We're going to build a new server with a bit more RAM -- 8 gigs. I know memory is somehow the bottleneck, but I don't think it's because I have too little. Performance monitors show a lot of swap activity during heavy query load testing, but available RAM is never maxed out, or even close. SOMETHING is starved for RAM, I just don't know what. We are running a web app here and have Apache and PG on the same box. I have PG buffers set very high, about 100K. Data set approaches 3 gigs. My gut feeling is I have RAM to spare, I'm just somehow not efficiently using what I have. I just wanted a gut-check reaction to this idea, and I don't hear anyone saying "NO! NEVER!". What I'm hearing is that it might, or might not, be practical, feasible, maintainable. Thanks for the replies. I'll look more at the memory usage and see what comes up. -- sgl > From: Chris Travers <chris@xxxxxxxxxxxxxxxxxx> > Date: Fri, 29 Jul 2005 09:50:07 -0700 > To: Steve Lane <slane@xxxxxxxxxxxxxxxxxxxxx> > Cc: <pgsql-admin@xxxxxxxxxxxxxx> > Subject: Re: [ADMIN] Indexes on RAM disk = insanity? > > Steve Lane wrote: > >> All: >> >> We have a postgres 7.4 server where we're trying to achieve some speedups. >> Right now, at least superficially, RAM appears to be the bottleneck -- lots >> of swaps in and out. >> >> There is another consultant beside myself in the mix and he asked this >> question: can we put the database indexes on a RAM disk? Won't that speed >> things up? >> >> > I am quite wary about putting the indexes on a RAM DISK for another reason. > > You say that RAM is your bottleneck, so putting anything unnecessary in > RAM seems like a good way to get less performance rather than more. > Additionally you have some overhead in tracking the files, etc. and they > will remain in RAM even when they are not used. This leaves the system > with less effective RAM for the memory intensive operations. > > Such a move might make a lot of sense if you had a LOT of RAM but disk > I/O was the bottleneck. However, if a lack of RAM is your problem, > putting more stuff in RAM doesn't seem very sound to me. > > Best Wishes, > Chris Travers > Metatron Technology Consulting
