Kris Kennaway wrote:
On Tue, Apr 17, 2007 at 02:51:55PM -0400, Kevin Way wrote:I built 7.0 as of 6 days ago, and ran the same test using 8-cores, ULE and 4BSD. The results are available at:http://blog.insidesystems.net/articles/2007/04/11/postgresql-scaling-on-6-2-and-7-0Unfortunately, I can't run the additional tests, as the machines in question have now been deployed to production. I should have similar equipment available in a few weeks, and will try the other tests at that point.That's better but still not comparable to my results. One important change you may have missed was to enable MALLOC_PRODUCTION in /usr/src/lib/libc/stdlib/malloc.c. This disables malloc debugging which may have significant overhead. My results from about a month ago: http://people.freebsd.org/~kris/scaling/nickel.png The big dip between 5-7 clients is believed to be due to a ULE bug that has known cause but is not yet fixed (on dual core CPUs processes are bouncing around between the two cores too much). The main point is that apart from this anomaly in the middle there is approximately linear scaling from 1 to 8 CPUs. On your system it is not scaling linearly beyond 4. I do have some other changes in my kernel but I don't think any of them are enough to explain this difference. Anyway, get back to us when you can retest with the above change.
Confirmed, I did miss MALLOC_PRODUCTION. I incorrectly believed that I just needed an empty malloc.conf. This is somewhat exciting news from my vantage point. I'll post a corrected re-test as soon as I have another available 8-core machine.
smime.p7s
Description: S/MIME Cryptographic Signature
