[OpenAFS] Re: Openafs performance with Apache/PHP

Nate Gordon nlgordon@gmail.com
Tue, 14 Aug 2007 11:42:33 -0500

So I loaded up ntop and isolated http, dns, ssh, afs, and mysql.  I
ran the previous test (ab -k -c <users> -t 120 <url>) again for
various numbers of users and here is what I came up with:
1 user: H:29.5M D:1.8M A:4.6K M:19.9M O:1.2M
2 users: H:50.3M D:3.0M A:2.0K M:33.7M O:241.9K
3 users: H:56.1M D:3.3M A:0K M:37.8M O:234.6K
4 users: H:56.2M D:3.3M A:0K M:37.9M O:127.9K
5 users: H:48.0M D:2.9M A:2.0K M:32.4M O:246.3K
8 users: H:23.4M D:1.4M A:2.3K M:15.8M O:135.6K
10 users: H:16.6M D:1.0M A:3.8K M:11.3M O:385.6K
20 users: H:16.7M D:1.0M A:0K M:11.7M O:235.5K
30 users: H:16.8M D:1.0M A:1.9K M:11.9M O:346.2K

H - http, D - dns, A - afs, M - mysql, O - other

This shows an interesting trend in that there is a peak level of
performance, as well as a worst case performance level as well.  It
also shows that afs traffic is essentially zero.  This would indicate
that my performance bottleneck is in accessing the afs cache.

If I can do further diagnostics/testing let me know.

On 8/13/07, Russ Allbery <rra@stanford.edu> wrote:
> "Nate Gordon" <nlgordon@gmail.com> writes:
> > I've looked through it before, but I usually get too annoyed when the
> > asfd process kernel panics my machine if I get dcache/stats too high.
> > My webservers deal with an annoying large volume of content (~35GB) and
> > defining a working set size seems to be a moving target to say the
> > least, but I'll take another stab at it.
> The other thing that would be very interesting to know is whether the time
> is going into data management in the cache manager or is really going into
> network traffic to the file server.  If you can use something like tcpdump
> to see how much Rx traffic there is to the file server under different
> loads and see if that increases with increasing request time or if it
> stays the same, that would be very interesting.  If it stays the same,
> that points to some sort of inefficiency in the cache manager itself with
> its data structure handling.
> --
> Russ Allbery (rra@stanford.edu)             <http://www.eyrie.org/~eagle/>

-Nathan Gordon

If the database server goes down and there is no code to hear it, does
it really go down?