SpMV performance for sparse graphs

Issue #6 new
Aydin Buluc created an issue

Should we put a warning somehow? Maybe if nnz < tn and splits == 0, print a warning message within CombBLAS/SpMV that says "Given how sparse your graph is, OpenMP parallelization without splitting is likely to significantly degrade the performance of your SpMV. Please consider running your SpMV operations after calling the ActivateThreading function on the matrix. For example: "A.ActivateThreading(nthreadsslack)" where slack is a small constant such as 4"

Better solution (proposed by Ariful): we can implement the propagation blocking algorithm that will split the matrix on the fly.

Comments (0)

  1. Log in to comment