How to parallelize integrating in Mathematica 8

问题

Somebody have idea how to use all cores for calculating integration? I need to use parallelize or parallel table but how?

 f[r_] := Sum[(((-1)^n*(2*r - 2*n - 7)!!)/(2^n*n!*(r - 2*n - 1)!))*
 x^(r - 2*n - 1), {n, 0, r/2}]; 


 Nw := Transpose[Table[f[j], {i, 1}, {j, 5, 200, 1}]]; 

 X1 = Integrate[Nw . Transpose[Nw], {x, -1, 1}]; 

 Y1 = Integrate[D[Nw, {x, 2}] . Transpose[D[Nw, {x, 2}]], {x, -1, 1}]; 

 X1//MatrixForm
 Y1//MatrixForm

回答1:

If one helps Integrate a bit by expanding the matrix elements first, things are doable with a little bit of effort.

On a quad-core laptop with Windows and Mathematica 8.0.4 the following code below runs for the asked DIM=200 in about 13 minutes, for DIM=50 the code runs in 6 second.

---

$starttime = AbsoluteTime[]; Quiet[LaunchKernels[]]; 
DIM = 200; 
Print["$Version = ", $Version, "  |||  ", "Number of Kernels : ", Length[Kernels[]]]; 
f[r_] := f[r] = Sum[(((-1)^n*(-(2*n) + 2*r - 7)!!)*x^(-(2*n) + r - 1))/(2^n*n!*(-(2*n) + r - 1)!), {n, 0, r/2}]; 
Nw = Transpose[Table[f[j], {i, 1}, {j, 5, DIM, 1}]]; 
nw2 = Nw . Transpose[Nw]; 
Print["Seconds for expanding Nw.Transpose[Nm] ", Round[First[AbsoluteTiming[nw3 = ParallelMap[Expand, nw2]; ]]]]; 
Print["do the integral once: ", Integrate[x^n, {x, -1, 1}, Assumptions -> n > -1]]; 
Print["the integration can be written as a simple rule: ", intrule = (pol_Plus)?(PolynomialQ[#1, x] & ) :> 
     (Select[pol,  !FreeQ[#1, x] & ] /. x^(n_.) /; n > -1 :> ((-1)^n + 1)/(n + 1)) + 2*(pol /. x -> 0)]; 
Print["Seconds for integrating Nw.Transpose[Nw] : ", Round[First[AbsoluteTiming[X1 = ParallelTable[row /. intrule, {row, nw3}]; ]]]]; 
Print["expanding: ", Round[First[AbsoluteTiming[preY1 = ParallelMap[Expand, D[Nw, {x, 2}] . Transpose[D[Nw, {x, 2}]]]; ]]]]; 
Print["Seconds for integrating : ", Round[First[AbsoluteTiming[Y1 = ParallelTable[py /. intrule, {py, preY1}]; ]]]]; 
Print["X1 = ", (Shallow[#1, {4, 4}] & )[X1]]; 
Print["Y1 = ", (Shallow[#1, {4, 4}] & )[Y1]]; 
Print["seq Y1 : ", Simplify[FindSequenceFunction[Diagonal[Y1], n]]]; 
Print["seq X1 0 : ",Simplify[FindSequenceFunction[Diagonal[X1, 0], n]]]; 
Print["seq X1 2: ",Simplify[FindSequenceFunction[Diagonal[X1, 2], n]]]; 
Print["seq X1 4: ",Simplify[FindSequenceFunction[Diagonal[X1, 4], n]]]; 
Print["overall time needed in seconds: ", Round[AbsoluteTime[] - $starttime]]; 

回答2:

I changed the integration of a list into a list of integrations so that I can use ParallelTable:

X1par=ParallelTable[Integrate[i, {x, -1, 1}], {i, Nw.Transpose[Nw]}];

X1par==X1

(* ===> True *)

Y1par = ParallelTable[Integrate[i,{x,-1,1}],{i,D[Nw,{x,2}].Transpose[D[Nw,{x,2}]]}]

Y1 == Y1par

(* ===> True *)

In my timings, with {j, 5, 30, 1} instead of {j, 5, 200, 1} to restrict the time used somewhat, this is about 3.4 times faster on my quod-core. But it can be done even faster with:

X2par = Parallelize[Integrate[#, {x, -1, 1}] & /@ (Nw.Transpose[Nw])]

X2par == X1par == X1

(* ===> True *)

This is about 6.8 times faster, a factor of 2.3 of which is due to Parallelize.

Timing and AbsoluteTiming are not very trustworthy when parallel execution is concerned. I used AbsoluteTime before and after each line and took the difference.

---

EDIT

We shouldn't forget ParallelMap:

At the coarsest list level (1):

ParallelMap[Integrate[#, {x, -1, 1}] &, Nw.Transpose[Nw], {1}]  

At the deepest list level (most fine-grained parallelization):

ParallelMap[Integrate[#, {x, -1, 1}] &, Nw.Transpose[Nw], {2}]

来源：https://stackoverflow.com/questions/8021501/how-to-parallelize-integrating-in-mathematica-8