I was trying to benchmark a program I have writing against Mathematica (or better said the internal Kernels used, like MKL).
Since my program is not threaded I wanted to compare with Mathematica 10 running in a single core (I have a i7 processor). From this I discovered a certain option that seems to control the number of threads used by Mathematica and to my surprise, if I use it, I get better performance from Mathematica if I force it to use one threads.
In[1] := SetSystemOptions["ParallelOptions" -> "ParallelThreadNumber" -> 4]; SetSystemOptions["ParallelOptions" -> "MKLThreadNumber" -> 4]; n = 1024; t = Table[RandomReal[], {i, 0, n}, {j, 0, n}]; Timing[Inverse[t]] // First Out[1] := 0.626792 The above is the default and it takes 0.62 seconds to invert the matrix. If I change the option I get better performance:
In[2] := SetSystemOptions["ParallelOptions" -> "ParallelThreadNumber" -> 1]; SetSystemOptions["ParallelOptions" -> "MKLThreadNumber" -> 1]; n = 1024; t = Table[RandomReal[], {i, 0, n}, {j, 0, n}]; Timing[Inverse[t]] // First Out[2] := 0.203677 I find hard to believe I found a bug in Mathematica, most likely I don't understand how this options work. Can somebody tell me if I am using these options correctly? Does the second line contain the correct code to force Mathematica to use only one thread?
AbsoluteTimingis the right way to measure wall time for functions. Still, I would like to see if this is the correct way to control the number of threads. $\endgroup$Timingmeasures (time spent by core 1) + (time spent by core 2) + (time spent by core 3) + (time spent by core 4).AbsoluteTimingmeasures the physical time elapsed. $\endgroup$AbsoluteTimingthe results are 0.134 and 0.2160 respectively. $\endgroup$