Design considerations for parallel performance tools

Abstract

In recent years there has been a shift in microprocessor manu facture from building single-core processors towards provid ing multiple cores on the same chip. This shift has meant that a much wider population of developers are faced with the task of developing parallel software: a difficult, time con suming and expensive process. With the aim of identifying issues, emerging practices and design opportunities for sup port, we present in this paper a qualitative study in which we interviewed a range of software developers, in both industry and academia. We then perform a systematic analysis of the data and identify several cross-cutting themes. These analysis themes include the practical relevance of the probe effect, the significance of orchestration models in development and the mismatch between currently available tools and developers' needs. We also identify an important characteristic of parallel programming, where the process of optimisation goes hand in hand with the process of debugging, as opposed to clearer distinctions which may be made in traditional programming. We conclude with reflection on how the study can inform the design of software tools to support developers in the endeav our of parallel programming.

Parallel Performance Problems on Shared-Memory Multicore Systems: Taxonomy and Observation

“…In interviews carried out for a previous study [23], this was found to be a difficult issue. The model does not provide clear indications to reliably support identification of this problem as the expert assessments were in disagreement.…”

“…Bug Classifications. It has been argued that within parallel program development, the distinction between performance problems and bugs is much less clear than in traditional software development [23]. However, the vast majority of parallel performance problems would receive only the broadest categorization under existing taxonomies.…”

“…The synchronization issue requires various events occurring within the processor to be correlated [46]. While some existing visualizations are potentially useful, there is a need for analysis of how such tools can aid in the diagnosis of problems [23].…”

“…In this section, we present one possible taxonomy, developed iteratively and derived from several sources. The initial starting point was an earlier qualitative study [23] in which interviews with parallel software developers were fully transcribed and coded following a qualitative research methodology, resulting in 582 open codes. While the focus of the qualitative analysis was a broad characterization of needs and practices, as a by-product of the analysis, a range of different types of performance problem emerged, including load balancing, lock contention and saturating memory bandwidth.…”

“…While the diagnosis of race conditions and deadlocks are important problems [72], [73], we aim to focus the taxonomy exclusively on performance problems, as opposed to correctness problems. However, it should be noted that the boundary between correctness and performance in concurrent systems is fuzzy [23]. Deadlock is a good example of this, where it can be seen as both correctness issue and a degenerate case of lock contention where multithreaded execution cannot progress without external intervention, either automated or manual.…”