ELSEVIER Parallel Computing 23 ( 1997) S-2 I PARALLEL COMPUTING Realistic parallel performance estimation Tony Hey *, Alistair Dunlop, Emilio Hemtindez Depurtment r$Elrctrcmics und Computer Science. Uniuersiry of Southampton. Sourhompton SOI 7 IBJ, UK Abstract This paper presents an account of two tools which enable realistic parallel performance estimates to be obtained in reasonable time. The first tool, PERFORM(performance estimator for RISC microprocessors), is a general purpose package for sequential program performance estimation on modem RISC microprocessor nodes with complex memory hierarchies. The novelty of the tool is that we have chosen an intermediate level of abstraction between simple statement counting and full simulation of the node architecture. The second tool, LEBEP, generates synthetic parallel programs from a simple communication specification. In particular, it can generate communication operations that incurs typical communication overhead, including data movements across different levels in the memory hierarchy. Keywords: Performance tools; Benchmarks; Communication; Message-passing; Memory hierarchy 1. Introduction Realistic parallel performance estimation depends critically on two key performance aspects of a parallel program - single node performance and communication costs. Simplistic approaches to node performance estimation using static statement counting and to communication costs using simple ‘ping-pong’ benchmarks fail to capture reality in several significant ways, and can result in ‘estimates’ being in error by more than an order of magnitude. The most important shortcoming is due to lack of attention to the complex memory hierarchies of modem RISC microprocessor nodes. Another, potentially important shortcoming, is concerned with possible network contention effects: for modern, dedi- cated parallel machines with their high bandwidth networks, any effects due to this cause are likely to be dwarfed by failure to incorporate an accurate representation of the ’ Correspondingauthor. 0167-8 I91 /97/S 17.00 Copyright 0 I997 Elsevier Science B.V. Ail rights reserved. PI1 SOl67-Sl9l(96)00093-2