Threaded FFTW3 on Gentoo vs. <a binary distro> [solved]

[diddly]

So I'm benchmarking an algorithm, the core of which is an FFT operation implemented by the FFTW library.

I am comparing both the single threaded and multithreaded APIs of FFTW version 3.2.2, running these 2 tests on both Gentoo and another distro. Both use glibc 2.9, both use gcc 4.3 (Gentoo is 4.3.2, other is 4.3.3). FFTW 3 is compiled with the same options/optimizations. I am running this all on the exact same hardware.

The tests just pound away running the FFT operation, and timing the result. I don't care about responsiveness, latency, anything desktop, just data throughput.

I am finding that for all intents and purposes, the single threaded benchmark is very close, with Gentoo slightly ahead (by maybe 5% or so)

However, in the threaded tests the binary distro tops Gentoo, in some cases by as much as 50-100%.

What could cause this to happen? Perhaps some kernel parameter that I am not aware of (I configured the Gentoo one myself, but might have missed something silly?)

Anything else?

Thanks in advance
_________________
Dave Flogeras

[diddly]

Well I think I found the cause.

In fact, FFTW was _not_ compiled with all the same options. On Gentoo it was using openmp for threads, while on the binary distro, it was not (which means pthreads). Apparently their pthreads implementation is quicker. When I recompiled FFTW with openmp on the binary distro, the performance was much closer to that of Gentoo.
_________________
Dave Flogeras