first time user with out of memory question
pault12345 at yahoo.com
Mon Jun 12 21:52:03 UTC 2006
19K hps certainly is "heavy pressure". In our current
setup we are not (yet) getting close to those numbers
with memcached. But then again - in our setup
memcached (latest stock memcached, built and started
with defaults) got crazy after 3 weeks.
19K hps of reads *does*not* subsitute for a
stress-test consisting of 10K hps reads mixed with 5K
hps of writes (for example). In async application
under pressure just changing pconnect to connect can
make a (critical) difference - changing the memory
Basically that was the only point I was trying to make
and it seems that we agreed on it several emails ago,
so I am sorry for being redundant.
BTW - the note that you made about USE_SYSTEM_ALLOC is
not 100% correct, because - from the code - the
USE_SYSTEM_ALLOC only affects part of the slabs system
(the 'heap' part, which is important part but it is
still - a part).
It is not actually that easy to entirely eliminate
slabs from memcached codebase.
Unfortunately, digging deeper will open another can of
worms. Linux is open source too - it envolves. And
frarmentation issues are with us since the days of
RT-11, so there is a good chance that memcached is
already fighting with the ghosts - in some setups.
Thanks for the numbers.
--- Steven Grimm <sgrimm at facebook.com> wrote:
> Paul T wrote:
> > I think that nobody on this list is running
> > under heavy pressure, though, so this bug can be
> > happily ignored in many cases.
> As I type this, we are running about 19,000 requests
> a second through
> some of our production memcached instances.
> There is a reason we spent time reworking
> memcached's network I/O to
> reduce its CPU consumption: we were getting
> uncomfortably close to
> running out of CPU capacity on our memcached servers
> (and these are not
> slow old machines). There's also a reason we added
> UDP support: our
> memcached instances were each handling tens of
> thousands of simultaneous
> client connections, enough that the per-connection
> overhead was starting
> to become significant. If that doesn't count as
> "heavy pressure" I'm not
> sure what does -- what does your load look like?
> (And what hardware do
> you run memcached on, that can support significantly
> heavier traffic
> than that?)
Do You Yahoo!?
Tired of spam? Yahoo! Mail has the best spam protection around
More information about the memcached