[openib-general] Re: cycles_to_units is incorrect in rdma_lat, rdma_bw.

Grant Grundler iod00d at hp.com
Wed Jun 15 18:03:23 PDT 2005


On Wed, Jun 15, 2005 at 11:55:07AM -0700, Shirley Ma wrote:
> > in case someone's firmware is less accurate than HP's.
> > Shirley, did the patch to get_clock.c fix the problem?
> 
> Cycles is not equal to the clock rate. If you have a processor with 
> internal doubling or pipelining or whatever, the clock rate isn't the 
> same.

This test assumes the ITC is running at some fixed rate.
Linux expects that rate to be reported in /proc/cpuinfo with "cpu Mhz".
Does your machine's cycle counter run at 1600 Mhz like the /proc/cpuinfo
output indicated?

hyperthreading, pipelining, superscaler, etc are all orthogonal issues.
The ITC is what matters.

> > Anyway, if get_cpu_mhz isnt reliable people can just dump raw
> > cycles data and do the math outside the tool.
> 
> Agree if the rdma_lat, and rdma_bw are only used for debugging/regression 
> testing not for performance measurement.

It would be nice if the test would determine the report result was
more than X% (e.g. 2%) off from the what the system clock reported.
It would be appropriate to warn the user about bogus results.

But this is a benchmarking issue. If rdma_lat is primarily a
micro-benchmark and secondarily an RDMA code example, then I
think it would be good to address this.

grant



More information about the general mailing list