Abstract

The elapsed time for external mergesort is nor-mally dominated by I/O time. This paper is focused on reducing I/O time during the merge phase. Three new buffering and read-ahead strategies are proposed, called equal buffering, extended forecasting and clustering. They exploit the fact that virtually all mod-ern disks perform caching and sequential read-ahead. The latter two also collect information during run formation (the last key of each run block) which is then used to preplan read-ing. For random input data, extended fore-casting and clustering were found to reduce merge time by 30 % compared with traditional double buffering. Clustering exploits any tem-poral skew in input runs to further reduce the number of seeks.

Citations