Topology-Aware I/O Caching for Shared Storage Systems

The main contribution of this paper is a topologyaware storage caching scheme for parallel architectures. In a parallel system with multiple storage caches, these caches form a shared cache space, and effective management of this space is a critical issue. Of particular interest is data migration (i.e., moving data from one storage cache to another at runtime), which may help reduce the distance between a data block and its customers. As the data access and sharing patterns change during execution, we can migrate data in the shared cache space to reduce access latencies. The proposed storage caching approach, which is based on the two-dimensional post-office placement model, takes advantage of the variances across the access latencies of the different storage caches (from a given node’s perspective), by selecting the most appropriate location (cache) to place a data block shared by multiple nodes. This paper also presents experimental results from our implementation of this data migration-based scheme. The results reveal that the improvements brought by our proposed scheme in average hit latency, average miss rate, and average data access latency are 29.1%, 7.0% and 32.7%, respectively, over an alternative storage caching scheme.