> After that point, cache probabilistically according to the node's > perception of its key distribution (a crude measure would be the > 256-buckets histogram, a more sophisticated one would be the ngrouting > estimator that ian and mjr have been working on).
As I said on IRC, I think it is dangerous to let nodes decide what their specialization is, and try to force further specialization on that basis - I propose we call this approach "introspective caching" - or "icaching". To illustrate why, consider a situation where - by chance - every node in the network decided to specialize in more-or-less the same area of keyspace. With your "forced specialization" proposal it would be difficult or impossible for any of those nodes to change their specialization. In the current mechanism (with or without pcaching), node's will specialize in areas where there is a "market" for expertise in that area. This should result - assuming that requested keys are randomly distributed (as they will be) - in good specialization coverage of the entire keyspace by nodes in the network. To further illustrate this point, pcaching could be seen like a market where individual sellers make sure that they stock the products for which they see a demand. This will naturally result in different sellers specializing in different product types in a manner which largely meets the demand of the market's customers. Icaching would be more like a market where sellers looked at what they have in-stock, and only purchased new items that were similar to what they already have, more-or-less regardless of which items they were actually seeing a demand for. In this analogy, there is a serious risk that the market as a whole will slowly degrade in its ability to meet consumer demand. Ian. -- Ian Clarke [EMAIL PROTECTED] Coordinator, The Freenet Project http://freenetproject.org/ Founder, Locutus http://locut.us/ Personal Homepage http://locut.us/ian/
pgp00000.pgp
Description: PGP signature
