Klavs Klavsen wrote:
>
> I hope squid-2.5 is as fast as 2.2-stable5.. - I like the rproxy features.
> a lot.
Thanks.
> Do you guys have any performance pointers, as to how and what I need to
> tune.. both on the squid.. and on the linux-box?
You can find some information at the following pages:
http://devel.squid-cache.org/hno/linux-lfd.html
http://www.swelltech.com/pengies/joe/squidtuneup/t1.html
> I read something about changing tcp_rmem and tcp_wmem so they can open up
> bigger tcp-windows.. I did that..
> but the cache mainly serves files in size of 10 - 110K.. and as the normal
> window on a linux-2.4 has a max of 256kb that shouldn't be a problem?
On Linux there is generally no need to increase TCP windows..
> how many aio processes do I need to run on a busy box such as this? - i
> started with 50.. I tried with 350..
> can't seem to see if that helps at all..
Not too many. And as your disks was nearly not loaded at all the problem
is most likely elsewhere. There may be need to increase the amount of
AIO threads if you see a lot of blocked processes while the request hit
latency is still within reasonable limits.
> I guess the interrupts are for the network? - unfortunately I can't see
> from netstat -s - how many of the resends there are - are from that box..
Make sure you use a NIC & driver that has packet queueing or else there
will be at least one interrupt per packet sent/received which quickly
adds up...
Also as Adrian said, if using IDE, make sure to enable DMA. PIO mode
will kill your interrupts and CPU. Even if PIO is sometimes faster at
bulk transfers than DMA, it serverly impacts other concurrent activity.
> But I do know this.. - while squid was freezing - and the vmstat output
> that you saw ran.. I could easily connect via ssh - no problems or delays..
> (that's how I got the vmstat output :-)
Could you also make connetions OUT from the box?
Regards
Henrik
Received on Tue Nov 13 2001 - 01:04:05 MST
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 17:04:09 MST