Re: [mark@reynolds.net.au: Re: [WAIX] squid 2.4 ?]

On Mon, 23 Jul 2001, Adrian Chadd wrote:

> Can someone cache-digest-clued see if what Mark is seeing
> is what happens?

Yes and no. Cache digest peer selection scans through all peers and
selects the best HIT peer based on round-trip time measurement (if
available). If no RTT measurements available, Squid selects the "next"
HIT peer since last scan. This "next" selection code is a bit iffy
(based on a global first_ping variable used by several peer selection
algorithms) and might attribute to uneven selection. Not sure. See
neighbors.c::neighborsDigestSelect().

Clearly neighbor selection code can be improved to make peer selection
algorithms more independent from each other and "cascade-able" so, for
example, one can configure Squid to use round robin _after_ cache
digests have selected HIT peers. IIRC, Henrik had some grand ideas
about improving this code.

Enabling level-5 debugging of section 15 will trace
neighborsDigestSelect() actions.

Alex.

> ----- Forwarded message from Mark Reynolds <mark@reynolds.net.au> -----
>
> X-Mailer: QUALCOMM Windows Eudora Pro Version 3.0.5 (32)
> Date: Mon, 23 Jul 2001 16:18:51
> To: Adrian Chadd <adrian@creative.net.au>
> From: Mark Reynolds <mark@reynolds.net.au>
> Subject: Re: [WAIX] squid 2.4 ?
>
>
> Hi Adrian,
>
> Would it be true to say, that when using cache digests
> only, and having say a dozen peers, that the top peers
> get more 'action' than the lower peers, because the
> first digest indicating a hit, gets the request?
>
> This is not very good, as it means that the main load
> is dumped onto the peers at the top of the peer list,
> and they get overloaded more than if the load was
> spread evenly across all peers with the objects.
>
> Is this true? Can it be fixed to be even handed?
>
> mark.
>
>
> At 11:42 28/03/2001 +0000, you wrote:
> >On Wed, Mar 28, 2001, Mark Reynolds wrote:
> >>
> >> Anyone done the
> >> 'squid 2.4 on production proxy server'
> >> thing yet?
> >>
> >> Looks like it has gone to the stable tree now.
> >>
> >> http://www.squid-cache.org/Versions/v2/2.4/
> >
> >There are a couple of little bugs that have been fixed
> >since stable1. I'll sit down with Henrik/Duane rsn to get
> >a set of patches. You can try the latest squid-2.4
> >snapshot, which has the little fixes in it.
> >
> >Oh, if you use FreeBSD - the squid24 port is currently
> >broken for diskd. Thats on my todo list as well. :-)
> >
> >Other than the above, I've had quite a lot of positive
> >feedback from squid-2.4 users. :-)
> >
> >
> >
> >
> >Adrian
> >
> >
> >--
> >Adrian Chadd "The fact you can download a 100 megabyte file
> ><adrian@creative.net.au> from half way around the world should be viewed
> > as an accident and not a right."
> > -- Adrian Chadd and Bill Fumerola
> >
> >--
> >To leave, email "unsubscribe" to nap-request@waia.asn.au
> >
> >
>
> ---
>
> Mark Reynolds
>
> Managing Director Reynolds Technology Pty Ltd
> Phone 1300 656 424 http://www.reynolds.net.au
> Phone 08 9474 1211 mailto:mark@reynolds.net.au
> Fax 08 9474 9592 PO Box 945 South Perth 6951 WA
> Pager 08 9480 5884 19 Lyall St South Perth 6151 WA
> ABN 73 078 831 740 ACN 078 831 740
>
>
>
> ----- End forwarded message -----
>
>
Received on Mon Jul 23 2001 - 08:02:20 MDT