Skip to Main Content
Emerging high speed, scalable networks allow machines to access remote memory faster than accessing local disks. Cooperative caching exploits this phenomenon and uses the I/O caches of remote workstations as an extra level of memory hierarchy. Most research has focused on improving the read performance of cooperative caching, but writes to disk are becoming a performance bottleneck. We propose a write-back mechanism that uses replication and logging to improve reliability as well as performance of a distributed system, based on a cache architecture called "Distributed, Efficient and Reliable (DEFER) Cooperative Cache". We present three strategies based on this architecture, and evaluate their performance using simulations. The simulations indicate a maximum speedup of 11.5x over the write-through baseline. The vast performance gains are a result of avoiding disk writes through the write-back policy.
Date of Conference: 22-26 April 2003