mc performance tuning; was: taint mode: insecure dependency in connect

[dormando]

Jeff McCarrell wrote:
> re: contention for memcache
> Well, we don't have a very standard kind of web app.
> The data show that running 8 memcaches is better for us than 1:
> 
> http://www.emptech.com/mc/image003.png

No no, I believe you see benefit when doing so :) I'm just not positive
something else isn't going on here :) I've yet to see a single instance
host actually overpower a memcached instance and not have it be some
weirdo bug.

It might be those multiget's, I guess.

> The graph shows the avg time to process the complete request.
> The X axis shows concurrent streams of requests being sent to apache.
> Each request makes several to small tens of memcache gets, and 1 multi-get.
> My next set of tests will compare 8 vs 16 memcaches and scale up the number
> of concurrent requests.
> 
> re: localhost:12211 vs unix domain sockets
> Um, yes, I can test tcp sockets vs. unix domain sockets if there is a good
> reason to.  My belief is that unix domain sockets are the fastest in
> general; is there something specific to memcache that suggests otherwise?

They're actually the same speed on linux these days, or thereabouts. Was
curious to rule it out, is all.

> re: multi-threaded memcache
> I read about multi-threaded mc with interest.
> In general, I'm leery of threading, but I would be willing to look further
> into it with a little encouragement. :-)
> We are running LAMP on linux 2.6 64-bit where the p is perl 5.8.5.
> Is there a consensus view that memcache threading works well on that
> platform?

Threaded mode is the new default. It works well enough for most of us.
Should be more (memory) efficient for you than multiple independent caches.

-Dormando