Re: [squid-users] 50 requests per second

Robin Stevens wrote:
> On Fri, Jun 21, 2002 at 01:08:00PM -0400, Robert Adkins wrote:
>
>>May I ask some specs on that one box that you are using? It would help
>>greatly to know the level of hardware required for such an install. It
>>might end up being more cost-effective and less of a headache if there is
>>system failure, to run two systems.
>
>
> While I can't speak for Joe's systems, we have servers capable of
> sustaining peak loads well in excess of 200 requests/second without
> noticeable loss of performance. In tests I've had them as high as 300/sec
> with significantly higher (but not unbearable) latency. But I've yet to be
> convinced that the current four servers will be enough to see us through the
> 2002/3 academic year if traffic continues to grow at the present rate...

The fastest machine I've shipped so far:

1.26GHz PIII
2GB RAM
2 x 15k RPM 36GB disks (approximately 36GB configured for cache_dir)

I benchmarked this box at 240 reqs/sec for four hours on a full cache,
and didn't have time to really push it beyond its limits. I wouldn't
expect it to sustain that in an ISP environment however, and would
recommend it for a significantly less demanding life (I'm comfortable
calling it a 180reqs/sec machine, with ability to handle peaks over
220--it lives in an ISP where it peaks at 160 reqs/sec, and usually
sustains around 120).

I've also shipped a dual PIII machine at 1GHz, with three 15k 18GB disks
and 2GB of RAM, but that requires dual processes to take full advantage
of the multiple CPUs--a nuisance, and the CPU affinity of current Linux
kernels is non-existent, so the CPU hopping is pretty costly. It is
still faster than the 1.26GHz machine, though, and lives at an ISP where
it has been seen doing 225 reqs/sec for about 3-4 hours sustained in the
evenings. I call it 'full'...and if their needs increase much we'll
have to cluster.

We've recently gotten word from the manufacturer that we can begin
shipping P4 and Xeon systems, so I'm looking forward to trying out a
2.4GHz machine, or a Xeon 1.7GHz with up to 12GB of RAM and 4 disks all
in a 1U chassis. ;-)

> Our hardware is based around Dell Poweredge servers: single PIII CPU, 1.5
> or 2GB RAM, 7x 10000 or 15000 rpm cache drives giving about 100GB of cached
> data per server. Software is based around Redhat Linux with 2.4.x kernel
> and reiserfs cache partitions (mounted noatime,notail) and squid 2.4Stable6.
> I'll probably be investigating 2.5 over the summer while we've got the spare
> capacity for me to perform tests.

These specs are fine. I wouldn't change a thing (except less cache_dir
and more RAM). ;-)

> Joe always seems to be recommending considerably more RAM for that amount
> of cache disk, but I guess this depends on the size of the average stored
> object. In our case this is around 20k - we get a lot of downloads into
> tens or even hundreds of megabytes.

My RAM recommendation assumes pretty strict disk I/O throughput
limitations (the 2-3 disks that we can fit into our 1U chassis don't
provide as much I/O bandwidth as is really needed to max out the CPU
once you get up to these speeds). You get around that limitation mostly
by providing 7 disks, which is a honking lot of throughput, and very
good disk availability. Another reason is that latency reduction is the
prime motivation for caching in most of my clients environments. ISPs
in the US are not strapped for bandwidth in most cases, but they do want
a 'leg up' on the competition in their area, by providing a 'snappier'
browsing experience (and if it saves them a few hundred bucks a month on
bandwidth, that's a nice bonus).

As I've always said, a safe number is 10MB of RAM for every GB of
cache_dir. In your case 100GB of disk is 1GB of RAM. You then double
it, and go zoom. You're fine, and not too far off from where I spec our
low end boxes, where performance isn't the primary concern. Though I'm
still only shipping 36GB or maybe 40GB of cache storage in a 2GB machine
(but those boxes are wicked fast). It would be safe to go higher, but
not so good for performance. If I ever have a client interested in more
disks, we can get a nice 2U chassis that will support 6 disks, so maybe
I've get to give one of those a go sometime with a 100GB of cache_dir.

It's all about balance. And RAM being dirt cheap these days.

-- 
Joe Cooper <joe@swelltech.com>
Web caching appliances and support.
http://www.swelltech.com