Re: Performance question

Hi all,

I do agree with you that value of q and q0 is high.
Actually I started with lower side 8192 and then
increase to see any gain and there is none.

But Squid is unable to handle more than 75/req for
more than an hour.
Polygraph is running on FreeBSD 3.0 with all suggested
tuning parametrs ( Thanks to Alex for that)
Polygraph is seems to be fine and can take load of
500/sec more than an hour.

I have increased the cache_mem from 256MB to 800MB but
squid at no time use more than 300MB( from top).

It is able to handle around 100 - 120req/sec just for
10 minutes not more than that.

Thanks ,

Gyan Prakash
Squid User...
 

> On Tue, 15 Jun 1999, Dirk Moerenhout wrote:
>
> ]First of all your settings seem to be way over the
> top. There is no reason
> ]to set tcp_conn_req_max_q and tcp_conn_req_max_q0
> to such high values.
>
> Unless you are fearing SYN floods, where SUN
> suggests to set q0 to about
> 10k. For benchmarkings as stated below, increasing
> the q is valid, too,
> though not *that* high.
>
> ]could ever have 16000 connections waiting, if you
> have you're screwed
> ]anyway. For the FD's you're way over the top too,
> if the thing has need
> ]for over 4000 FD's there's something wrong.
>
> I agree to this. Even while benchmarking there was
> no need for me for more
> than 4096 FDs. 'Upgrading to Linux' is not always an
> option. Setting FDs
> too high is said to have a negative impact on
> performance, too, e.g. 4096
> should suffice.
>
> ]> FD is set to 12288.
>
> There are a few more toggles and dials to fiddle
> with, especially when
> benchmarking (so what systems (plural!) did you use
> to run your polygraph
> suite from), but most won't do much difference.
> However, if you stuck to
> the "how to polygraph" step-by-step as outlined by
> Alex,
>
> http://www.terena.nl/tech/d2-workshop/d2cache99/
> (at the very bottom)
>
> you should have discovered during level 0 tests that
> you had to tune your
> TCP buffers to 64k, otherwise you must have noticed
> that you weren't able
> to achieve the netperf speed seen with FreeBSD.
> Until your level 0 tests
> work out, all higher level results are bogus.
>

===

_________________________________________________________
Do You Yahoo!?
Get your free @yahoo.com address at http://mail.yahoo.com
Received on Tue Jun 15 1999 - 09:44:22 MDT