Let us suppose I have to do the following computation $A_i\cdot B_i^{-1}\cdot C_i$, where $A_i,B_i,C_i$ are conformable matrices(dim=$50\times 50$).

For that, I have 3 lists, each of size 1 by 10,000 , where in list $A$, each component is a matrix $A_i$, and similarly for list $B$, we have $B_i^{-1}$, and for $C$.

  1. I thought of doing a ParallelTable. Or should I go for a loop? Is there a faster way to create a list where each component is $A_i\cdot B_i^{-1}\cdot C_i$? Is there a way to vectorize this ?
  2. What if I do a huge loop where I build the lists A,B and C, and as I build them, I also calculate $A_i\cdot B_i^{-1}\cdot C_i$?

If this question is too broad, or really needs a working example, tell me, don't just vote to close. Thanks ;)


A few possibilities:

n = 10000;
m = 50;
a = RandomReal[{-1, 1}, {n, m, m}];
b = RandomReal[{-1, 1}, {n, m, m}];
c = RandomReal[{-1, 1}, {n, m, m}];

result = MapThread[#1.LinearSolve[#2, #3] &, {a, b, c}]; // AbsoluteTiming // First
result2 = Table[a[[i]].LinearSolve[b[[i]], c[[i]]], {i, 1, Length[a]}]; // AbsoluteTiming // First

DistributeDefinitions[a, b, c]; // AbsoluteTiming // First
result3 = ParallelTable[a[[i]].LinearSolve[b[[i]], c[[i]]], {i, 1, Length[a]}]; // AbsoluteTiming // First

On my Haswell Quad Core CPU, this produces the following timings:





This highlights that moving all the data between several core is the major bottleneck here. So, in total, MapTread seems to be the better option. In fact, this makes me wonder whether MapTread is parallelized (the documentation does not state that).

I am currently not aware of a better way to parallelize this. Unfortunately, Compile with options RuntimeAttributes -> {Listable} and Parallelization -> True does not work because LinearSolve cannot be compiled and thus needs calls to MainEvaluate which makes parallelization inpossible.

  • $\begingroup$ Henrik, thank you for your answer. I still have a few questions. Why is there a need for DistributeDefinitions? Couldn't you just use ParallelTable? $\endgroup$ Feb 13 '19 at 8:53
  • 1
    $\begingroup$ Oh, you can use ParallelTable directly, at least in newer version of Mathematica. When you run the code result3 = ParallelTable[...] for the first time, it will take around 7 seconds; when you do it for the second time (with the same definitions of a, b, and c), it will take only 0.45 seonds.That puzzled me quite a lot (this is why the first version of my post states that ParallelTable were faster). So I decided to make DistributeDefinitions explicit in order to show that it is the bottleneck here. $\endgroup$ Feb 13 '19 at 9:02
  • $\begingroup$ Ah... I get it now! Thanks for the help ;) $\endgroup$ Feb 13 '19 at 9:54
  • $\begingroup$ You're welcome. $\endgroup$ Feb 13 '19 at 10:00
  • 1
    $\begingroup$ That's due to rounding errors. When B is ill-conditioned, this may be come quite severe. Actually, I expected that using Method -> "Cholesky" in LinearSolve would improve upon that but to no avail... =/ $\endgroup$ Feb 13 '19 at 16:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.