This paper introduce a mapping strategy for RAID as shown in above figure, based on the following observations:
- Data usage is highly skewed with a small percentage of blocks being heavily accessed.
- Working-sets remain stable over long durations.
This paper makes the following contributions: we prove that using a large cache-like partition that uses all storage devices can be better than using dedicated devices due to the improved parallelism, in some cases even when the dedicated devices are faster. Additionally, we demonstrate that information about hot data can be used to reduce the overhead of rebalancing a storage system.