Tuning Linux IPv4 route cache (2011)

The dst/route cache has effectively been replaced by the TCP metrics cache. Values like MSS, SRTT, and cwnd are still useful to preserve across streams between peers. See ip-tcp_metrics(8) for a decent summary.

David Miller has a summary of the route cache removal in http://vger.kernel.org/~davem/columbia2012.pdf. The cache exhaustion argument was .. odd. As I recall the cache entry was created on handshake and was keyed off of src ip, dst IP, dst port. Given that the attacker can (effectively) only vary the src ip thats a lot of addresses to control. The other argument was that not hitting a cache was "only" 10% slower in microbenchmarks.

Personally Im dubious of the route cache removal, though I havent tested. The 10% hit rate, as I recall, was from Google front end servers which Id expect to have a high cardinality of peers along with a simplistic route table. For servers with multiple interfaces, variable route properties, multiple tables, etc and a high packet rate a fast cache should be quite valuable.

Nothing replaces it. Lookup performance of the routing table has been increased by using an LC trie since 2.6.39 and the various information which were previously in the cache have been moved to other structures.

The performance of the cache was said to be bad (only 10% hit, but that doesn't match at all my experience where the performances of the cache were pretty good, see the article) and being an hash table, it was the target of DOS attacks. Moreover, many information were tied to the cache, triggering some additional problems, like the "neighbor cache overflow".

This post is 2 years old.