Performance Tuning

I have a couple of performance tuning questions, and would like to get any
opinions that would help me to make this cache as fast and efficient as
possible.

I am running Redhat 6.1 on a Dell Poweredge 4400 server with a 667Mhz P3X,
512MB (133MHz SDRAM), and 5 x 9gb 10k rpm U160 drives connected to a raid
controller with 64 meg ram and configured for write caching. 1 GB of each
drive is striped with raid 5 for the operating system. The rest of each
drive is set up as a standalone volume to create an OS volume of 4GB and 5
x 8GB cache volumes.

I am running squid 2.2s5-hno with delay-pools, memory pools, async-io (32
threads), and dl-malloc. There are 5 cache_dirs, each with 6200kb, 16/250
directories. Cache_mem is 64MB. During business hours, the cache is
averaging about 700 concurrent open files (configured for 16000 maximum).
During peak times I've seen the request rate go up to 65/sec (6900/min).

This machine only runs squid and named. The resolve.conf is configured to
split the named load between this machine and another machine. The named
on this machine is reserved for squid use.

In the immediate future, we will be installing an additional 512MB of
memory because we have recently had problems with running out of memory.
Delay pools will also be turned off once our internal and external
networks are upgraded (hopefully soon). We are also expecting our internet
usage to double within the next year.

The first question that I have deals with the number of async-io threads.
I have noticed that under moderate load squid will report that all of the
async-io threads are in use. I have also noticed that the hard drives are
very active during this time. I use xosview to keep an eye on this server
and the total disk io during these times is usually around 300kb/sec. Also
from watching the activity lights, the drives are not as active as they
could be. Would increasing the number of async-io threads help in this
case? How much memory does each async-io thread use (ps always shows me
the total for squid)?

The other question that I have deals with CPU usage. I've noticed that the
CPU usage will peak up to about 50%, with about 15% due to squid, and the
rest due to system time. Is there anything that I can do to drop the
percentage of system time? Would a second processor allow more room for
growth?

Any suggestions on how I can get the most from this system would be
greatly appreciated.

Thank You,

Erik
Received on Thu May 25 2000 - 10:47:07 MDT