On 18/07/11 21:05, Francesco wrote:
Hello,
i have got a Squid version 3.1.8 running on a CentOS 5.x.
Since it works for url and content filtering, in conjunction with
Dansguardian in front, for some hunderd of users, the load average of that
machine is sometimes very high (also 5.0 or 8.0...).
Check your config for regex ACLs. They slow squid down. A lot. Simple
tuning such as dstdomain instead of domain regex can raise the req/sec
rates up easily.
The biggest process is squid, both on CPUs and Memory; i noticed, by
Squid-3 is not very efficient with CPU yet. We are still optimizing that
down after the new feature additions. It also uses RAM cache a lot more.
entering top -s and press "1" - that often the system is in wait, perhaps
it waits for the disk IO subsystem to write and manage Squid's cache.
Very likely. Check your disk I/O stats as well.
I tried to make some customization in cache tuning, thinking that the
bottleneck are disks, so i reduce the cache size.
Okay, confirmation that disk is one of the bottlenecks for you.
Something is getting better, but the system is still very loaded.
Are there some general suggestions to privilege speed and not bandwidth,
for systems with very high load?
* AUFS cache format on CentOS. This one is specific.
* JBOD configuration spread over more disks. Avoid RAID unless you have
high quality H/W controllers. Squid thrashes the disks with mostly
random-access writes and byte-hit-ratio proportion of random-access reads.
* Faster disks, you can get a good +50% speed out of 15Krpm over 10Krpm
according to the I/O benchmarks from last year.
* Possibly an SSD drive for small objects cache_dir (use min-size and
max-size to tune).
If this is still a big problem after tuning you may want to get in touch
with The Measurement Factory about a build of squid 3.2 with rock
storage. That is the new replacement for COSS and does SMP support as
well as disk backed slices.
Amos
--
Please be using
Current Stable Squid 2.7.STABLE9 or 3.1.14
Beta testers wanted for 3.2.0.9
