Logo of Science Foundation Ireland  Logo of the Higher Education Authority, Ireland7 CapacitiesGPGPU Research Projects
Ireland's High-Performance Computing Centre | ICHEC
Home | News | Infrastructure | Outreach | Services | Research | Support | Education & Training | Consultancy | About Us | Login

Publication

Title:Enabling Smeagol on Xeon Phi: Lessons Learned
Authors:Alin M. Elena, Ivan Rungger, 2014
Abstract: The aim of the work is to port the quantum electron transport code Smeagol to the Intel Xeon Phi (MIC) based hybrid architecture. Smeagol is based on density functional theory, and utilises the non-equilibrium Green's function formalism (NEGF) to calculate the charge density at an applied bias voltage. We successfully ported Smeagol to run in native mode on the MIC, however for small systems the run-time on the host is always significantly shorter than the one on the MIC. It is not possible to run large systems on a single MIC due to memory limits. For large systems we adopt a different strategy for efficiently using this hybrid architecture: we propose to run most of the calculation on a number of host processors, distributed over a set of nodes, using MPI, and to port only the most time-consuming operations to the MIC for each host. This allows us to take advantage of the large memory on the host, which enables calculations for large systems. We show that in Smeagol the most time-consuming operation is the inversion required to obtain the Green's function within the NEGF formalism. Our results indicate that for sufficiently large systems our implementation of the recursive Green's functions (RGF) inversion algorithm can indeed run faster on the MIC plus host setup than on the host. This is also true for the whole Smeagol code utilised in such an offload mode.
ICHEC Project:
Publication:PRACE Whitepaper
URL: http://www.prace-ri.eu/IMG/pdf/wp134.pdf
Keywords: Xeon Phi, Smeagol
Status: Published

return to publications list