Thanks Lukas and Baptiste for your advices, which are very helpful. It is clear to me that, the "peers" setup is needed for scaling the multi-master haproxy instances. According to Peers doc at http://cbonte.github.io/haproxy-dconv/configuration-1.5.html#3.5, the stick-table is used for sync multi haproxy instances. Do you know if this stick-table is viewable and manageable? If yes, could you tell me how?
Thanks Again, Q.Xie On Monday, December 9, 2013 3:18 PM, Lukas Tribus <luky...@hotmail.com> wrote: Hi, >> Hello Experts, >> not sure if this subject was already discussed or not, like to hear the >> advices and suggestions. >> If a single HAProxy instance as a load-balancer could not handle the >> high-load traffic, how to scale multiple instances as a group of >> load-balancers to handle the high-load? > > [...] > > You can Load-Balance your HAProxy servers using LVS, which is more or > less a packet forwarder. > It's dumb but very fast and can sustains millions of connections > (since it manages only packets, it requires much less memory than > HAProxy). If you need to grow even larger, some DNS tricks may come in handy, like active-active round-roubin and geolocation based redirect. If even that is not enough and you need to scale horizontally, then you can load balance the incoming traffic at your routers via ECMP. This scales as long as your router has enough bandwidth and ports. Also, some folks (cloudflare) use anycast even with TCP traffic like HTTP and HTTPS to scale. But you need to consider the downsides of anycast with TCP very carefully and design your network to compensate for it. You do not want to "just switch this on". Regards, Lukas