Skip to main content
NASA Advanced Supercomputing (NAS) Division World-class modeling, simulation, and supercomputing in support of the NASA mission

The Parallel Research Kernels, a Tool for Parallel Systems Investigations - Part II

AMS Seminar Series

Speaker: Rob Van Der Wijngaart, Intel Corporation

Presentation

Seminar Slide Deck (PDF-3MB)

Abstract

The Parallel Research Kernels (PRK) are a suite of kernels that can be used to obtain insight into the performance of complete HPC platforms, particularly parallel computers. Each kernel is a proxy for an important pattern in HPC. Traditionally, we selected kernels that could be load balanced perfectly statically. However, with the imminent arrival of exascale computers and their expected levels of system inhomogeneity and noise, the need for dynamic load balancing emerged. Due to the complexity of load balancing at the application level, a number of runtime solutions have been proposed that provide automatic dynamic load balancing. To be able to rate and compare the efficiency of these approaches, we need workloads with controlled levels of load imbalance. This seminar will present two new PRKs, PIC and AMR, that proxy the two main patterns of load imbalance in HPC. Important difficulties in developing such kernels are the need for simplicity and compactness, and the need for analytical solutions. We will show how the new kernels satisfy these requirements, and will also show salient performance results for our reference implementations.

Biography

Rob Van Der Wijngaart spent 12 years at NASA Ames Research Center, mostly in the NAS division, where he was instrumental in the implementation of the MPI versions of the NAS Parallel Benchmarks, co-created the NAS Grid Benchmarks and the BT-IO Benchmark, and designed the NAS Parallel Benchmarks Multi-Zone versions. While there he also designed and implemented Charon, a library supporting incremental porting of serial applications to distributed environments. After his stint at NASA he joined Intel, where he has spent the last decade on various research projects, all related to parallel High Performance Computing. Among others, he developed applications and runtime libraries for some of Intel’s many-core research processors, including the 80-core Polaris research processor and the 48-core Single-chip Cloud Computer. Most recently he has focused on studying advanced runtimes enabling exascale computing.