Re: Clustered Store

--MimeMultipartBoundary
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit

Andres Kroonmaa:

> I'm running squid with these patches for quite some time now. Cache is

Me too (of course :-).

> very full, and "emergency LRU" is not so rare. What can I say about
> performance? Compared to default squid its a tremendous speedup, as
> Arjan has pointed out on his web page. Dayly (10-20h) average time
> for open(RO) is at 6.4 mSec, open(WR) is 12 mSec.
>
> I've added some measurements for unlink() calls into squid, and found
> that unlink() calls take 40 mSec on average (with some maxes at 100mSec).
> So, unlinks are really very expensive.
>
> Now, if compared to Arjan's results, this seems not quite so good, but
> I'm not sure if Arjan's cache was filled and into LRU problems by the
> time he did his samples. Perhaps he can clarify. When our cache was

Earlier results were taken just after introducing the patch. I have more
data now, see below.

> not full, the times for RO/WR were pretty close at 4 mSec for both.
> (Our cache size is 14GB (1M objects) and doing 50K TCP reqs/h, NOVM.18 train)

I've been gathering statistics from two weeks before I implemented the new
storage method up til now and here are the results. All measurements are
done daily from 08:00-18:00 (business hours on working days only).

                before (2 weeks) after (2.5 months)
----------------------------------------------------------
open for read 20-25 msec/call 15-20 msec/call
open for write 27-35 msec/call 7-10 msec/call

The cache size is only 4GB (2*2GB) now, but a 20GB RAID array is waiting to
be connected to it. The cache has 1.0 to 1.4 million hits per day at the
moment.

Arjan

-- 
Arjan de Vet, Eindhoven, The Netherlands            <Arjan.deVet@adv.IAEhv.nl>
URL: http://www.IAEhv.nl/users/devet/       for PGP key: finger devet@IAEhv.nl
--MimeMultipartBoundary--