Supercomputing Wales - Task 008

Work Package:
Completed

Description

Using records retained from the time period, and having conversations with staff at the Supercomputing Wales partner institutions, we will document the approaches used for software performance characterisation and improvement at the Supercomputing Wales partner institutions, and prepare a document outlining them, including the principles underpinning them and lessons learned from them.

Outcomes

Fit to programme

This task has been identified by the working groups as part of the agenda behind WP 1.1.

The task number is 008.

Description

Supercomputing Wales (SCW) has a long-standing record of software performance characterisation featuring staff at both Hub Sites – Cardiff and Swansea. Staff at Cardiff have built on the original Architecture Comparison Exercise (ACE [1]), a UK based initiative aimed at developing a qualitative methodology for understanding and predicting the performance of a series of large-scale HPC applications across a spectrum of massively-parallel HPC architectures. This has evolved into the regular CIUK presentations focused on the performance analysis of a variety of multi-core high-end cluster systems from Intel and AMD. Swansea University developed the BSMBench benchmark designed to evaluate supercomputers, a code that transitioned into Sombrero, that became part of the UKRI ExCALIBUR benchmarking suite and has in fact provided a template for developing other benchmarks. Swansea also pioneered the use of AccelerateAI, a GPU compute cluster designed to enable high-speed AI/ML computations, available to all SCW institutions.

Outcomes

  • Document existing work on ‘software performance characterisation’ from Supercomputing Wales.
  • Publish summmary on SHAREing webpage.

This project has now finished. The outcomes can be found in a report on the SHAREing webpage.

Literature

  1. A. Gray et al: Mapping application performance to HPC architecture, CPC 183 (2012) pp. 520–529