journée du groupe de travail LaMHA - Automne 2018

In my past time I like to watch old movies, not because they are old per se, but because they must be really good if after all that time, people still remember them. Similarly, I like old parallel algorithms. In this talk I like to talk about a combination of two classics: Bruck's all-to-all and Valiant's two-phase randomized routing. Both describe fast ways of communicating between processes in a parallel computer. By combining the two you get a new communication algorithm, which effectively transforms a high-throughput network into a low-latency network. Although Bruck's and Valiant's algorithms are true classics, I will present them both and, of course, also their combination: Randomized Bruck. Cost analysis is included, but also demonstrated with some benchmark results on a small HPC cluster.

BSML and Multi-ML are functional parallel programming languages “à la ML” based of the respectively the bsp and Multi-BSP bridging models. Multi-BSP extends bsp to take into account hierarchical architectures. For both models, it is possible to predict the performances of algorithms thanks to embedded cost models. To do so, we propose formal a operational semantics with cost annotations for the two aforementioned languages. This work has been done in a incremental manner. First we recall the cost semantics of core-ml language. Then, we adapt it to bsml and we adapt it to Multi-ML. It is then possible to evaluate the cost of a program following the annotated semantics. Finally, we compare the theoretical approach with the current implementation on a code example.