Linux/PostgreSQL scalability issue - problem with 8 cores

Hi all,

we have a PostgreSQL dedicated Linux server with 8 cores (2xX5355). We
came accross a strange issue: when running with all 8 cores enabled
approximatly once a minute (period differs) the system is very busy for
a few seconds (~5-10s) and we don't know why - this issue don't show up
when we tell Linux to use only 2 cores, with 4 cores the problem is here
but it is still better than with 8 cores - all on the same machine, same
config, same workload. We don't see any apparent reason for these peaks.
We'd like to investigate it further but we don't know what to try next.
Any suggenstions? Any tunning tips for Linux+PostgreSQL on 8-way system?
Can this be connected with our heavy use of listen/notify and hundreds
backends in listen mode?

More details are below.

Thanks,

Kuba

System: HP DL360 2x5355, 8G RAM, P600+MSA50 - internal 2x72GB RAID 10
for OS, 10x72G disks RAID 10 for PostgreSQL data and wal
OS: Linux 2.6 64bit (kernel 2.6.21, 22, 23 makes little difference)
PostgreSQL: 8.2.4 (64bit), shared buffers 1G

Nothing else than PostgreSQL is running on the server. Cca 800
concurrent backends. Majority of backends in LISTEN doing nothing.
Client interface for most backends is ecpg+libpq.

Problem description:

The system is usually running 80-95% idle. Approximatly once a minute
for cca 5-10s there is a peak in activity which looks like this:

vmstat (and top or atop) reports 0% idle, 100% in user mode, very low
iowait, low IO activity, higher number of contex switches than usual but
not exceedingly high (2000-4000cs/s, usually 1500cs/s), few hundreds
waiting processes per second (usually 0-1/s). From looking at top and
running processes we can't see any obvious reason for the peak.
According to PostgreSQL log the long running commands from these moments
are e.g. begin transaction lasting several seconds.

When only 2 cores are enabled (kernel command line) then everything is
running smoothly. 4 cores exibits slightly better behavior than 8 cores
but worse than 2 cores - the peaks are visible.

We've tried kernel versions 2.6.21-23 (latest revisions as of beginning
December from kernel.org) the pattern slightly changed but it may also
be that the workload slightly changed.

pgbench or any other stress testing runs smoothly on the server.

The only strange thing about our usage pattern I can think of is heavy
use of LISTEN/NOTIFY especially hunderds backends in listen mode.

When restarting our connected clients the peaks are not there from time
0, they are visible after a while - seems something gets synchronized
and causing troubles then.

Since the server is PostgreSQL dedicated and no our client applications
are running on it - and there is a difference when 2 and 8 cores are
enabled -  we think that the peaks are not caused by our client
applications.

How can we diagnose what is happening during the peaks?