From: "John B. Pormann" Newsgroups: comp.parallel.pvm Subject: Re: PVM Performance on SGI Power Challenge Date: Fri, 4 Dec 1998 13:03:45 -0500 Organization: Duke University, Durham, NC, USA Message-Id: References: <36675533.74EFF212@its-gipps1.cc.monash.edu.au> Mime-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII In-Reply-To: <36675533.74EFF212@its-gipps1.cc.monash.edu.au> We have an old Power Challenge in our lab and I've seen the same problems. FYI: ours has 75MHz R8k chips, so it is not blindingly fast to begin with. Of course, this shouldn't affect the parallel speed-up. One thing to make sure of is that all of your processes are really getting split onto all of your processors. It is quite possible for you to allocate 2 slaves and have them BOTH placed on the same CPU - thereby doubling the run-time of the code instead of halving it. I think that when the OS "sees" other processes (editors, calculators, etc.) running on CPU1, it assumed that it was loaded and hence put the next couple of processes (the PVM codes) on CPU2. Makes sense in terms of overall throughput of the system, but doesn't help if you're trying to do timings of parallel runs. Does anyone know of a way around this? can you "force" PVM to put one slave on CPU1 and the next on CPU2, etc.? (I haven't used the machine in a while, so I forget what utility you can use to view per-CPU loading) As mentioned by others, check that your algorithm is not doing too much communication and not waiting at barriers unnecessarily. Also give some consideration to I/O - it may be conflicting with your other communications. John Pormann