Jonathan, I am interested in what you mentioned about using 1.5 "peers". I googled it, but a little confused about "peers". Is "peers" used for Failover or for Load-Balancer scaling? I do not find a detail information how "peers" is working. If you have one in handy, could you recommend it to me?
Thanks Again, Q.Xie On Friday, December 6, 2013 10:17 AM, Jonathan Matthews <[email protected]> wrote: On 6 December 2013 17:50, Qingshan Xie <[email protected]> wrote: > Godbach, > Thanks for the quick reply and suggestions. > To enable multi-process mode does increase the capacity but limited by host > NIC bandwidth. Can HAProxy be scaled to multi-node to host the same > traffic? Yes it can, but *you* have to make sure the traffic arrives at multiple nodes, distributed correctly. HAProxy plays no part in the decision of "which physical HAProxy node will handle this traffic/packet/request?". I believe that you can then use 1.5 and its peer support to co-ordinate the multiple nodes' actions once the traffic /has/ been distributed, but that's your job to do - not HAProxy's. [ I believe this is the case, but only from reading this mailing list; I've not run 1.5 in production yet. ] Of course, you could combine a "dumb" layer 4 proxy (perhaps HAProxy in TCP mode?), sitting in front of a more involved layer 7 proxy running on more hosts. But this is only a fix for CPU contention: you'll still be constrained by the NIC in the front proxy, unless you arrange for the traffic to be distributed across nodes before it arrives, somehow. You may be interested in reading Willy's 2006 load balancing paper, which I still find interesting and useful to this day. It has sections talking about your problem: http://www.exceliance.fr/sites/default/files/biblio/art-2006-making_applications_scalable_with_lb.pdf Jonathan
