Re: [eigen] Updating the Benchmarks

[ Thread Index | Date Index | More Archives ]

I'm also running the bench suite on a core i7 (Xeon E5540) and I'm
surprised by the behavior of Eigen for level 1 routines, e.g.:

We can clearly obverse 2 step for vector sizes of 256 and 4096. They
correspond to an amount of 2048 Bytes and 32kB respectively (size *
sizeof(float) * 2 vectors). I can reproduce this behavior on my core
i7 server too.

So the second step is easily explained by the L1 cache which is of
32kB - perfect match :)

However, the first one (at 2048kB) puzzle me. Let me also add that it
does not appear on my core2 machines... Does anybody have an idea to
explain this behavior?


On Sun, Jan 30, 2011 at 12:34 AM, Milian Wolff <mail@xxxxxxxxxx> wrote:
> Gael Guennebaud, 26.01.2011:
>> btw, if you plane to bench matrix size up to 9000x9000, make sure you
>> disable the non optimized libraries like GMM, ublas, blitz, etc.
>> because they will take forever. Also make sure you disable the result
>> verifications when running the tests (BTL_CONFIG="--nocheck).
>> BLAS and lapack libs are very very painful regarding
>> configuration/linking Depending on how the libs have been generated
>> you might need very different settings. That's why you have these
>> trouble. That's also a non negligible advantage of Eigen.
> I've finally done the benchmarks on my i7 machine. Should I add that to the
> wiki?
> Bye
> --
> Milian Wolff
> mail@xxxxxxxxxx

Mail converted by MHonArc 2.6.19+