Prefetching Data From Disk Into Main Memory

Student: Vivek Vellanki

Prefetching Without Hints: A Cost-Benefit Analysis for Predicted Accesses
Vivekanand Vellanki and Ann L. Chervenak, submitted to OSDI'99

To address the gap between memory and disk access times, we will prefetch data from the Personal Terabyte magnetic disk into main memory.

This type of prefetching will be essential to achieving high performance on the Personal Terabyte. Over time, CPU speeds are increasing over 50% per year, while disk access times are improving less than 10% per year. To avoid an I/O bottleneck, we must prefetch disk data into memory so that it is available to the processor without incurring a page fault. Today, a page fault stalls the CPU for millions of cycles; as clock speeds increase, this penalty will increase to tens of millions of cycles.

We are performing a prefetching study that predicts future accesses based on past access patterns and then performs a cost-benefit analysis to determine whether the benefit of prefetching a block exceeds its cost.