Re: chekcpoints, BUFFER settings, and optimization

David Williams <djw@smooth1.demon.co.uk> wrote: > No! Surely the whole idea is to minimize disk I/O. When possible, yes. But a certain amount of disk i/o is inevitable. And we don't want to do it all in one lump during which all other activity is halted - you want to do it a bit at a time, behind the scenes, when other activity is not blocked. Pay me now, or pay me later with interest. > >>the write cache. But as long as you are not thrashing the hell out of your >>disks, who cares? Writes are not blocked from the transaction's perspective, >>so the impact is general and thus typically diffuse in nature (as opposed to >>writes occuring during checkpoints, during which time transactions are blocked >>and thus drastically affected). > > But these writes still slow checkpoints. Huh? I think you miss my point completely. You want to minimize the amount of writing at ckpt time, as ckpts have an adverse affect on throughput. If this means doing more i/o in between ckpts, so be it. If you are using KAIO, which most should at this point, the kernel should be optimizing the i/o such that thrashing is minimized. If you assume optimistic cacheing, then the writes should not have too much of an impact on query performance (i.e. reads). Dave "That's just my opinion - I could be wrong." - Dennis Miller