PhD opportunities

Hybrid code synthesis from a pure distributed memory program

Area of expertise Real-time computer science, robotics, systems and control - Fontainebleau
Doctoral School SMI - Sciences des Mtiers de l'Ingnieur
Title Hybrid code synthesis from a pure distributed memory program
Supervisor M. Claude TADONKI
Co-supervisor M. Claude TADONKI
Contact TADONKI Claude
Research unit Mathematics and Systems
CRI - Centre de Recherche en Informatique
Keywords parallelism, memory
Abstract With the advent of multicore architectures we need us to reconsider the way we organize the tasks of our parallel code and plan their cooperation. Current (and future) supercomputers are built up with multicore processors probably coupled with accelerators such as GPUs. Regarding the connectivity, latency and network bandwidth have certainly evolved significantly, but the gap between the virtual topology and physical topology greatly increases the effective cost of interprocessor communications, especially for applications involving a high connectivity like stencil computation. Therefore, it becomes crucial to really consider hybrid implementations, which consider distributed memory model on top and shared memory model on the nodes. However, several factors are likely to discourage this effort: the required skill to design and implement hybrid codes is not common; in addition, number of quite complex codes have already been written based on the distributed memory model exclusively, and it is hard to consider modifying them; moreover, the belief of a clear effective reward from a hybrid code is moderate, because getting a benefit as close as expected is indeed not trivial. Therefore, designing a quasi-systematic methodology and a corresponding framework, that allow to easily switching from a pure MPI code to an efficient corresponding MPI + OpenMP version would be a valuable contribution. In this thesis, we propose to tackle the problem quite rigorously, on the basis of a tasks graph model and a clearly quantified accounting of major hardware and network aspects. The first goal will be to understand the criteria for an efficient and scalable hybrid code, followed by a methodology for a quick and pragmatic implementation. Applications envisaged for conducting this study are: Lattice QCD; pattern matching on a large sequence of high resolution images; partial differential equations; shortest paths computation in a graph.
Funding Concours pour un contrat doctoral
Partnership
Starting date October 1st 2017
Date of first publication January 6th 2016