The HECToR Service is now closed and has been superceded by ARCHER.


This Distributed Computational Science and Engineering (dCSE) project was to improve the performance of the GWW component of Quantum Espresso (QE), which is a freely available collection of applications for ab-initio simulation. GWW is an addition to QE which uses a basis of localised Wannier orbitals within the GW approximation to calculate polarisation.

The project concentrates on the performance of the 3D Fast Fourier Transform (FFT) implementation, which is the main performance bottleneck. The work was split into two parts. The first part was to investigate optimisations to the current 3D FFT implementation, in particular the global communications required during data transposition. The second part was to evaluate and incorporate a new data decomposition, switching from using a 2D plane decomposition to a more scalable 1D pencil decomposition.

The individual achievements of the project are summarised below:

  • Three alternative methods for global communication have been tested with a 400% speedup achieved (for phase 2a) in isolation, translating into a 4-36% improvement in the application (part 1).
  • A checkpoint mechanism has been added to the code so that 12-hour plus jobs can be restarted (part 1).
  • A 3D FFT pencil decomposition has been developed and tested in isolation and presented in non-application-specific detail in the report. The report also shows how this approach improves scalability (part 2). (However, the pencil decomposition has not been fully incorporated into the GWW codebase.)
  • Optimising the configuration of the virtual 2D grid of processes has shown a 27% performance improvement in 3D FFT performance (part 2).

Please see part 1 PDF and part 2 PDF or part 1 HTML and part 2 HTML for the reports.