Disk-Locality in Datacenter Computing Considered Irrelevant
amplab.cs.berkeley.edu
amplab.cs.berkeley.edu
"This paper takes the position that disk-locality is go- ing to be irrelevant in cluster computing, and considers the implications this will have on datacenter computing research. The quest for disk-locality is based on two as- sumptions: (a) disk bandwidths exceed network band- widths, (b) disk I/O constitutes a considerable fraction of a task’s lifetime. We show that current trends undermine both these assumptions."