Shared-Memory Parallel Probabilistic Graphical Modeling Optimization: Comparison of Threads, OpenMP, and Data-Parallel Primitives

Abstract

This work examines performance characteristics of multiple shared-memory implementations of a probabilistic graphical modeling (PGM) optimization code, which forms the basis for an advanced, stateof-the art image segmentation method. The work is motivated by the need to accelerate scientific image analysis pipelines in use by experimental science, such as at x-ray light sources, and is motivated by the need for platform-portable codes that perform well across many different computational architectures. The primary focus of this work and its main contribution is an in-depth study of shared-memory parallel performance of different implementations, which include those using alternative parallelization approaches such as C11-threads, OpenMP, and data parallel primitives (DPPs). Our results show that, for this complex dataintensive algorithm, the DPP implementation exhibits better runtime performance, but also exhibits less favorable scaling characteristics than the C11-threads and OpenMP counterparts. Based upon a set of experiments that collect hardware performance counters on multiple platforms, the reason for the runtime performance difference appears to be due primarily to algorithmic efficiency gains: the reformulation from the traditional C11-threads and OpenMP expression of the solution into that of data parallel primitives results in significantly fewer instructions being executed. This study is the first of its type to do performance analysis using hardware counters for comparing methods based on VTK-m-based data-parallel primitives with those based on more traditional OpenMP or threads-based parallelism. It is timely, as there is increasing awareness of the need for platform portability in light of increasing node-level parallelism and increasing device heterogeneity.

Performance Analysis of Traditional and Data-Parallel Primitive Implementations of Visualization and Analysis Kernels

“…One primary difference between these previous works, except for Perciano, et al, 2020 [17], and our work here is the deeper introspection provided by using detailed hardware performance counters. These additional metrics offer the ability to better understand why a given code performs better or worse in a particular set of circumstances, and also helps to provide a more sound basis for performance analysis.…”

“…The three cases we present all exhibit different aspects of why a method might have better or worse runtime than another. In some cases, the way an algorithm is implemented, such as VTK vs. VTK-m, can have a dramatic impact on overall number of instructions, a fact that is corroborated by other recent studies (c.f., [17]). In other cases, the buffer management needed to implement a complex, multi-stage processing pipeline may trigger more memory movement instructions, which may be more expensive and result in higher CPI values, and we see evidence of this in two of the examples.…”

A parallel Breadth-First Search using shared memory level-synchronization

“…The emphasis is not merely on parallelization but also on meticulous optimizations. These refinements strategically curtail layer synchronization overheads and mitigate the intricacies tied to race conditions, ensuring the algorithm's robustness [3]. Concurrently, a discerning evaluation measures the algorithm's performance enhancements, specifically gauging the speedup in relation to the count of engaged threads.…”