Squid_ The Definitive Guide - Duane Wessels [99]
cache_peer neighbor.host.name parent 3128 3130
To switch over to HTCP, the line becomes:
cache_peer neighbor.host.name parent 3128 4827 htcp
Sometimes people forget to change the port number, and they end up sending HTCP messages to the ICP port. When this happens, Squid writes warnings to cache.log:
2003/09/29 02:28:55| WARNING: Unused ICP version 23 received from 64.216.111.20:4827
Squid doesn't currently log HTCP queries as it does for ICP queries. HTCP queries aren't tracked in the client_list page either. However, when you enable HTCP for a peer, the cache manager server_list page (see Section 14.2.1.50) shows the count and percentage of HTCP replies that were hits and misses:
Histogram of PINGS ACKED:
Misses 5085 98%
Hits 92 2%
Note that none of the current Squid versions support HTCP authentication yet.
Cache Array Routing Protocol
CARP is an algorithm that partitions URI-space among a group of caching proxies. In other words, each URI is assigned to one of the caches. CARP maximizes hit ratios and minimizes duplication of objects among the group of caches. The protocol consists of two major components: a Routing Function and a Proxy Array Membership Table. Unlike ICP, HTCP, and Cache Digests, CARP can't predict whether a particular request will be a cache hit. Thus, you can't use CARP with siblings—only parents.
The basic idea behind CARP is that you have a group, or array, of parent caches to handle all the load from users or child caches. A cache array is one way to handle ever-increasing loads. You can add more array members whenever you need more capacity. CARP is a deterministic algorithm. That is, the same request always goes to the same array member (as long as the array size doesn't change). Unlike ICP and HTCP, CARP doesn't use query messages.
Another interesting thing about CARP is that you have the choice to deploy it in a number of different places. For example, one approach is to make all user-agents execute the CARP algorithm. You could probably accomplish this with a Proxy Auto-Configuration (PAC) function, written in JavaScript (see Appendix F). However, you're likely to have certain web agents on your network that don't implement or support PAC files. Another option is to use a two-level cache hierarchy. The lower level (child caches) accept requests from all user-agents, and they execute the CARP algorithm to select the parent cache for each request. However, unless your network is very large, many caches can be more of a burden than a benefit. Finally, you can also implement CARP within the array itself. That is, user-agents connect to a random member of the cache array, but each member forwards cache misses to another member of the array based on the CARP algorithm.
CARP was designed to be better than a simple hashing algorithm, which typically works by applying a hash function, such as MD5, to URIs. The algorithm then calculates the modulus for the number of array members. It might be as simple as this pseudocode:
N = MD5(URI) % num_caches;
next_hop = Caches[N];
This technique uniformly spreads the URIs among all the caches. It also provides a consistent mapping (maximizing cache hits), as long as the number of caches remains constant. When caches are added or removed, however, this algorithm changes the mapping for most of the URIs.
CARP's Routing Function improves on this technique in two ways. First, it allows for unequal sharing of the load. For example, you can configure one parent to receive twice as many requests as another. Second, adding or removing array members minimizes the fraction of URIs that get reassigned.
The downside to CARP is that it is relatively CPU-intensive. For each request, Squid calculates a "score" for each parent. The request is sent to the parent cache with the highest score. The complexity of the algorithm is proportional to the number