Newsgroups: comp.parallel.mpi From: James Cownie Subject: Re: MPI and Financial Modelling (Debugging, Deadlock) Organization: Dolphin Interconnect Solutions Date: Mon, 29 Jun 1998 09:12:01 +0100 Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Message-ID: <35974C51.B786D5C4@dolphinics.com> Greg Lindahl wrote: > > James Cownie writes: > > > As well as all the obvious things for parallel codes (acquire them all, > > run them all together, stop them all when bpts are hit, etc), we also > > have explicit support for > > MPI to allow you to see the pending messages and send/receive operations > > in each communicator. Since you can attach to a deadlocked code this is > > great for finding what > > went wrong. > > Although the last bit really should be provided by your MPI vendor. > mpich, lam, and Legion's mpi layer all do that -- it's the moral > equivalent to "printf debugging" in mpi... > > -- g Perhaps you could expand a little, Greg, I'm unclear what "the last bit is". If you mean that these systems detect deadlock and give you a trace of the pending messages/operations, I don't think you're right. Speaking for MPICH which is the MPI I'm most familiar with, there is definitely no explicit deadlock detection in MPICH. What MPICH does do (and the others may also do) is dump pending operations and messages if you finish your MPI code with pending messages lying around. A very different thing... -- Jim James Cownie Dolphin Interconnect Solutions +44 117 9071438 http://www.dolphinics.com/toolworks