<div class="gmail_quote">On Mon, Aug 29, 2011 at 10:08, Matthew Knepley <span dir="ltr"><<a href="mailto:knepley@gmail.com">knepley@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div>As I said, no one knows</div><div>how to do this for Krylov methods (and everyone has tried).</div></blockquote></div><br><div>There are methods and even simply running multiple independent Krylov solves concurrently would be good for memory traffic (matrix entries get reused for multiple vectors). The problem with block Krylov methods (that try to share information between multiple simultaneous solves) is loss of orthogonality between the subspaces generated by each vector. And it's not so much that there are no ways to detect and account for this, but robustness is still a problem and making efficient software for it is more tricky and AFAIK, has not been done.</div>