Christophe Jaillet

MultiGPU computing using MPI or OpenMP

The GPU computing follows the trend of GPGPU, driven by the innovations in both hardware and programming languages made available to nongraphic programmers. Since some problems require an important time to solve or data quantities that do not fit on one single GPU, the logical continuation was to make use of multiple GPUs.

Solving the Langford problem in parallel

In this paper, the parallel resolution of the Langford problem is studied. Two different approaches are developed. First, an explicit construction of all the solutions is done using a shared memory. The application associated to this approach is written in C using the standard OpenMP library. Second, a parallelization of the algebraic method introduced by Godfrey is proposed. The application is taking advantage of MPI and has revealed efficient up to 128 processors. This solution opens up some new perspectives such as solving the already resolved instances of the problem more quickly and solving the next two open instances of the problem in a near future.

BFS Traversal on Multi-GPU Cluster

The High Performance Computing field is striving for the next computational step: exascale (1018 floating point operations per second). The first machine at this scale is expected at horizon 2020 and some French companies such as Atos/Bull are already working on it. Their supercomputer, Sequana, will deliver the expected computational power using either Xeon Phi or GPU accelerators. However these machines are built to produce an exaFLOPS, they are exclusively based on computation. The question is how they will behave when confronted with highly irregular problems such as graph theory ones. Graph algorithms are considered to be highly irregular with intensive communication and require high memory resources (but also I/O). In this study our goal is to provide an optimized implementation of the Graph500 benchmark to rank the Kepler GPU based supercomputers (a specific class of HPC architectures) in the Graph500 list.

A New Memory Allocation Model for Parallel Search Space Data Structures with OpenMP

OpenMP is a shared memory programming API (see [2] or http://www.openmp.org ), brought on UMA and CC-NUMA architectures, which supports multithreaded applications. The present study deals with the management and arrangement of memory, which may cause useless memory blocks reloading from the shared memory to the processors caches. This occurs for some memory access sensible applications, when programming with OpenMP: cache faults may congest the system and induce a communication over-cost between the processors. A new memory allocation model is presented, enabling us to solve the memory contention phenomenon in the specific case of shared memory programming with OpenMP.

A parallel shared memory simulator for command and control

Introduces a military application in the command and control field. The main feature of this study is the parallelization of the simulator. The simulator is object-oriented and written in C++. It uses the OpenMP standard for the parallel version. To produce an efficient parallel simulator, we have to deal with the dynamic load balancing problem.