Domain‐specific virtual processors as a portable programming and execution model for parallel computational workloads on modern heterogeneous high‐performance computing architectures

Abstract

We advocate domain‐specific virtual processors (DSVP) as a portability layer for expressing and executing domain‐specific computational workloads on modern heterogeneous HPC architectures, with applications in quantum chemistry. Specifically, in this article we extend, generalize and better formalize the concept of a domain‐specific virtual processor as applied to scientific high‐performance computing. In particular, we introduce a system‐wide recursive (hierarchical) hardware encapsulation mechanism into the DSVP architecture and specify a concrete microarchitectural design of an abstract DSVP from which specialized DSVP implementations can be derived for specific scientific domains. Subsequently, we demonstrate, an example of a domain‐specific virtual processor specialized to numerical tensor algebra workloads, which is implemented in the ExaTENSOR library developed by the author with a primary focus on the quantum many‐body computational workloads on large‐scale GPU‐accelerated HPC platforms.

Implementation of Relativistic Coupled Cluster Theory for Massively Parallel GPU-Accelerated Computing Architectures

“…The ExaCorr module provides two distinct implementations of coupled cluster methods, one intended for execution on a single shared-memory node with an optional GPU acceleration and another one for execution on many such nodes (distributed parallelism), thus supporting a broad variety of computer platforms, from simple workstations to leadership HPC systems. Both implementations use the ExaTENSOR library 24 as a massively parallel GPU-accelerated processing backend for numerical tensor algebra operations, although there are some differences in the interface between the single- and multinode API. For the single-node runs (with OpenMP multithreading and/or GPU acceleration), only the single-node component of ExaTENSOR, the TAL-SH library, 53 is used.…”

“…These extensions of the ExaTENSOR library are implemented as extensions of an abstract tensor transformation class provided by the ExaTENSOR interface. Once this is done, the ExaTENSOR parallel runtime (domain-specific virtual processor 24 ) is started within a provided MPI communicator. After initialization, ExaTENSOR will begin accepting commands to perform distributed tensor algebra operations that realize a given coupled cluster algorithm.…”

“…In our reimplementation, we therefore do not consider molecular symmetries but instead focus on data and compute parallelism. The ExaCorr implementation that we describe here is based on the ExaTENSOR library, 24 a distributed numerical tensor algebra library for GPU-accelerated HPC platforms developed at the Oak Ridge Leadership Computing Facility (OLCF).…”

Special Issue on Emerging Architectures in Computational Chemistry

“…However, new algorithms with potentially new numerical representations are required, and new programming approaches must be adopted. Lyakh introduces abstract that promise to increase performance portability and hide from the scientist some of the complexity of programming current heterogeneous high‐performance computer designs. Mironov et al describe the OpenMP+MPI hybrid parallelization of the Fragment Molecular Orbital (FMO) method in GAMESS and its excellent performance on current NSF and DOE supercomputers.…”

Implementation of relativistic coupled cluster theory for massively parallel GPU-accelerated computing architectures

“…In our reimplementation, we therefore do not consider symmetry but instead focus on data and compute parallelism. The ExaCorr implementation that we describe here is based on the Ex-aTENSOR library, 23 a scalable numerical ten-sor algebra library for GPU-accelerated HPC platforms developed at the Oak Ridge Leadership Computing Facility (OLCF).…”