joomla templates

Sunday, 26 Mar 2023

A Library of Automatically Tuned Sparse Matrix Kernels for Graphics Processing Units

E-mail Print

A project that has progressed from Educational to Preparatory and then to Production access.


Professor Walid Abu-Sufah's computational group from the University of Jordan started using LinkSCEEM resources through educational access provided to them through a cluster of PC's which also possessed graphic cards. Educational Access was then given to them on Euclid which was a more structured hybrid CPU-GPU cluster upon which the group could test and run CUDA code.


The group later applied for a number of Preparatory Access projects where development, parallelisation and optimisation of their code was carried out. Coupled to this was the scalability testing required for the production access application completed by the group in the 3rd LinkSCEEM/Cy-Tera call.


The group was allocated twelve thousand GPU hours on Cy-Tera to design, implement, and test fundamental sparse matrix computations kernels for high-performance execution on graphics processing units.