Portable heterogeneous computing on distributed HPC systems

Abstract

Heterogeneous hardware has been increasingly adopted at all scales of computer systems to meet the high and specific demands of current applications. Given the wide diversity of accelerators available, cross-platform programming models such as SYCL have emerged to offer a unified abstraction layer. This thesis addresses the development of portable heterogeneous applications using Intel's implementation of SYCL, Intel oneAPI, and evaluates their performance on distributed HPC systems. To this end, a microbenchmark was developed to solve iterative problems, applying dynamic workload balancing and considering CPU+GPU and CPU+FPGA configurations. Furthermore, the degree of automation provided by SYCLomatic for migrating CUDA to SYCL was examined, analysing the performance portability of native and migrated code.