0 Replies Latest reply on Mar 14, 2005 3:06 PM by mranga

    JBoss Cache in Scientific Cluster computing.

    mranga

      JBoss Cache currently uses eager consistency which means - when the TX ends, updates are propagated. This is necessary for fault tolerance which is after all the main reason for its existence.

      However, consider the needs of a compute cluster with coarse grained parallelism (and not fault tolerance) being the objective. For this, it would be interesting if a lazy model were followed where, a lock is aquired at the start of the tx - resulting in network traffic. When the tx commits that generates no network traffic. Thus is you are iterating on a node, on a shared variable, and the other nodes sharing the variable are much slower, then no network traffic is generated.

      Such things have been investigated in the past ( for example, consider Peter Keleher's work on Lazy Release consistency). It would be interesting to move the support into jboss cache given the current interest in grid computing .

      Just some thoughts which I thought I could share.


      Ranga.