From: Erich F. <fo...@es...> - 2001-10-10 09:15:55
|
> But I would expect that even on architectures allowing it, half-full > nodes would be quite rare. After all, memory latency is better within > a node than between nodes, so performance is going to be best if you > cram as many CPUs into each node as you possibly can. Therefore, > if someone configures a machine to have half-full nodes, they are > already accepting a sub-optimal result. Why impose added complexity > on other algorithms to help someone who has deliberately configured > their machine for suboptimal performance? > > Thoughts? A machine with half-full nodes is not necessarily suboptimal. The typical HPC loads I'm used to are memory-bandwith hungry, are often single threaded or MPI parallelised. So if the processors in a node share a bus, each can provide higher bandwidth to memory if it has to share the processor bus with less CPUs. Isn't this the reason why IBMs new Power4 comes with 4 CPU cores per MCM instead of 8 for the high performance computing configuration? Anyhow, I wouldn't boot the system with full nodes and hot-remove half of the CPUs... Regards, Erich --- Erich Focht <ef...@es...> NEC European Supercomputer Systems, European HPC Technology Center |