On Fri, Mar 2, 2012 at 2:51 PM, Ripduman Sohan <ripduman.so...@gmail.com> wrote: > On 2 March 2012 01:55, Wendy Cheng <s.wendy.ch...@gmail.com> wrote: >> Hi, >> >> We plan to start tuning work for our Memcached setup. The binaries >> were built on top of "engine_pu" branch (storage engine interface). A >> quick test run (everything on default, including storage engine >> itself) using >> >> "/usr/local/bin/memslap --server xyz --test=get --concurrency=1024" >> >> got a bunch of "Failured on read " errors. Using "gdb" to dig a little >> bit, nothing in "default_get()" shows failure. Before we continue the >> trace tomorrow, if folks can pass some tuning tips (particularly >> around the network stacks), it'll be greatly appreciated. >> >> Thanks, >> Wendy > > Wendy, > > Solarflare did some work previously on network stack tuning which may > be useful to you: Link is at: > http://10gigabitethernet.typepad.com/network_stack/2011/09/memcached-goes-faster-with-onload.html > > --
Thanks .. (your link) look like a NUMA machine ... Anyway, our problem seems to be a DNS lookup scaling issue (fail at getaddrinfo() with error code 51, not sure how "51" exactly means though) . We are now running with 2 test clients each with 800 concurrency. We'll come back to this issue eventually. But for now, 2-clients setup works for us (as the focus is with server's storage engine). -- Wendy