Re: ICMP Feedback Request (Was: Re: Efficient public peer access control *without ACLs*)

Hmm. Lets see...

The good:

I used to have some "lag" for objects occasionally that weren't cached.
Thew lag only appeared to be a problem when accessing an object through a
parent. This was discussed here a few weeks/months ago and I don't think
anyone could come up with a definate reason for it. I think it had to do
with packet loss myself, but I'm not entirely sure. I tried the ICMP stuff
to see if it would help the problem (after I tried a bunch of other stuff
without success), and found that it DID! Woohoo! Every now and then there
is still some lag (probably from some of my parents that don't run the
ICMP stuff), but it is so small a problem now that I no longer worry about
it.

Of course, Physics tells us that every action will have an equal and
opposite reaction (and Murphy's law tells us that the reaction must be bad
:-), which brings us to the bad:

The ICMP stuff makes Squid like the caches that return ICMP info MUCH more
than those that do not, EVEN IF YOU WEIGHT THE NON-ICMP CACHES MUCH
HIGHER! Doh! If ANYONE knows a solution to this **PLEASE** tell me. To
give an idea of what I mean, I have the following lines in my squid.conf:

cache_host cache.merit.net parent 3128 3130 weight=128
cache_host cache.ann-arbor.cic.net parent 3128 3130 weight=100
cache_host pb.cache.nlanr.net parent 3128 3130 weight=32
cache_host uc.cache.nlanr.net parent 3128 3130 weight=32
cache_host_domain cache.merit.net !merit.net !merit.edu !umich.edu

Cachemgr reports that I fetch 13% from cache.merit.net, 4% from
cache.ann-arbor.cic.net (which has been down a lot and isn't a good number
right now), 29% from pb.cache.nlanr.net and 42% from uc.cache.nlanr.net.
Granted uc and pb (especially uc) have MUCH higher hit rates than the
other two, the merit cache is only 5 hops away from me, while the two
NLANR caches 11+ hops away. Ping times are 85 mS for merit, 97 mS for CIC,
108 mS for pb, and 107 mS for uc.

The Merit cache admin tells me that he will compile in the ICMP stuff when
he gets a chance, and the CIC cache admin never answers my email. The two
NLANR caches are currently the only parents I have that return ICMP info.
Overall I have seen a performance improvement with the ICMP stuff so I
have left it.

        -Bill

On Sun, 19 Oct 1997, Dave Zarzycki wrote:

> On 10/19/97 11:43 AM, Bill Wichers (billw@unix0.waveform.net) wrote:
>
> >On Fri, 17 Oct 1997, David J N Begley wrote:
> >
> >> Sounds like an interesting idea, but please remember that not everyone has
> >> or wants the ICMP pinging stuff compiled in (ie., as you're designing this
> >> don't make the assumption that everyone will be running it so don't break
> >> Squid or ICP to the point that it *has* to be run). Thanks.. :-)
> >
> >I just though some of the code for the ICMP pinging stuff would help
> >development. I would like to see the final option (if it comes to be :-)
> >as part of Squid that is seperate from the other compile time options. I
> >run the ICMP stuff myself since it solves a few lag problems I was having,
> >but it also messes up the weights for parents that don't run the ICMP
> >stuff. I guess I just can't have it all ;-)
>
>
> Could you and other squid users please send me some of your experiences
> of using ICMP? The good and bad please.
>
> Thanks,
> davez
>
> ----------------------------------------------------------------------
> Dave Zarzycki Student
> Workgroup Server QA Tester San Jose State University
> Apple Computer, Inc. zarzycki@ricochet.net
> zarzycki@apple.com dave@zarzycki.ml.org
> ----------------------------------------------------------------------
>
> PGP Fingerprints:
> DSS/Diffie-Hellman: CB9E 2621 B4BA 3F96 3516 B312 15B4 D842 3809 EF99
> RSA: 8AF2 1040 8A9C D025 47BE 70DD A51C C887
>
> Contact pgpkeys.mit.edu for my public keys.
>
Received on Mon Oct 20 1997 - 20:25:13 MDT