Ian: I am currently thinking of doing a Sparse-Dense matrix matrix multiplication, where the Sparse matrix is a larger symmetric matrix, and the dense matrix is a skinny matrix. The sparse matrix is used mainly because of the memory issue, and the sparsity is just around 10~20%. Thus I am thinking of using MPISBAIJ type to reduce the memory cost. Does MPISBAIJ support MatMatMult with MPIDense matrix? If so, how would the data be distributed over processors? Would it be the same as MPIAIJ (row are by default evenly distributed over processors, but that does not make sense to MPISBAIJ as I suppose the lower part of the matrix is not stored)? We do not support MatMatMult with MPISBAIJ and MPIDense matrix, only MPIAIJ and MPIDense matrix.
Also, just another quick question, in your experience, considering the overhead cost for sparse algorithm, how sparse should a matrix be for the sparse algorithm to outperform dense algorithm in terms of computation time? Sparse matrix normally is defined as the sparsity less than 10%, or less than 1%. I would consider 20% as a dense matrix. The complexity of C=A*B is O(nnzA*nnzB), nnzA = num of nonzeros of A. Hong
