Re: [Fwd: [Fwd: [squid-users] Update: #redirectors decays overtime]]

On 22 Aug 2001 10:52:12 +0200, Mike Kiernan wrote:
> Robert,
>
> This leaves me with a few questions:
>
> - why is the default to use sockets rather than pipes
> for local ipc ?? [I come from solaris where normally
> you'd switch from sockets to doors on the local box]

No comment :]. It's been like this since I starting hacking on bits of
squid - that was nearly 18 months ago. There are comments in the code
about OSF/1 and poll() and FIFO's being bad bad bad. In fact IPC doesn't
default to sockets - redirector IPC defaulted to sockets. Other IPC
using bits of squid have other defaults, including diskd - which is
quite heavy on load :} that looks like
#if HAVE_POLL && defined(_SQUID_OSF_)
    /* pipes and poll() don't get along on DUNIX -DW */
        IPC_TCP_SOCKET,
#else
        IPC_FIFO,
#endif

> - is there a perf difference between the two ipc mechanisms
> on linux ?

I'll defer this to a linux guru. My *guess* would be that there will
always be more overhead with local tcp sockets - tcp checksums etc -
than pipes. How much overhead is a different question, as is whether the
extra overhead will cause problems like you saw.

> - why is redirectors, dnsservers etc. hardcoded so low?
I don't think they are _hardcoded_ so low :]. Some sites need more, some
need less. The default squid.conf is really a "well it works"
configuration, as opposed to a optimum configuration. You are hitting
performance limitations with squirm. Someone with a faster redirector -
say squidguard - will need less.
> - in these days of continuous uptime etc etc why does
> squid decide to exit in when the situation is far from
> beyond repair?
Continuous uptime? Whazzat? ;]. Seriously speaking the metric in the
helper heuristics for when to quit may be too low for you. However not
many folk have a pool size of 90 helpers, so 56 queued requests means
that they typically have at least one queued request and one in-progress
for every helper, which indicates a quite busy box. Maybe you should
increase the value tested against for your machine.

The idea of quitting is that things are at the point the client is
already suffering performance wise.
 
A max queue length config option might make sense.

> I don't mind patching source when needbe, but in the long
> term (ie. upgrading) it's all very time-consuming going
> through the patch/test cycle. I'd rather put effort into
> ensuring changes that make sense for everyone go into
> the base source release. Is there a facility for logging
> RFE's for squid ??

www.squid-cache.org has bugzilla on it. Log a feature request there.
 
Rob
Received on Wed Aug 22 2001 - 03:30:46 MDT