No ;) I said "Usually No" to the question "Is memcached UNABLE to handle large number of connections?". It is negative of a negative and actually positive answer.
And yes you are right about all the things you said about the context in which memcached is used. I didn't meant to say anything against memcached and I think it is a wonderful piece of software solving real world problems. But I do think 256MB limit is an overkill. It might be useful for testing max throughput, but it can cause practical problems to users. Bug in one the client machines can hold up memory in server, causing problems to other clients. Thread death, busy cpu, memory pressure, anything that slowes down few of the clients can add up memory pressure on memcached. The reason I mention is this: I can think of this as the ONLY cause of memcached not able to handle large number of connections. I have read the code. Multiple times. Nothing else. thanks! rohitk On Mon, Jul 29, 2013 at 12:30 AM, dormando <[email protected]> wrote: > > Yes. > > OS doesn't pre-allocate the full buffer. > > > > Yes. > > Memcached should easily handle million concurrent idle connections > (given enough memory). > > > > Except, if user has only allocated 512MB/1GB RAM for TCP, it is > "possible" at runtime only four connections have eaten up all the memory, > which > > could lead to connection refused or some other errors. > > > > Lets revisit the question. > > Is memcached unable to handle large no of concurrent connection? > > Usually no. Under some circumstances. Yes. > > > > And yes, don't guess. Just try it. Open <server_tcp_mem/256MB> > connections with large multi-get request (response > 256MB) and don't read > them at > > client side. And then try opening more connections. > > Usually no?? Are you sure you mean usually it can't handle it? That is > insanely wrong. > > Memcached runs on LAN's almost all of the time. There are almost NO > buffers stuck in use because of the low latency. This isn't an internet > facing tool, wherein you have to tune that more carefully and leave a lot > more free memory for TCP retransmits. Connections to memcached use a > handful of kilobytes. > > So very few people are going to run into this problem, complaining about > it is nothing short of alarmist. > > It's also never going to be 256MB: The actual memory used is limited (in > linux) by the tcp_rmem and tcp_wmem set of sysctl's. Even when people > aggressively tune those, they set the maximums around 16 megabytes. > Usually it's much lower than that. SENDBUF can't use more than what's in > wmem. > > -- > > --- > You received this message because you are subscribed to the Google Groups > "memcached" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > For more options, visit https://groups.google.com/groups/opt_out. > > > -- --- You received this message because you are subscribed to the Google Groups "memcached" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/groups/opt_out.
