Re: [squid-users] Re: Re: Performance Question

Thanks for the reply.

> Anything under ~ 1 seconds is probably fine for misses, and even up to 2
> seconds depending on congestion and latency on your link.
>
Seems my squid setup is not that bad after with a miss time of 250ms :-)
I get those numbers only during the day when the cache is fairly loaded
with enogh "fast" misses.
During the night response times are sometimes much higher as one "bad"
request has much more influence at the average.

>
> Look at the refresh_pattern setting in your squid.conf. You will probably
> get the most benefit by tuning settings for image files, which tend to be
> large, downloaded often, and seldom changed.
>
I will.
>
> The ufs store method is probably a major (if not the only) bottleneck here.
> Try to solve the problems you were having with diskd - switching to it will
> give you a performance boost over ufs.
>
> diskd has its own FAQ section with configuration tips - maybe it will help:
>
> http://www.squid-cache.org/Doc/FAQ/FAQ-22.html

The o'reilly book helps a lot as well. Seems like I didn't configure the
queues properly.

>
>>pppoe on openBSD is not the fastest of his kind, probably there is room
>>for improvement. I may try to switch over to debian/linux on this
>>machine as well.
>
>
> If you switch to Linux, use aufs instead of diskd (it performs better).

The child proxy is running under Linux. That is the one where the users
connect to. I will try aufs first before having another try with diskd.

Only the parent (the box with connects to the adsl line) is using
openbsd. This machine is using a memory cache only, the disks are only
used for logging. IMHO it is useless to have a big cache_dir here, as
everything is cached again on the child proxy.

>
> How do you make that switch when the parent goes down?
>

the bsd-box is the default-parent with no-delay for the linux box.
with "always_direct deny all" all request go the parent at full speed,
but only as long as the parent is up. If the parent is dead, the child
is free to fetch them direct (via the default gateway) through another
but smaller line. This line is used for other purposes as well and must
not be used more than 50%. This is where the delay_pools kick in.
I would need different delay_pool levels for parent and direct connects,
but this is impossible, that is why I can not use delay_pools with the
parent at all.

This setup may seem a bit odd, but think of the ADSL-Line & the
openbsd-box as a fast & cheap but probably unreliable bypass for the
squid traffic.

By the way: I may have found a way to do some benchmarking with the
whole system (both squid boxes, firewalls & the ADSL-line). This will be
the definite test :-)

Regards, Hendrik
Received on Fri May 21 2004 - 10:57:14 MDT