On 2010-04-28 19:10, Alex Forrow wrote: > We're looking to upgrade our HAProxy hardware soon. Does anyone have any > recommendations on the things we should be looking for? e.g. Are there > any NICs we should use/avoid?
Hi Alex, I'm just writing down here what comes to my mind. Sorry if it looks a bit unorganized... Haproxy itself is not very demanding. A two core system will suffice. Check to have enough RAM to hold all your sessions, but since it's rather cheap to get 4 or 8 Gigs you should be safe here :) Always think about the resource demands of the TCP stack. ON large loadbalancer instances (esp. with many short connections), the TCP stack will consume much more resources than your Haproxy. Some NICs allow to offload some of the responsibilities like calculation of packet checksums to silicon. Interrupt mitigation is something you most probably want to have. Normaly, each packet will trigger an interrupt which will eat away all your ressources if there are many of them. Some NICs allow to cap the number of interrupts per second which might increase latency a bit but saves your load balancer from dying :) Make really sure your intended NIC is very well supported by your intended OS. Many show suprising behaviour under stress. So the best advice would possibly be to have a look in the vendors hardware support lists and ask in the respective channels. You most probably want to stay away from most on-board NICs from vendors like Broadcom or SiS. Dedicated PCIe NICs from Intel are normally safe (you find them also on some server boards from e.g. Supermicro). But make sure to check the individual capabilities. As a loadbalancer is always IO bound, check your data paths. Most interestingly is the speed from and to the NIC (in a way that the network-line is always the bottleneck) and between memory (ECC of course) and CPU. Harddisks are obviously uninteresting :) Hope this helps, --Holger
