RE: [squid-users] TCP_MEM_HIT performance

On Wed, 7 Apr 2004, Francisco Lopez wrote:

> I don't agree with this point. Since:
> A) there are (on the top users test) 150 simultaneous users
> B) squid doesn't put in memory objects not received from the network

Still you only have 500k data. This absolutely for sure fits in memory on
any server.

If not in the Squid hot object cahce then it is in the filesystem buffers.

> All of those requests 150 (ok, not all of the users are simultaneously
> requesting pages but the number should be around that) must be read from
> disk. This is pretty high i/o and filesystem performance should be a
> factor.

With this test profile there should only a minimal burst of disk I/O at
the start of the test. Then nil (or very close to).

> Again, I think that on high load AUFS should perform better and 150
> users sending requests is generating a lot of i/o on the server (unless
> the store manager is keeping some kind of memory pool, in that case I
> agree that even when there are 150 requests only 10 of them are unique;
> since there are only 10 different pages in my test).

Number of concurrent users is not of relevance. Total volume of the
frequently requested objects is, and to some extend how many different
objects are requested at the same time.

As long as the total volume of your objects fits in memory there is no
disk I/O. Only when the disk I/O is significant does aufs make a benefit
compared to ufs.

> According to the tests published in the Squid book by Duane Wessels,
> AUFS should be faster. My system has 2 PIII, 1Ghz, and I don't think CPU
> is an issue here.
> Or is slower because of the cache magnitude you say ?

Correct.

Your test is not using the disk, only memory. As memory is very fast
compared to disk the tradeoff made by aufs does not give any benefit and
you only have the additional overhead of sending I/O requests to threads.

> Any ideas of what could be wrong ?

Make sure you are not swapping.

   vmstat 5

page in/out columns is a good monitor.

the fact that you have free memory is not and indication that your system
is not swapping.

Regards
Henrik
Received on Wed Apr 07 2004 - 19:04:00 MDT