custom build of atlas only uses ~70% CPU

[zarquon42]

First of all, the option to easyily compile a multithreaded atlas is just great!
However, I compared atlas to the OpenBLAS library and recognized it to be approx. 30% slower. When examining the CPU-load during calculation, I saw atlas only using 70% of the CPU horsepower (while OpenBlas uses full 100%) and wondered, if there was an option during compiling that can trigger atlas to gain full speed.

I attached a screenshot of the CPU monitoring.

Thank you in advance
Stefan

[Sylvestre Ledru (sylvestre)]

Some codes could help here...

You should report a bug upstream if you want to see it fixed.

S

[zarquon42 (zarquon42)]

Sorry for the incomplete report. As my understanding on how atlas works is very limited, I didn't know if you (the packager) added some debian/ubuntu specific compiler options leading to this result. So I asked here first because I wondererd if there maybe was a trivial answer (like setting a specific flag during compilation).

So here is what I did:
disabled CPU throtteling in BIOS
apt-get source atlas
fakeroot debian/rules
installed the package

Open R:

a<-list();for(i in 1:10){a[[i]]<-matrix(rnorm(1e6),1000,1000)} ## create a list containing large matrices
system.time(lapply(a,solve)) ### measure time of inverting 10 matrices

I also monitored the CPU usage using the gnome system monitor and compared with the result using the OpenBLAS library.
The latter uses 100% on all cores.

My hardware specs:
CPU: Intel Core2 Quad CPU Q9400 @ 2.66GHz × 4
RAM: 8GB

Ubuntu 12.04 64bit