DATE: Thursday, November 30, 2006
TIME: 12:00 pm - 1:00 pm

Ahmed Amer
U. Pittsburgh

Prediction for Power and Cache

Data access prediction has been proposed as a mechanism to improve storage system performance, often through the prefetching of data. Effective implementations of prefetching caches are faced with the challenge of performing predictions in a timely manner, otherwise such schemes are either ineffective or detrimental. Fortunately, there are several effective approaches to effectively employ access prediction with storage, and in at least one case to reconcile traditionally contradicting performance goals.

In this talk I will discuss recent research where my students and myself have had success experimenting with applications of data access prediction to improve cache performance and reduce device energy consumption. One such algorithm reduced average response times by approximately 50% compared to a basic LRU cache, while requiring less than half the I/O operations that traditional predictive prefetchers would require to achieve similar hit rates. This is particularly significant when considering the energy savings that can be achieved by avoiding excessive and unnecessary device activity. Such a result considered disks that could enter low-power states when inactive, but we were also successful in finding a mechanism to reduce energy consumption in a disk without exploiting periods of inactivity, through the use of predictive grouping to improve data layout. In this particular application it is interesting to notice that increased power savings go hand-in-hand with reduced data access latencies.

Ahmed Amer received his Ph.D. from the University of California, Santa Cruz, where he worked primarily on predictive data management. His work has focused on the applications of access prediction to construct adaptive caching schemes, self-optimizing systems algorithms, and the analysis and enhancement of storage device power management. His most recent work has looked at the theoretical limits of predictive data grouping, and its impact on disk energy consumption and responsiveness.

For Further Seminar Info:
, or visit