Re: icp_no_misses

martin@mrrl.lut.ac.uk writes:

>I was thinking that for clusters of closely coupled caches dealing
>with a very high volume of requests it might make more sense to have a
>way of turning off the returning of ICP misses to nominated servers.
>
>This could put the requesting server in a sticky position, so I think
>it would need to know that a particular peer wasn't going to be
>returning ICP misses - say via an option in its cache_host directive.
>It probably would also be necessary to be able to specify a
>neighbor_timeout in terms of milliseconds rather than seconds, but I
>don't know whether this is feasible.
>
>I'm wondering if there are any obvious flaws in this as an idea -
>overlooking the actual implementation detail for now. What do people
>think?

I think the only possible problem would be with the timeouts.

For Squid, the timeout handlers are now only checked once per select()
or poll(). However, the ICP socket is checked much more frequently;
for most versions its checked between every other filedescriptor we are
selecting on. Squid-1.2 tries to be smarter than that....

It could probably work by moving the FD timeouts to the event linked
list, giving the event list sub-second resolution and checking it much
more frequently than we do currently.

Duane W.
Received on Tue Jul 29 2003 - 13:15:44 MDT