ExCALIBUR HES Benchmarking for Performance Portable Applications
Lead Research Organisation:
University College London
Department Name: Centre for Advanced Research Computing
Abstract
Application benchmarking is a crucial activity in the UK's path to Exascale. It ensures future Exascale
systems are understood by the community so that UK Science Applications, particularly those
developed as part of ExCALIBUR, can take advantage of the scientific opportunities at Exascale. Linear
solvers are a ubiquitous pattern in many ExCALIBUR codes, and this project will develop a suitable
benchmark applicable to many of UK codes. Application benchmarks are part of the enabling software
stack that allows the performance of hardware to be assessed for scientific workloads .
It is vital to rigorously measure the performance of benchmarks in a systematic way to enhance the
transparency and enable reproducibility. International conferences in this space are now mandating
reproducibility information. This project provides the tooling to make collecting and analysing
performance data straightforward. This will allow application owners to develop the skills and
experience to measure the performance of their code across the gamut of the UK's supercomputer
facilities, and improve their code based on benchmarking as they target Exascale.
The increasing diversity of computer architectures now requires applications to strive to be
Performance Portable (PP); that is, achieve a high level of efficiency on a range of different processors
with minimal changes to the source code. Measuring PP using analysis such as those in our recent
studies needs to become as common in HPC as scaling tests or roofline analyses. This is especially
true for ExCALIBUR codes, where the likely Exascale architectures may not be known for several years.
This project will ensure this vital PP analysis can become ubiquitous in HPC software development.
The ReFrame tool has been developed by the CSCS and ETH Zurich to aid in the automated collection
of performance data. ReFrame is an already established approach within ExCALIBUR. This project will
augment the rudimentary performance analysis available in ReFrame to include our own PP
Methodology developed with collaborators at Intel . By incorporating this alongside the Benchmarking
Initiative, we will empower the supercomputing community in the UK to measure the performance
portability of their code across the UK's HPC infrastructure. This aligns with the H&ES mission to ready
UK codes and software for Exascale.
systems are understood by the community so that UK Science Applications, particularly those
developed as part of ExCALIBUR, can take advantage of the scientific opportunities at Exascale. Linear
solvers are a ubiquitous pattern in many ExCALIBUR codes, and this project will develop a suitable
benchmark applicable to many of UK codes. Application benchmarks are part of the enabling software
stack that allows the performance of hardware to be assessed for scientific workloads .
It is vital to rigorously measure the performance of benchmarks in a systematic way to enhance the
transparency and enable reproducibility. International conferences in this space are now mandating
reproducibility information. This project provides the tooling to make collecting and analysing
performance data straightforward. This will allow application owners to develop the skills and
experience to measure the performance of their code across the gamut of the UK's supercomputer
facilities, and improve their code based on benchmarking as they target Exascale.
The increasing diversity of computer architectures now requires applications to strive to be
Performance Portable (PP); that is, achieve a high level of efficiency on a range of different processors
with minimal changes to the source code. Measuring PP using analysis such as those in our recent
studies needs to become as common in HPC as scaling tests or roofline analyses. This is especially
true for ExCALIBUR codes, where the likely Exascale architectures may not be known for several years.
This project will ensure this vital PP analysis can become ubiquitous in HPC software development.
The ReFrame tool has been developed by the CSCS and ETH Zurich to aid in the automated collection
of performance data. ReFrame is an already established approach within ExCALIBUR. This project will
augment the rudimentary performance analysis available in ReFrame to include our own PP
Methodology developed with collaborators at Intel . By incorporating this alongside the Benchmarking
Initiative, we will empower the supercomputing community in the UK to measure the performance
portability of their code across the UK's HPC infrastructure. This aligns with the H&ES mission to ready
UK codes and software for Exascale.
