[squid-users] performances ... again

Hi there,

I've been looking around for performance issues and possible enhancement
of generic squid configuration.
We (in our company) are running a linux box with Squid (2.7Stable1, I
try to stay up-to-date).
The box is a PIII-1GHz with 512Mb of RAM and 3x18Gb hard-drives.
Web access is provided by a 10Mb dsl line (download speed. Upload speed
is 768Kb)

The box handles about 250 clients.
Monitoring shows around 1200/1600 HTTP requests (5min average) and 400
HTTP hits (25% hits)
DSL line is used between 1.5 and 2 Mb/s.
 
Beside the fact that the hit rate is low, response time are way too long
for users (cache-misses median service times are around 200ms and
cache-hits are around 3ms)

I made changes to the cache_dir :
- lowered it from 10240Mb to 1024;
- moved from aufs to diskd
- moved from diskd to coss
- added a second cache disk :
-- run both with aufs
-- run both with diskd
-- run one with diskd and the second with coss

I tried the null storage too, in order to check if disk I/O is involved
in this situation.

Then I tried to move squid to a "newer" box with two 2.4GHz CPUs, 2Gb of
RAM and 5 300Gb SCSI disks.
The configuration run with 1 system disk and 4 cache disks with diskd.
(1024Mb of cache)

Tried the null storage again ...

Responses times are not better, in any case.

Well, in fact, I got a *tiny* improvement when I installed CACHEBOY 1.0
(sorry Henrik, but I had to give it a try ...)

As it does not seem to be a squid issue, I thought of modifying the
kernel, particulary the NR_OPEN and NR_FILE, updating these values from
the original 1024 to 8192.
Are these the right values to modify ?
Should I dig somewhere else in the system ?

Many thanks for your help,

Ionel
Received on Thu Jun 05 2008 - 16:12:11 MDT