3 ms·
Running a VoIP platform, we've "discovered" that instead of running one big process (which spawns lots of threads), we run one process per core or socket (stil
by MichaelGG 13y ago
Running a VoIP platform, we've "discovered" that instead of running one big process (which spawns lots of threads), we run one process per core or socket (still experimenting). We force affinity so the scheduler can't bounce stuff around. We've gotten somewhere between 200% and 500% efficiency from doing so, as well as minimized variation (that is, the one-process-per-server machines see huge CPU spikes, whereas the affinity locked systems don't).
The way I see it: You've gotta figure out your scale-out or load balance method. It might not hurt to just pretend each NUMA node is a separate server and just treat it like that.
- seanmcdirmid 13y agoThat is not the point of NUMA at all. If you are just going to use RPC to integrate a bunch of computers in a cluster, why would you want to bother with the Non Uniform Memory Access abstraction at all? NUMA is kind of out of fashion though. For everything it claims to do, there are better alternatives (MPI, map reduce, actors, ...).
- kasey_junk 13y agoWhat in the world do actors & map/reduce have to do with memory locality?
- seanmcdirmid 13y agoThey involve alternative communication mechanisms that are not shared memory.
- kasey_junk 13y agoNo they are abstractions around shared memory. Actors are a good example of where having an understanding of NUMA can have serious performance implications. If you architect your actor system so that all of your actors are on the same socket, all message passing (assuming messages are small enough) can happen in socket local cache. This will be much faster than having to cross socket memory or even worse going to main memory. Conversely, if you aren't paying attention to memory locality you could see unexpected performance degradation as you add actors, because even though your system is more parallel there is more overhead in memory access.