Analysing an SQL application with a BSPlib call-graph profiling tool

This paper illustrates the use of a post-mortem call-graph profiling tool in the analysis of an SQL query processing application written using BSPlib [4]. Unlike other parallel profiling tools, the architecture independent metric of imbalance in size of communicated data is used to guide program optimisation. We show that by using this metric, BSPlib programs can be optimised in a portable and architecture independent manner. Results are presented to support this claim for unoptimised and optimised versions of a program running on networks of workstations, shared memory multiprocessors and tightly coupled distributed memory parallel machines.