On Thu, 2003-07-24 at 19:41, Ian Clarke wrote:
> On Thu, Jul 24, 2003 at 07:39:20PM -0400, Scott Young wrote:
> > There has been a lot of talk lately about algorithms to make routing
> > better. What about also using a better algorithm for deciding which
> > content to keep on a node, instead of the simple LRU cache? The new
> > algorithm would determine the key least likely to be requested again,
> > and that key is the next one that gets deleted. With this algorithm
> > being improved, data retention across the network should also improve.
>
> Good question, although it is hard to think of an improvement upon the
> current approach, which is to assume that the least-recently-accessed
> key is the key least likely to be requested again.
>
> Ian.
Analyzing how well the data-removal algorithm works would require a
little information stored about each deleted key (such as time of
deletion and the key). The parameter to be minimized is the probability
of a deleted key being requested again. Some potential algorithms for
this could correlate the routing algorithm's time-estimates with the
keys in the DataStore. For example, if the minimum time for requesting
a key is high, requests for that part of the keyspace might possibly be
less frequent, and therefore it might be better to delete from that part
of the keyspace. It would be up to the algorithm to learn such
correlations ('cause i'm too lazy to analyze the data myself and update
alchemy variables accordingly).
Another possible correlation might be in file size. High bandwidth
nodes would be good at transferring large files, but low bandwidth low
latency nodes would be like express lanes for small files. With the new
CHK specification, the file size could be used as a variable to learn
with in the routing and data deletion algorithms.
Scott Young
_______________________________________________
devl mailing list
[EMAIL PROTECTED]
http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/devl
