Re: Performance question

Duane Wessels wrote:
>
> On Sun, 13 Jun 1999, Andy Dills wrote:
> >
> > Here is the setup I will have: ServerIron, PII450 512MB/Ram, 6 10,000 RPM
> > 9 gig cheetahs, 3 2940 controllers (one for the system, three cache disks
> > on each of the other two controllers). FreeBSD, with noatime.
> >
> > I have two questions:
> > 1) Does anybody have an estimate of how many reqs/s I can serve well with
> > this setup?
>
> Your maximum request rate will depend mostly on how full you let the
> disks become. If you use 90% of the filesystems, then you'll probably
> be down in the 25/sec range. If you use much less, like 25%, then
> you can take probably 50/sec.

Based on my experience, your performance ought to be much more than
this. I have, according to most of what is said, a very low-performance
squid setup and I easily handle 25 req/sec on each of my boxes:

  FreeBSD 3.2 currently, 2.2.7 until recently
  Pentium II 266 / 384M RAM
  2 x 12.6MB Quantum Fireball ***IDE***

My demand peaks at about 25 req/sec, so I don't know the max, but at
that level squid spends only about 30% of it's time in disk wait.

I have two cache directories of 4GB each on each drive, 16GB total,
which I keep 91% full at all times. I use noatime and "tunefs -m 0"
and set the cache_dir size in squid to keep capacity at 91%. I use
100MB of storage memory.

> > 2) About how many reqs/s would I be looking at if I have 5 mb/s of general
> > ISP bandwidth?
>
> 5 Mb/s of HTTP traffic is about 50 requests per second.
> If you have other traffic in the 5 Mb/s, then its probably
> more like 30-40 req/sec.

My traffic may be atypical, but I get 50 req/sec on about 2.5 Mb/sec
total, of which about 60% is HTTP.

> > If I need to cluster, how many should I use? And how do I share the
> > content among the cache servers?

I have a cluster of two, and I avoid sharing content as much as
possible by directing unique requests to each server. I do this both
through transparent redirection and also an autoconfig file for some
clients. Currently, I am sending requests for servers with IP addresses
of 0.0.0.0-205.255.255.255 to one server and 206.0.0.0-223.255.255.255
to the other. This splits the HTTP load about in half.

I use ICP between them, although each proxy only hits about 2-3% of
it's misses on the other server, due to the servers mostly fielding
unique sets of requests. My LRU runs about 18 days with this setup.

Hope this helps - I guess it just points out that real results can
vary a lot from benchmarks and from one site to another.

Dave
Received on Mon Jun 14 1999 - 18:14:00 MDT