> > But speaking of writing parallel matrix vector products in native > Julia, this might be a great use case for shared arrays (although > right now I think only dense shared arrays exist). Amit, can you > comment on this? >
Actually you could use a shared array for both the sparse matrix and output vector. A CSC/CSR sparse matrix is just 3 dense arrays. If you load up the matrix row-wise into a shared array, you could assign each worker to do a subset of row-vector dot products, outputting into a shared vector. This could potentially work reasonably well as long as the memory bandwidth scales. You'll have to drop down to a low-level coding style (no more A*b), but this can probably be implemented in < 15 lines. > > I see it's possible to link Julia with MKL. I haven't tried this yet, > but if > > I do, will A*b (where A is sparse) call MKL to perform the matrix vector > > product? > No, it will not. There was a PR for this ( https://github.com/JuliaLang/julia/pull/4525) so you could probably use that code. I have can't comment on how efficient MKL is for multicore sparse mat-vec.
