Karl Rupp <[email protected]> writes: > Hi, > > >> When using good preconditioners, spMV is essentially never the >>> bottleneck and hence I don't think a separate communication thread >>> should be implemented in PETSc. Instead, such a fallback should be part >>> of a good MPI implementation. >> >> SpMV is an important part of most of those scalable preconditioners. In >> multigrid, those are grid transfer operators, residuals, and Chebyshev >> or Krylov-accelerated smoothers. > > From the context I was referring to SpMV for a 'full system matrix' as > part of the outer Krylov solver as it was the topic in the paper (cf. > third paragraph). In preconditioners you may use different storage > formats, not communicate across nodes, etc.
It's exactly the same operation in the Chebyshev smoother, for example. There is no reason to use a different format (it doesn't) and "not communicating across nodes" in the preconditioner is an egregious crime that prevents a scalable method. So I would say that fine-grid SpMV performance is of similarly high performance with a "scalable preconditioner" as with Jacobi.
pgpgxTLZkb6bU.pgp
Description: PGP signature
