Matlab on Microsoft 2008 compute server

I'm trying to take advantage of a compute server with 4 quad-core processors and 128 GB of RAM running Microsoft Server 2008 (Matlab R2010b). When I try the benchmark routine at the command line, it's very slow:

>>bench
ans =
  5.4150   0.0812   0.1598    34.6092   0.4319   0.8718

We have the parallel computing toolbox, and when I try the benchmark routine within a parallel for loop, each individual instance is much faster.

>>matlabpool; 
parfor i=1:8; 
  foo(i,:) = bench; 
end; 
matlabpool close; 
mean(foo)
ans =
  0.2253   0.2894   0.2894   0.343   0.4630   0.0292

The code that I'm running normally takes advantage of the two quad-cores that I have in my macPro and is not written using the parallel computing toolbox. Is there any way to get matlab to take advantage of the 4 processors (i.e., 16 cores) by default?

Answer 1

[Daniel Shub]

I am not sure what bench does with a compute server. In particular I do not know if the LU, ODE, and Sparse tests can use multiple cores. I know the FFT implementation uses multiple cores, so it is not surprising to me that it is faster. I am not sure what happens with a headless machine (not that your necessarily is) and a parfor loop with the 2-D and 3-D tests.

I wonder if the parfor loop divides the memory in such a way that it results in these tests becoming faster.

Instead of looking at the speeds for general tests, I would suggest testing with a procedure that is specific to what you want to compute.

Answer 2

[Fangjun Jiang]

First of all, run bench a few times in a row till you get stable results. My result of first run of bench is much worse than the second run.

Second, check your settings: menu File->Preferences ... ->General->Multithreading

[Brock Kirwan]

I've tried that. The first call to bench is pretty representative of multiple calls, such as: bench(10)