Forum Discussion

Chris_Rizzo's avatar
Chris_Rizzo
Icon for Nimbostratus rankNimbostratus
Oct 28, 2010

GTM Load Balancing

In my GTMs, I've implemented the least connections method of balancing. (My wide IPs each contain one pool. Within each pool is a number of virtual servers, from various datacenters.)

 

 

This seems to be working well, as the connection numbers are very evenly matched within any given pool. My problem is this...I have some situations where the number of servers, in some of these pools are not even. So, for example in a particular pool, site A has 50 servers, and site B has 75 servers. The connections, for the pool, are evenly divided because of the least con method, but I would like it to more reflect a dynamic ratio that maybe could include connections per server, or at the very least the number of servers, considering all servers are equal (from iqdump, it looks like there is a value for this - 25.

 

 

Put in another situation, if I have 100 servers evenly divided between sites, and I lose 25 servers in site A (being extreme), I don't want all connections to be balanced, as one site has has 25 servers now, and the other has 50. Site A may get overloaded, so I would want site B to get a higher ratio of connections.

 

 

It looks like the only dynamic method for this would be QOS, but does anyone have any good suggestions/ideas for something like this?

 

  • I'd consider having a pool for each datacenter, that way you can have two layers of load balancing, first between datacenter at the wideIP level, then between virtuals within that datacenter at the pool level.
  • I'll have to try this. But, would this allow for the dynamic type of balancing that I mentioned? (I'm looking at this now...)
  • Either way will allow for the dynamic type, but as you indicated, you'd need to define the qos algorithm that strikes the right balance of connections/throughput/etc. It's really knowing your environment and working out the layers, if they're needed, to suit the resources. For example, if you have two datacenters and want to load balance connections between them, and then load balance within the datacenter, least connections at the wideIP level might not be a good idea if you have 10 servers in DC1 and 50 servers in DC2. But you would most likely want to do least connections within the datacenter, assuming each virtual server fronts an equivalent number for servers.
  • Thanks. This is exactly what we are doing. We use least connections in the LTM layer, and also in the GTM layer, except in cases where we have mismatched environments. Now the problem is trying to figure out the best values for the qos algorithm, for those "odd" cases. Thanks again.