2018-01-15
Title: Programming Multi-BSP Algorithms in ML
Speaker: Victor Allombert
Abstract: From personal computers using an increasing number of cores, to supercomputers having millions of computing units, parallel architectures are the current standard. The high performance architectures are usually referenced to as hierarchical, as they are composed from clusters of multi-processors of multi-cores. Programming such architectures is known to be notoriously difficult. Writing parallel programs is, most of the time, difficult for both the algorithmic and the implementation phase. To answer those concerns, many structured models and languages were proposed in order to increase both expressiveness and efficiency. Among other models, Multi-BSP is a bridging model dedicated to hierarchical architecture that ensures efficiency, execution safety, scalability and cost prediction. It is an extension of the well known BSP model that handles flat architectures. We introduce the Multi-ML language, which allows programming Multi-BSP algorithms “à la ML” and thus, guarantees the properties of the Multi-BSP model and the execution safety, thanks to a ml type system. To deal with the multi-level execution model of Multi-ML, we defined formal semantics which describe the valid evaluation of an expression. To ensure the execution safety of Multi-ML programs, we also propose a typing system that preserves replicated coherence. An abstract machine is defined to formally describe the evaluation of a Multi-ML program on a Multi-BSP architecture. An implementation of the language is available as a compilation toolchain. It is thus possible to generate an efficient parallel code from a program written in Multi-ML and execute it on any hierarchical machine.
The DISCOVERY initiative aims at exploring a new way of operating Utility Computing (UC) resources.
methodological needs of HPC application and runtime developers and to allow to study real HPC systems both from the correctness and performance point of view. To this end, we gather experts from the HPC, formal verification and performance evaluation community. website : 
o propose to revisit the principles of existing schedulers after studying the main factors impacted by job submissions. Then, we will propose novel efficient algorithms for optimizing the schedule for unconventional objectives like energy consumption and to design provable approximation multi-objective optimization algorithms for some relevant combinations of objectives (performance, fairness, energy consumption, etc.). An important characteristic of the project is its right balance between theoretical analysis and practical implementation. The most promising ideas will lead to integration in reference systems such as SLURM and OAR as well as new features in programming standards implementations such as MPI or OpenMP. We expect MOEBUS results to impact further use of very large scale parallel platforms.