|
Although much attention has been devoted to increasing I/O bandwidth,
low latency has remained an elusive goal. We are building next
generation storage systems and file systems based on programmable
disks and fast switched networks. Our goal is achieving micro-second
latency.
By reading from or writing to a number of disks in parallel, a traditional RAID system delivers large bandwidth for large I/O's. This is analogous to a Single Instruction Multiple Data (SIMD) computing model where a number of different processing elements perform the same operation on a number of different data streams simultaneously in lock step. Although the SIMD model has large bandwidth potential, it does not help improve latency of small I/O operations. In contrast to the SIMD model, a Multiple Instruction Multiple Data (MIMD) computing model allows different processing elements to execute different instructions for different data streams. We propose to extend the traditional SIMD RAID to a MIMD RAID. In a MimdRAID, the different disk heads not only can operate in lock step to deliver large I/O bandwidth, they can also be decoupled when necessary to perform different duties. This extra degree of flexibility allows a MimdRAID to further take advantage of device intelligence to improve small I/O latency by exploiting the parallelism in the storage system. This research is supported by NSF Career Award: Low Latency I/O and Ubiquitous Storage (CCR-9984790). |