> > Could you give us some more infos about the box' performance while you > > run the PG benchmark? A few minutes output of "vmstat 10" maybe? What > > does "top" say? > > > > Here, an extract from the vmstat 3 during the test, you can see that > my problem is probably a very high disk usage (write and read). > > procs -----------memory---------- ---swap-- -----io---- --system-- ----cpu---- > r b swpd free buff cache si so bi bo in cs us sy id wa > 0 11 92 128344 9224 2428432 0 0 287 9691 2227 685 4 3 0 93 > [...] Yes, as is the case most of the time, disk I/O is the bottleneck here... I'd look into everything disk releated here... > > How are you using the 3 disks? Did you split pg_xlog and the database > > on different disks or not? > > > > Data are on disk 1 et 2. Index on disk 3. Perhaps i'm wrong but fsync > = off, pg_xlog are running with that ? Yes, pg_xlog ist also used with fsync=off. you might gain quite some performance if you can manage to put pg_xlog on its own disk (just symlink the directory). Anyway, as others have pointed out, consider that with fsync = off you're loosing the "unbreakability" in case of power failures / os crashes etc. > > Can you say something about the clients? Do they run over network from > > other hosts? What language/bindings do they use? > > > > Client is another server from the same network. Clients are connected > with JDBC connector. ok, don't know about that one.. > > When they do inserts, are the inserts bundled or are there > > single insert transactions? Are the statements prepared? > I use prepared statements for all requests. Each transaction is about > 5-45 requests. sounds ok, could be even more bundled together if the application is compatible with that. Bye, Chris.
