User concurrency thresholding: where do I look?

Folks,

I've run into this a number of times with various PostgreSQL users, so we
tested it at Sun. What seems to be happening is that at some specific number
of connections average throughput drops 30% and response time quadruples or
worse. The amount seems to vary per machine; I've seen it as variously 95,
1050, 1700 or 2800 connections. Tinkering with postgresql.conf parameters
doesn't seem to affect this threshold.

As an example of this behavior:

Users Txn/User Resp. Time
50 105.38 0.01
100 113.05 0.01
150 114.05 0.01
200 113.51 0.01
250 113.38 0.01
300 112.14 0.01
350 112.26 0.01
400 111.43 0.01
450 110.72 0.01
500 110.44 0.01
550 109.36 0.01
600 107.01 0.02
650 105.71 0.02
700 106.95 0.02
750 107.69 0.02
800 106.78 0.02
850 108.59 0.02
900 106.03 0.02
950 106.13 0.02
1000 64.58 0.15
1050 52.32 0.23
1100 49.79 0.25

Tinkering with shared_buffers has had no effect on this threholding (the above
was with 3gb to 6gb of shared_buffers). Any ideas on where we should look
for the source of the bottleneck?

--
Josh Berkus
PostgreSQL @ Sun
San Francisco