[Pgpool-general] replication stability under heavy loads, many connections?

[Andre Felipe Machado]

Hello,
What are suggested configurations for 
num_init_children
max_pool
for heavy replication mode loads?
we have 10 webservers connected to pgpool2.3.4-unreleased_cvs to two
backends pg 8.3, all Debian Lenny virtual machines over XenServer 5.6.
We may configure ram and cpu up to 32 Gb and 16  true cores "if needed",
but currently using 4 GB and 8 cores to the pgpool Debian Lenny VM.
The pg backends were tuned at another Debian Lenny VMs (at other
physical Xen servers) and cope with the load.
The php app is not well behaved, and it is not tolerant to almost any pg
latencies. 
Its profile is 95% selects and 5% inserts/updates.
So we started with thousands connections (with kernel adjustments) and
max pool left 4, then 3, then 2, then 1.
Also reduced children reuses
We also reconfigured tcp_tw_reuse kernel to 1 and abandoned tw recycle
for network speed.

After reducing to 1k children, it worked for a week with low latency.
When pgpool crashes it locks processes needing a killall to zap them.

What are recommended confs for high replication loads, low latency?
Should we go the opposite, increase max pool and reduce init children?
the php web app concurrently connects as the same db user.
suggestions?
Regards.
Andre Felipe