08-01-2009 04:37 AM
I do have different real servers from the performance point of view. I want to reduce the load of the low power machines and I would like to send some more sessions to the high performance sessions.
Thanks for your help.
08-01-2009 05:34 AM
This is not a simple question because it depends on the real server, the traffic type and the type of clients. I would suggest to use weights in case all clients requests / connections are very similar. Similar requests will result in a similar load per request and 10 requests do result in 10 times more load than a single request. I would add some weights to your config and I would move to weighted round robin or maybe weigted (enhanced weighted). Use higher values as weights because that is more granular. I tend to use weight 50 for an averga servers, lower values for slower servers and higher values for a faster server.
It is a little different in case the real servers are different and the clients are different as well. 5 requests coming from so called power clients might eat up as much resources @ your server as 100 normal requests. You might want to look at the "dynamic weighted predicto" which is using values retrieved from the real servers (via SNMP) as weights - this allows it to adjust weights dynamically.
It is as well an option to use static weights in such a scenario but you might to limit the maximum mount of connection per real to keep the load per real at an acceptable level. And you might want to look at PBSLB to bind power users to power servers - PBSLB is Layer 4 load balancing with some restrictions:
There is no solution which fits in every case as you can see - try to check different predictors and look for the best one.