When we are under high load (10k+ requests per second) we start
getting communication failures (connection timouts,closed connection
during header) to memcached.

After a small period of time in this state we start getting mangled
returns from the cache.  Some returns will contain two pages data,
some will simply be the wrong entry.

We can recover from this by resetting the memcache servers.

Is there some connection concurrency limit we may be bumping into
under high load that would cause communication channel mismatches,
such that request 1 will somehow receive request 2's answer?

We are using memcached 1.4.11 with the following config on x86_64

-d -p 11211 -u memcached -m 1024 -c 1024

Reply via email to