanoseconds.
>From the comments in time.c I expected that this is compensated by running the tests more
often and divide the time by the number of runs.
>From a first glance using stck seems to require the CPU frequency and this is not that
easy to determine. The only way I can think of is putting a cpu table into the code which
would need updating whenever a new CPU appears. So for now clock_gettime appears to be a
bit easier to me. Btw. libatlas also seems to rely on clock_gettime - however their tests
seem to run much longer.
More information about the gmp-bugs