ORCHA -- A Performance Portability System for Post-Exascale Systems

TL;DR

ORCHA is a system that uses code synthesis and code generation to enable performance portability of complex scientific applications across diverse post-exascale hardware platforms, demonstrated on the Flash-X simulation software.

Contribution

It introduces a novel toolchain with three components for expressing control flow, data variants, and runtime orchestration, tailored for scientific software on heterogeneous systems.

Findings

Successfully mapped Flash-X to three different hardware configurations

Demonstrated performance portability across GPU-centric, balanced, and concurrent layouts

Enabled exploration of configuration space for optimal performance

Abstract

Heterogeneity is the prevalent trend in the rapidly evolving high-performance computing (HPC) landscape in both hardware and application software. The diversity in hardware platforms, currently comprising various accelerators and a future possibility of specializable chiplets, poses a significant challenge for scientific software developers aiming to harness optimal performance across different computing platforms while maintaining the quality of solutions when their applications are simultaneously growing more complex. Code synthesis and code generation can provide mechanisms to mitigate this challenge. We have developed a toolchain, ORCHA, which arises from the needs of a large multiphysics simulation software, Flash-X, which were not met by any of the existing solutions. ORCHA is composed of three stand-alone tools -- one to express high-level control flow and a map of what to execute where on the platform, a second one to express variants of data structures and arithmetic operations in the solvers in a unified fashion, and a third one that manages the runtime orchestration of the data and computation. We use an application-specific interface layer that uses code generation and code synthesis to stitch together the application. In this paper, we describe the interface layer for the application Flash-X and demonstrate the use of ORCHA in exploring possible configurations from which the optimal one can be selected for production, including a case study in which a single simulation recipe is realized on three distinct hardware mappings -- a GPU-centric, a CPU/GPU balanced, and a CPU/GPU concurrent layouts -- highlighting the breadth of configurations ORCHA enables.