אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

Nowadays, multi- and many-core architectures are becoming the focus of affordable High Performance Computing (HPC). In this context, one of the programming models (and paradigms) that rapidly gains popularity is OpenCL, originally designed and developed to exploit heterogeneous (massive) parallelism. In this talk, I will briefly introduce our directions of research involving OpenCL, many-core architectures, and HPC, and ranging from low-level memory access patterns analysis to large-scale graph processing. However, the main focus of my talk is a discussion on OpenCL's portability. While the inter-platform portability is currently one of the model's most attractive features, performance portability - i.e., obtaining similar performance on different platforms using the same OpenCL code - remains a matter of dispute in both its usefulness and its feasibility. Therefore, I will introduce the main challenges that inter-platform performance portability poses and present empirical solutions to tackle them. Furthermore, based on our results, I will present the "paradox of portability", show what are the limitations of the OpenCL's portability model, and open the discussion about the steps that could be taken to address these limitations.