Need help on an elegant and fast way using MPI for a Parallel Matrix Problem

hamsha_ir — Thu, 10 Aug 2006 16:34:55 GMT

Hi, every body

I should be very obliged if some one can help me on this: Suppose for simplicity that there is a parallel loop, and two matrices A & B, both having dimension m*n, and also we have P processors. Matrix A is distributed row-wise and B is distributed column-wise on P processors. Can you possibly give an elegant and fast way for doing the following pesudo-code using MPI commands (like MPI_Gather & MPI_Type_struct &...) in C language?

for k=1,...,KMAX // parallel loop on all P processors

// some parallel calculation on B to produce new results

// dim A= dim B= m*n

//A is distributed row-wise on P processors for example in local_A

//B is distributed colum-wise on P processorsfor example in local_B

A=B // ???? how to do thisfast in the parallel, using MPI

end for

For an example suppose that P=3 (p0,p1,p2) , m=n=4 and note that first row of A and first column of B is stored on p0,..... (see the following diagram). Please note that each process has stored diiferent amount of A & B and we want direct setting i.e. a(i,j)=b(i,j) (no transpose). The bottleneck is that we want to set A=B in a loop for many times and each process has stored diiferent amount of A & B.

Thanks very much in advance

Best regards,

Ham. Sha.

		Mat. A			set A=B in MPI		Mat.B
						p0	p1	p2	p2
p0	a11	a12	a13	a14		b11	b12	b13	b14
p1	a21	a22	a23	a24	=	b21	b22	b23	b24
p2	a3 1	a32	a33	a34		b31	b32	b33	b34
p2	a41	a42	a43	a44		b41	b42	b43	b44

topic Need help on an elegant and fast way using MPI for a Parallel Matrix Problem in Intel® oneAPI Math Kernel Library

Need help on an elegant and fast way using MPI for a Parallel Matrix Problem