RE: Testing Squid with Polygraph

From: Duane Wessels (wessels@measurement-factory.com)
Date: Wed Oct 03 2001 - 13:19:58 MDT


On Wed, 3 Oct 2001, Ahuja Sadhna (NRC/Boston) wrote:

> Hi Duane,
> Thanks for your prompt reply. Here are answers to the questions
> you asked:
>
> > How many disk drives in your Squid? What kind?
>
> Just one drive - an IBM ATA drive, 9GB.

50 req/sec is probably close to the limit for just one
disk drive.

> >
> > Did you try a proxy-only (all replies uncachable) workload?
>
> Yes, there are no errors in this case.

and you got better throughput?

>
> >
> > Did you recompile Squid after increasing kern.maxfiles? Does
> > Squid see the new filedescriptor limit?
>
> I hadn't done this earlier, but now I have. Squid sees the new limit as
> its SQUID_MAXFD value. With this change, after running the tests for
> some time, I see the following message continuously on Squid:
> comm_open: socket_failure: (55) No buffer space available
> But, the Polygraph tests continue to run normally (or at least they look
> normal to me).
>
> >
> > Did Squid run out of network mbufs? What does 'netstat -m' say?
> > Did you increase NMBCLUSTERS on the Squid box?
>
> I didn't increase NMBCLUSTERS. Here is the output of netstat -m, while
> Squid was complaining about no buffer space:
>
> 1158/1968/4096 mbufs in use (current/peak/max)
> 115 mbufs allocated to data
> 1043 mbufs allocated to packet headers
> 98/930/1024 mbuf clusters in use (current/peak/max)
> 2352 bytes alloacted to network (76% of mb_map in use)
> 0 requests for memory denied
> 0 requests for memory delayed
> 0 calls to protocol drain routines

You should increase NMBCLUSTERS in the kernel. Maybe you can
do it with sysctl.



This archive was generated by hypermail 2b29 : Mon Feb 06 2006 - 12:00:20 MST