Performance analysis and Optimisation of the Met Unified Model on a Cray XC30

TL;DR

This paper analyzes and optimizes the UK Met Office's Unified Model on a Cray XC30 supercomputer, demonstrating performance improvements through MPI and OpenMP optimizations, with implications for heterogeneous architectures.

Contribution

It presents a detailed performance analysis and optimization strategy for the Unified Model on a Cray XC30, including the use of Cray tools and parallelization techniques.

Findings

MPI configuration improvements yield 5-12% performance gains.

Adding OpenMP directives results in 5-16% speedup.

Optimizations save tens of millions of core-hours in climate simulations.

Abstract

The Unified Model (UM) code supports simulation of weather, climate and earth system processes. It is primarily developed by the UK Met Office, but in recent years a wider community of users and developers have grown around the code. Here we present results from the optimisation work carried out by the UK National Centre for Atmospheric Science (NCAS) for a high resolution configuration (N512 $\approx$ 25km) on the UK ARCHER supercomputer, a Cray XC-30. On ARCHER, we use Cray Performance Analysis Tools (CrayPAT) to analyse the performance of UM and then Cray Reveal to identify and parallelise serial loops using OpenMP directives. We compare performance of the optimised version at a range of scales, and with a range of optimisations, including altered MPI rank placement, and addition of OpenMP directives. It is seen that improvements in MPI configuration yield performance improvements of between 5 and 12\%, and the added OpenMP directives yield an additional 5-16\% speedup. We also identify further code optimisations which could yield yet greater improvement in performance. We note that speedup gained using addition of OpenMP directives does not result in improved performance on the IBM Power platform where much of the code has been developed. This suggests that performance gains on future heterogeneous architectures will be hard to port. Nonetheless, it is clear that the investment of months in analysis and optimisation has yielded performance gains that correspond to the saving of tens of millions of core-hours on current climate projects.