CS262B Reading Summary

Pond: the OceanStore Prototype

Sean Rhea et al.

Summary by Feng Zhou
4/7/04

Strong points of the paper are:

  1. The motivation for OceanStore is to provide "truly durable, self-maintaining storage". This is a very useful service and only made possible by latest technology advances like universal Internet connectivity and cheap storage and computing hardware.
  2. The consistence model of OceanStore is unusual. All versions of each object are archived by the system. This enables time travel and simplifies object update. Because no in place modification is done, reads are always repeatable. The update process looks very like Bayou.
  3. The division of primary and secondary replicas is important. Primary replicas are responsible for the availability/correctness of data, while secondary replicas exist for performance. To ensure fault-tolerance, the inner ring of each object that maintains the primary replica uses a Byzantine fault-tolerant protocol to replicate there state. This makes sense considering the not fully-trusted environment in which the system works.
One major flaw.

What machines will form the OceanStore service is still not clear. From one perspective it looks like NOW for storage. And from the NOW experience, using idle workstation to provide computing seems not a very good idea because of manageability and software complexity problems. The same may also apply for storage. So the key question is still: why is better than a centralized solution, just like Google.