A Practical Approach for Performance Analysis of Shared-Memory Programs

Abstract

Abstract-Parallel programming has transcended from HPC into mainstream, enabled by a growing number of programming models, languages and methodologies, as well as the availability of multicore systems. However, performance analysis of parallel programs is still difficult, especially for large and complex programs, or applications developed using different programming models. This paper proposes a simple analytical model for studying the speedup of shared-memory programs on multicore systems. The proposed model derives the speedup and speedup loss from data dependency and memory overhead for various configurations of threads, cores and memory access policies in UMA and NUMA systems. The model is practical because it uses only generally available and non-intrusive inputs derived from the trace of the operating system run-queue and hardware events counters. Using six OpenMP HPC dwarfs from the NPB benchmark, our model differs from measurement results on average by 9% for UMA and 11% on NUMA. Our analysis shows that speedup loss is dominated by memory contention, especially for larger problem sizes. For the worst performing structured grid dwarf on UMA, memory contention accounts for up to 99% of the speedup loss. Based on this insight, we apply our model to determine the optimal number of cores that alleviates memory contention, maximizing speedup and reducing execution time.

Towards Modelling Parallelism and Energy Performance of Multicore Systems

“…Using extensive measurement analysis on state of the art multicore systems [8], [9], we conclude that for weak-scaling programs the last-level misses, number of cycles unrelated to memory contention and work cycles do not change significantly when n changes. Furthermore, we study the pattern of burstiness of the memory requests and conclude that large parallel programs do not exhibit bursty memory traffic [9].…”

“…We define ω(n), the memory contention factor, as the number of threads busy due to memory overhead to the number of threads busy due to useful work. With these notations, following the derivations described in [8], the speedup of a shared-memory program is:…”

“…The average prediction error across more than two hundreds experiments with four problem sizes, and core counts ranging from 2 to 48 is 9% for UMA systems and 14% for NUMA systems. The accuracy of the model is strongly determined by the problem size, with best accuracy for larger problem sizes, due to large steady-state compute phases [8] and uniform memory access patterns [9].…”

“…The practicality of the models stems from their generality across programming languages/threading packages and architectures, and from new insights on the decrease of memory burstiness in multicore systems, for large problem sizes. These insights are based on extensive experiments on state-of-the-art multicore systems, with core counts up to 4 in ARM systems, and 48 in commodity UMA and NUMA systems [8], [9]. 2) Using predictions of the model, we optimize execution of parallel programs by adjusting the number of cores or the core frequency in systems with dynamic frequency scaling.…”

“…2) Using predictions of the model, we optimize execution of parallel programs by adjusting the number of cores or the core frequency in systems with dynamic frequency scaling. The optimality criteria ranges from fastest execution time to minimum energy usage, and trade-offs between these two [8].…”