[f-nsp] Response-time balancing

[Vlad]

Hello,

we are using ServerIron XL to load-balance http traffic configured
with DSR. Real servers hardware is pretty different, so I thought
using response-time predictor logic is adequate. I've written a simple
script that checks real server load average and delays it response
proportionally to that LA. That script is specified like this for each
real server:

 port http keepalive
 port http url "GET /cgi-bin/la.cgi"


what I've being observing is ServerIron distributes requests pretty
evenly according to each real server processing capacity. However
every once in a while, all of the sudden SI would send tons of new
connections to a single real server, overloading it for few moments.

To fight that I tried to setup balancing by "Least connection and
server response time weights", but documentation is not quite clear.

I assume that each real server should be given "weight X Y"
configuration option, where X represents connections weight, Y - its
response time weight. It's not clear what "predictor" should be set to
account for both values - to weighted or least-connection or
response-time... I tried every single with no luck - it either
balances on connections or response time but not combination.

Did anyone has a similar problem or ideas on optimal load balancing
when real servers has very uneven processing capacity.

--
Vlad