Reducing Disk Latency through Replication
Gordon B. Bell and Morris Marden
Abstract: Today's disks are inexpensive and have a large amount of capacity. As a result, most disks have a significant amount of excess capacity. At the same time, the performance gap between disks and processors has widened to the point that many workloads have become disk bound. To improve the performance of disks, we propose using the excess capacity of disks to replicate blocks. To do this, the disk controller observes sequences of requests to blocks and replicates blocks on disk so that they are in the same order on disk as in the sequences. By doing this, when the sequence occurs again, no seeks are needed between accesses to blocks in the sequence. Our work shows that these sequences can be reused a large number of times, so they potentially can yield a large benefit. We also have an algorithm, which we implemented in the DiskSim simulator, for detecting these sequences and performing replication.
Available as:
Postscript
or PDF
Disk Traces:
Limit study program
Modified DiskSim simulator (~6 MB)
Links: