> Not 100% relevant but I found this to be interesting if you're nodes are > doing heavy disk I/O: > > http://rackerhacker.com/2008/08/07/reduce-disk-io-for-small-reads-using-memory/
There are some pitfalls though, or at least there were the last time I was tweaking such stuff for a PostgreSQL database. One issue I remember is that, at least at the time, there was a limit to how many dirty buffers *per inode* that the background dirty buffer flusher would write. So for lots of outstanding random I/O to the same large file, tweaking ratio triggers was not necessarily enough and your eventual fsync() still ended up pushing more data than you may have wanted. I also read a pretty good article (can't find it right now) that made a good point that some of the more obscure settings one might be tempted to tweak, should not necessarily *be* tweaked be cause the kernel had hard-coded assumptions about them (this does not apply to just the dirty background ratio though; IIRC it was things like the frequency with which the flusher runs. I never experimented with that though. -- / Peter Schuller
