Hi Lukas, Thanks very much for your confirmations and the Stackoverflow post. I just want to expand a little on the last question relating to websockets connections.
Let's say I have a backend which exclusively deal with long-lived connections: websockets and xdr-streaming. These connections tend to stay open for 30-60 mins, during which both the server and clients send multiple JSON msgs. The server processing/response time stays on average <10ms per msg. In a way, you can think of this backend as a game server which reacts upon receiving a msg from a client. Currently, I have only one process of this backend running with HAProxy on the same machine (actually VPS). The machine has 8-core cpu and 1g memory. I don't expect huge amount of traffic for now. Given this setup, what would be a sensible 'maxconn' number I should set for the server option of this backend? Of course, I can estimate roughly the number of concurrent websockets/streaming connections to my website at any point of time, and use that number for 'maxconn' on the server. But I feel there must be a better way to figure out this number optimally. Some pointers will be much appreciated. Thanks! Best, Merton On Tue, Apr 16, 2013 at 3:52 PM, Lukas Tribus <[email protected]> wrote: > Hi Merton, > > > Please let me know if the above understanding is correct. > > Yes, thats the concept. > > > > > if I have multiple backends (multiple server options), does the > > sum of 'maxconn' of their server options have to be no more than > > the 'maxconn' of their corresponding frontend? > > Yes that does make sense, although you are probably able to do a > little oversubscription. > > > > does specify a low 'maxconn' for a server option help improve the > > performance of that particular server in any way? > > Yes it does, because it makes sure your backend servers is never > overloaded (by dispatching requests to backends with lower load or > queueing them). In fact iirc some of the ruby folk runs their backends > with maxconn 1, so a server always handles just one request at the > time. > > > > if the connections to a particular backend are long-lived, e.g. > > websocket connections, does it imply that I should set a relatively > > large 'maxconn' for the server? > > Probably yes. The question is, how many of those long-lived sessions > are concurrently active on how many of them can your backend serve? > > > Here are a few thoughts from another user and a response from Willy about > queueing: > > http://stackoverflow.com/questions/8750518/difference-between-global-maxconn-and-server-maxconn-haproxy > > > > Regards, > Lukas
