Memory-Contention Responsive Hash Joins
Diane L. Davison, Goetz Graefe · 1994
davison @ cs.colorado.edu In order to maximize system performance in envi-ronments with fluctuating memory contention, memory-intensive algorithms such as hash join must gracefully adapt to variations in available memory. Mixed workloads, creating fluctuations of erratic frequency and magnitude, make respon-siveness to memory contention particularly impor-tant. Previous studies on adaptable hash joins have focused on lowering I/O costs by reducing the I/O volume, as measured in the number of pages, by spilling partitions from memory to disk and then restoring them into memory if more memory becomes available. In this paper, we pre-sent memory-contention responsive hash joins that (i) reduce the amount of time spent on I/O by us-ing large I/O buffers, or clusters, (ii) dynamically vary the cluster size in response to fluctuations in memory availability, and (iii) employ earlier tech-niques of dynamic destaging and restoration. Our simulation results demonstrate that these com-bined techniques provide better performance than previous algorithms, particularly in environments with medium to high memory contention or with very frequent changes in memory availability. 1.