Program

In this talk we examine how high performance computing has changed over the last 10-year and look toward the future in terms of trends. These changes have had and will continue to have a major impact on our software. Some of the software and algorithm challenges have already been encountered, such as management of communication and memory hierarchies through a combination of compile--time and run--time techniques, but the increased scale of computation, depth of memory hierarchies, range of latencies, and increased run--time environment variability will make these problems much harder. We will look at five areas of research that will have an importance impact in the development of software and algorithms. We will focus on following themes:

• Redesign of software to fit multicore and hybrid architectures
• Automatically tuned application software
• Exploiting mixed precision for performance
• The importance of fault tolerance
• Communication avoiding algorithms

Weak-scaled parallel execution has demonstrated dramatic advances in the last two decades through the combination of MPP and commodity cluster architectures with communicating sequential processes programming methods (e.g., MPI). Increased performance in that time frame for suitable applications (e.g., Linpack) has been observed in excess of four to five orders of magnitude. However, many applications requiring reduction of execution time for fixed sized data sets have exhibited less favorable progress. These strong scaled applications are constrained by limited parallelism, high overheads, delays due to remote access latencies, and contention for shared resources. Conventional practices mitigate such challenges through regular coarse grained tasks avoiding significant effects of these sources of performance degradation. This presentation will discuss recent results of experiments in strong scaling using lightweight synchronization constructs including dataflow and futures objects combined with message-driven computation and dynamic thread scheduling. The ParalleX execution model is an experimental synthesis of these and other methods that expose and exploit parallelism inherent to the meta-data of dynamic graph-based applications. A proof-of-concept reference implementation of ParalleX, HPX-3, is a runtime system that serves on conventional SMPs and commodity clusters with Unix-like operating system interfaces. The goal of this research has been to investigate the potential of constraint-based synchronization for detection and dynamic scheduling of this form of parallelism for strong scaling. An advanced Adaptive Mesh Refinement code used for numerical relativity studies was originally developed in MPI and has been ported to the HPX library. Initial measurements suggest that such methods may exhibit significant improvements of scalability compared to conventional practices. This talk will present these findings and conclude with requirements for future work.

TSUBAME2.0 is the latest incarnation of the series of clusters that have been built at Tokyo Institute of Technology, and has become the first supercomputer in Japan to reach the Petaflops plateau. TSUBAME 2.0 embodies many unique features derived from years of research into HPC, especially keeping in mind retaining or improving bandwidth scalability, fault tolerance, green, using the latest hardware components such as GPUs and SSDs, as well as employing some of the latest software research results from labs at Tokyo Tech. I will also touch upon its possible use to simulations of natural disasters that have hit Japan recently, demonstrating that despite its relatively small size as well as adoption of hybrid architectures it scales well to the use of thousands of GPUs as well as demonstrates performance topping the largest machines such as the ORNL Jaguar.