[openib-general] [PATCH] udapl: PPC64 cpuinfo change

Grant Grundler iod00d at hp.com
Fri Oct 7 10:05:50 PDT 2005


On Thu, Oct 06, 2005 at 11:01:21PM -0500, Troy Benjegerdes wrote:
> Oh boy.... is there some reason 'gettimeofday' does not work?

In general, it doesn't work as well.

> Trying to infer timebase/clock/rtsc frequency is going to be a mess.

Using cycle counters is quite portable today and provides
accurate results (with caveats on it's use). I'm open to
using the next best thing once it's clear the cycle counters
do NOT work.

> Think cpus that dynamically change frequency.. Laptops do now..
> how long before something with infiniband does and breaks this
> code horribly? (think embedded systems)

I don't buy this argument. Most of the tests load the CPU and it
essentially runs at a fixed frequency.

A better argument is how to benchmark under virtualized environment.
I think that is totally broken today regardless of what method
one uses to measure time.

> There are a couple of implementations of gettimeofday fully in userspace
> that hide the details and still read the high-res hardware counters. Google
> for 'vDSO gettimeofday'.

Well, I'm sure Michael is open to patches on this for userspace/perftest
stuff and like wise for James Lentini for uDAPL.

grant



More information about the general mailing list