Setting the number of ATLAS threads at runtime
Brought to you by:
rwhaley,
tonyc040457
Menu
▾
▴
#997 Setting the number of ATLAS threads at runtime
open-later
1
2015-07-30
2015-06-25
No
Hello,
I am trying to install ATLAS for an Intel-Haswell cluster with 20 cores per node. I am getting close to MKL performance, which is quite impressive.
Now, I'd like to limit the number of threads at runtime to some nuber N <= 20 to enable concurrent execution of multiple multithreaded programs on the same node (SGE supports this nicely with its Parallel Environments).
The FAQ seems to imply that ATLAS uses less than ATL_NTHREADS threads for small problem sizes, so I suppose the number is not fixed. The FAQ also states that you cannot change the number of threads dynamically, which is OK; I only need to set it statically at the beginning of a program run.
Is there (short of installing multiple versions of ATLAS with different values of ATL_NTHREADS) any way to achieve this?
Kind regards and many thanks, Michael Fink
Unfortunately, to make ATLAS use less than the full machine as max# cores does indeed require an install for each max number you want to support, using the -t flag. If you have to use only particular coreIDs, see:
http://math-atlas.sourceforge.net/atlas_install/node21.html
Unfortunately, I can't warrant anything is fully working for threads: I'm still working on rewriting the threaded framework to handle massive parallelism, and I know the tid list thing does not work presently during the install.
I also know MKL currently is much faster in parallel than ATLAS on a 24-core haswell-ep (12-core/chip) for LU and square GEMM (not sure of other things).
I was working on a better algorithm for GEMM when I had to switch to other projects due to student scheduling problems. Several things have been added to the queue since then. So, feel free to try ATLAS, but unless you've got a case MKL isn't handling well, I'm guessing ATLAS will currently disappoint. I'll probably be posting 24-core Corei3 numbers to the developer list once I can work on the problem again, but for now, our threading support is in a state flux . . .
Sorry,
Clint
The GEMM tuning should now respect the -tl tid list in 3.11.35.