Re: What's the best configuration for this setup?

From: Oskar Pearson <>
Date: Fri, 22 Nov 1996 15:32:34 +0200 (GMT)

Daniel O'Callaghan wrote:

> On Fri, 22 Nov 1996, Ross Wheeler wrote:

> > I noticed a particularly unpleasant problem I'd appreciate ANY input on,
> > and that is that when someone requests a page through our proxy, and it
> > is NOT cached, it checks the "parent". IF upstream has it, (and here's
> > the catch), it comes roaring down the 128K and consumes most of our
> > bandwidth while it does so. If it's a large file (big web page, or worse,
> > a large file that's being FTPd), this can go on for quite some time.
> This was discussed on freebsd-hackers not so long ago. It is possible to
> bandwidth limit connections by a simple

> while () {
> read();
> sleep();
> }

> If course, sleep(), or usleep() are not appropriate for squid, but how
> hard would it be to implement an algorithm which says "OK, I'm supposed
> to read max 3000 chars/sec; I just read 1500 chars; I must not read again on
> this connection for at lease 500 ms."??
Or you could set a counter on each filehandle, decrease the number on the
filehandle each time that you were "supposed" to transmit the data to them,
and store data in a buffer until the counter=0 When it reaches 0, put X bytes
of data in an outgoing queue and set counter=10 (for eg)

Of course, if you have a T3 to the net and someone downloads Netscape via modem
through the cache and is bandwidth limited, you have to hold the whole of
netscape in those buffers (yes, the whole thing is anyway, but on a really big
cache you are going to have to swap your current position in memory so fast
that you are going to get latency from changing pages... Yes this is valid, has this problem)

I think there are much more important things to be done. (I don't know if
some of them have been done in 1.1bX - I haven't been watching)

If (for eg) has 9 IP's, try connect to each one if you
        get something like a "cannot connect to remote host" message.
Respect the TTL's that are returned with DNS responses
?Objects living forever in neighbor caches? (Must have been worked out?)
Documentation (If someone has some kind of base framework, or chapter
list that is reasonable, I am volunteering...)

> My guess is that it would need a bandwidth and an "earliest time for next
> read" to be stored for each (outbound?) tcp connection.
> One could also limit write()s so that delivery is throttled.

It wouldn't work if the slow person has the middle cache set as a perent
(unless I am a fool ;) hey, no comments in the corner there!

He is fed by ISP-A, he is ISP-B, he feeds ISP-C

presumably he wants to feed ISP-C as fast as they can get, so he can't
put his patch on himself. If he gets ISP-A to bandwidth limit, they are
going to limit him too, as he is a PARENT to ISP-C.

ISP-C requests doc from ISP-B
ISP-B requests doc from ISP-A
ISP-A limits ISP-B? WRONG!

So ISP-C would have to have be set up with ISP-B as a neighbor and ISP-A
as a parent. Squid on ISP-A would have to then bandwidth limit ISP-C directly.


'experience made art, but inexperience luck.'
Received on Fri Nov 22 1996 - 05:33:01 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:33:36 MST