memcached performance on Mac OS 10.4

[Brad Fitzpatrick]

If you rebuild libevent with kqueue support, and disable TCP_NOPUSH
support in memcached (which it tries to auto-detect), then memcached will
be fast on OS X, but you're limited in your number of concurrent
connections.  Or rather, it'll eat more CPU as you have more clients
connected because it has to poll()/select().

The problem, as I understand it, is:

   kqueue is auto-detected, but sucks on OS X.

   NOPUSH is auto-detected, but sucks (on OS X and FreeBSD?) or its
     semantics diff from TCP_CORK on Linux.  not quite sure.

In any case, a coworker of mine using memcached on OS X (for development,
not production), had success with the above.

- Brad


On Sun, 15 Jan 2006, Peter Bengtson wrote:

> I'm experiencing the previously reported 200 ms latency on each
> memcached query. As five queries per second is far, far worse than
> scrapping memcache altogether and running without memory caching, I
> was wondering if the problems on Mac OS 10.4 have been addressed and/
> or worked around in any way since Gregory Block wrote the following
> to this list in March last year:
>
> > I had a bug open on Radar regarding memcached causing core dumps when
> > used in poll(), kevent(), or anything other than select.  I've been
> > informed by the team that there's a fix in place for Tiger, and that
> > they've tested under tiger and poll() without seeing the kernel
> > panics.
> >
> > So...
> >
> > Until that fix is in place, it's memcached + select() on mac os x, or
> > watch your kernel go tits up.
>
> Of course, libevent since v1.1 has a "workaround", but memcache when
> using libevent 1.1a and started in the following fashion:
>
> 	/usr/local/bin/memcached -k -uroot
>
> still reports:
>
> 	[warn] kq_init: detected broken kqueue; not using.: Bad file descriptor
>
> and then proceeds to respond exactly five times a second.
>
> Also, James Robinson wrote, on Aug 30:
>
> > stores of payload between 14000 and 195846 bytes all happen on OS X
> > without delay. Apparently OS X's TCP_NOPUSH is just plain busted, as
> > seems to be the net wisdom garnered from google.
>
> So, I'm wondering what is happening here. Are the Mac OS 10.4
> implementations of kqueue and TCP_NOPUSH still broken? Are there any
> patches from Apple or any third-party solutions? And is the slow
> speed we're seeing the result of select() being inherently slow on
> Mac OS X? Is there another way of forcing memcached use poll()?
>
> 	/ Peter Bengtson
>
>