Collaborative Object Caching for Heterogeneous OSD Clusters

With the emerging of 10 Gbps Ethernet, modern storage networks allow the hosts to access remote cache nearly as quickly as they can access from their local storages. Aiming at improving the performance and load-balancing of large-scaled object-based storage systems, we propose an objectbased caching scheme that works collaboratively by distributing redundant objects among neighboring object-based storage devices (OSDs). Our objective is to maintain valuable objects in the caches (either local or remote) to reduce the disk I/O bandwidth demand, and minimize the cache response time. Trace-driven simulations show that the cache hit ratio of our object-based caching scheme is superior to the general block-based cache replacement algorithms by a factor of more than 2.6 and the cache response time can be improved by about 25 folds.