IPERF : A Framework for Automatic Construction of PerformancePrediction

Performance prediction models at the source code level are crucial in optimizing compilers, programming environments, and performance debug-ging tools. For each performance prediction task, minimal-cost models are needed that deliver the required accuracy. Current techniques to derive the desired models are ad hoc. This paper discusses a new framework for automatic construction of cost-eeective performance prediction models for diierent target systems at the program source level. A target system consists of a target optimizing compiler, a target operating system, and a target architecture with a multi-level memory hierarchy. Preliminary results for a small computation kernel on a set of target systems indicate the eeectiveness of the proposed framework. 1 Motivation Performance models are needed to guide the compiler or programmer in the code optimization process by supporting the decision about which optimization transformation to apply next. The ideal performance model matches exactly the decision process for an optimization or a set of optimizations, i.e., it ranks the optimization alternatives correctly according to their expected performance beneets. Since a more detailed performance model is typically more expensive to compute, the ideal performance model is also minimal in the sense that it ignores target system components that are not needed to distinguish the optimization alternatives. Finding a matching, ideal performance model for an optimization and target system is hard, or even impossible since such a model may not exist. Recent trends in architecture and language design pose new challenges to optimizing compilers and their performance models. For example, the success of In-tel's and Hewlett Packard's new IA-64 architecture (Merced) will depend on the ability of optimizing compilers to take advantage of the parallelism provided by the machine, putting an even greater burden on the compiler to achieve eecient program execution than for current superscalar architectures. Another new challenge for optimizing compilers are dynamic compilation strategies for languages such as Java. Dynamic optimizations are performed at program execution time and therefore have more severe compile time constraints than static compilation systems. In the context of this paper, a target system is a target compiler consisting of a set of optimization passes, a target operating system, and a target architecture with a multi-level memory hierarchy. The diiculty of performance prediction is a result of the diierent sets of optimizations performed by the target compiler and their interactions with the advanced operating system and architecture features of the target system. It is important to note …