Task-Based Programming on Emerging Parallel Architectures for Finite-Differences Seismic Numerical Kernel

Abstract

The recent advances on microprocessors architecture have brought a large amount of computational power through numerous heterogeneous computing devices. For instance, a typical modern supercomputer comprises a set of interconnected computing nodes, each of which having several multi-core processors and hardware accelerators such as Nvidia GPU and Intel MIC. Moreover, moving forward to exascale platforms, estimations indicate that computing nodes will feature up to 1000 cores. However, to fully take into advantage these resources, a profound shift on the implementation of numerical applications has to be performed. In this paper, we study the design and implementation of a seismic wave propagation simulator, based on finite difference scheme, and specifically tailored for massively parallel architectures. The application data-flow is built on top of PaRSEC, a generic task-based runtime system, targeting distributed heterogeneous architectures. Considering the memory-bound nature of the stencil scheme, we designed the numerical kernels to maximize data reuse and thus increasing their arithmetic intensity. Such a strategy allows to efficiently exploit large SIMD units available in modern CPU cores. To illustrate the performance of the application, we conducted a strong scalability study on a cluster of Intel KNL processors. The obtained results compare favorably to an existing seismic wave propagation code.

Improving parallel executions by increasing task granularity in task-based runtime systems using acyclic DAG clustering

“…Task-based implementations to solve the transport equation on a grid (i.e. structured and regular mesh) have already been proposed by Moustafa et al (2018). The authors have created a version on top of the ParSEC runtime system where they partitioned the mesh and avoided working on the graph of tasks.…”

“…Moreover, the well-known and widely used OpenMP standard (OpenMP Architecture Review Board, 2013) also supports the tasks and dependencies paradigm since version 4. The advantage of the method to achieve high-performance and facilitate the use of heterogeneous computing nodes has been demonstrated by the development of many applications in various fields (Sukkari et al, 2018;Moustafa et al, 2018;Carpaye, Roman & Brenner, 2018;Agullo et al, 2016;Agullo et al, 2017;Agullo et al, 2015;Myllykoski & Mikkelsen, 2019).…”

Peer Review #2 of "Improving parallel executions by increasing task granularity in task-based runtime systems using acyclic DAG clustering (v0.1)"

“…Task-based implementations to solve the transport equation on a grid (i.e. structured and regular mesh) have already been proposed by Salli Moustafa et al [10]. The authors have created a version on top of the ParSEC runtime system where they partitioned the mesh and avoided working on the graph of tasks.…”

“…Moreover, the well-known and widely used OpenMP standard [8] also supports the tasks and dependencies paradigm since version 4. The advantage of the method to achieve high-performance and facilitate the use of heterogeneous computing nodes has been demonstrated by the development of many applications in various fields [9,10,11,12,13,14,15].…”

Untitled

“…Moreover, these approaches cannot adapt their executions to the noises generated by the OS or the hardware. This is why most task-based applications use runtime systems that are powered with dynamic scheduling strategies [21,22,23,24,9]. In this case, the scheduler focuses only on the ready tasks and decides during the execution on how to distribute them.…”