Re: Response time increases over time

Thanks for that Mario, I will check it out.

@All:
Anyway, I have compiled 9.1.2 from source, and unfortunately the performance haven't got better at the same load, it is consistently quite low (~70 ms average transaction time with 100 clients) on this Debian. I am quite surprised about this, it is unrealistically high.
I have run pg_test_fsync, and showed about 2600 fsync/sec, which means HDD has write caching on (it is a 7200 rpm drive, there is no HW RAID controller). However my other machine, the simple Win7 one, on which performance was so good and consistent, fsync/sec was a lot lower, only about 100 as I can remember, so it probably really flushed each transaction to disk.
I have also run load simulation on this Debian machine with InnoDb, and it performed quite well, so the machine itself is good enough to handle this. On the other hand it is quite poor on Win7, but that's another story...

So there seems to be something on this Debian machine that hinders PostgreSQL to perform better. With 8.4 I logged slow queries (with 9.1 not yet), and almost all were COMMIT, taking 10-20-30 or even more ms. But at the same time the fsync rate can be quite high based on pg_test_fsync, so probably not fsync is what makes it slow. Performance seems to degrade drastically as I increase the concurrency, mainly concurrent commit has problems as I can see.
I also checked that connection pooling works well, and clients don't close/open connections.
I also have a graph about outstanding transaction count over time, and it is quite strange: it shows that low performce (20-30 xacts at a time) and high-performace (<5 xact at a time) parts are alternating quite frequently instead of being more even.
Do anybody have any idea based on this info about what can cause such behaviour, or what I could check or try?

Thanks in advance,
Otto