(would it be possible for you to configure your mail software to add attribution lines when quoting? it makes it much easier to follow who said what)
On Mon, Aug 11, 2003 at 05:22:11PM -0700, Jim Sibley wrote: > >I think that the word you're searching for is "pathological". > > Its only pathological in the sense that is forces the condition to > emphasize its effect. We did not start with this case. It's still pathological, even if it was done intentionally (perhaps especially so). > What we did notice is that a data base random write was much slower than a > data base random read for a large data base. Examination showed that no > I/O was actually taking place because of the Linux caching of buffers and > retention of dirty buffers. (and as I said before, no paging). > > We first eliminated the I/O entirely by reading/writing to /dev/shm and > saw the same effect - the random reads were much slower than the > sequential reads to the data base. First you were comparing random write to random read, then random read to sequential read. Is this correct? That would seem like (potentially) a different problem. > Throwing more hardware at it won't solve this poor performance problem. > Adding CP's or memory may aggravate it! This is not generally true, though it may be in a consolidated scenario. For example, if your working set is 1GB, and you can efficiently segment that working set and forward requests to N physical servers rather than a single one, each of those servers will have a smaller working set and be more efficient in itself. If these are instead virtual images, the effect is less predictable (at least for me). > The only solution that I see is reducing locality of reference (working > set size). Reducing working set size increases locality of reference (which is what you want). -- - mdz
