Programming Challenges for Many-Core Computing

Many-core architectures face significant hurdles to successful adoption by ISVs, and ultimately, the marketplace. One of the most difficult is addressing the programmability problems associated with parallel computing. For example, it is notoriously difficult to debug a parallel application, given the potential interleavings of the various threads of control in that application. Another problem is that predicting performance, even at coarse accuracy, is extremely inaccurate. I will explain why a chip company like Intel is interested in advanced programming languages research and believes this is critical to adoption of many-core architectures. Intel's Programming Research Lab is addressing these issues for both client and server computing, in particular media and gaming workloads. We are implementing a high-level programming abstractions based on transactional memory, data parallel programming models and functional languages. In this talk, I will briefly discuss a language based on Nested Data Parallelism (NDP) called Ct. NDP models have the advantage of being deterministic, meaning that the functional behaviors of sequential and parallel executions of an NDP program are always the same for the same input. Data races are not possible in this model. Furthermore, NDP models have an easy to understand coarse performance model, which can be made more accurate for specific architectural families. This enables the programmer to comprehend the performance implications of their code well-enough to make well- informed algorithmic choices. About the speaker Anwar Ghuloum earned degrees at the University of California, Los Angeles (B.S., Computer Science and Engineering) and Carnegie Mellon University's School of Computer Science (Ph.D., Computer Science, 1996), where his thesis introduced concepts of Nested Data Parallel idioms to traditional parallelizing compilers. Anwar has been a Senior Staff Scientist with Intel's Programming Systems Lab since joining in early 2002, working on diverse topics such as optimizing memory system performance, parallel architecture evaluation, parallel language and compiler design, and multimedia applications. Before that, he co-founded and was the CTO of a fab- less semiconductor startup called Intensys that built programmable, highly parallel image and video processors for the consumer electronics market. Prior to that, Anwar developed novel predictive drug design software for early lead optimization using 3D surface pattern recognition techniques for a biotech startup called MetaXen (acquired by Exelexis Pharmaceuticals). He has also served as a post-doctoral research associate at Stanford University's Computer Science department. A recurring theme in Anwar's work has been to bridge high-level application knowledge and low-level parallel architecture constraints with careful parallel language and compiler design