squid & cache routing

Hi, all,

As I have already reported at 3d Int. WWW cache Workshop and TF-cache
meeting this June, we are carrying out researches on cache routing.

I think that Squid (especially 1.2beta) icmp-pinging tohether with
netprobe database is good enough (and may be the best) for realization
of our ideas and for choosing closest parent with the shortest path
to destination server.
But we would like to have some additional features which are absent
(or realized not good enough) in the current beta version of squid
which I would like to discuss now.

1. Connectivity between neighbors in Russia is not so good as in the West
and is not very stable (RTT and even routing can vary in wide range from
time to time). I consider that squid should choose closest parent with
the best sum of RTT/hops to neighbor and from neighbor to destination
rather than only from neighbor to destination as realized in squid-1.2.beta23.
In addition, this way permits to exclude possible loops if several neighbors
have equal or close connectivity that can arise in current squid-1.2.beta.
Moreover, I think that the same approach should be used when choosing
neighbor having 'closest-only' option: if squid receives ICP_HIT reply
from a neighbor with 'closest-only' option it should send TCP GET request
only if RTT to neighbor is less than DIRECT to destination.

2. In a hierarchy of cache servers from different networks the situation
when all or many cache servers use cache-digest is possible. I think that
'no-cache-digest' option in cache_peer definition is needed, especially
to avoid possible false hits if neighbor does not permit miss_access.

3. Now, it is practically impossible to configure the same neighbor
as parent for one kind of requests and as sibling for another.
For example, I would like to have some neighbor as parent only for requests
from particular hosts and as sibling for requests from all others.
neighbor_type_domain seems do not work if cache_peer_access exists.
I think that to have separate cache_peer_access rules for parent and
sibling definition would be useful.

4. I don't sure, but by my feeling, 'closest-only' option (in squid-1.2.beta23)
seems to have higher priority than 'cache_peer_access' rules, and squid
sends request to neighbor with 'closest-only' even if this requests must
be denied by cache_peer_access rules.

I would like to know your opinion about features mentioned and
ready to be 'guinea-pig' if somebody has patches realizing them.

Sincerely, Serge Krashakov
Received on Tue Aug 18 1998 - 13:56:34 MDT