Section: New Results
Distributed Shared Memory approach for the steering of parallel simulations
As a different approach of EPSN, we conceived and developed an in-transit visualization framework for interfacing an arbitrary HPC simulation code with an interactive ParaView session using the HDF5 parallel IO library as the API. The library called H5FDdsm is coupled with a ParaView plugin ICARUS (Initialize Compute Analyze Render Update Steer).
Because our interface is based on files, stored in a distributed shared memory (DSM), we sought during this year different redistribution strategies to optimize the bandwidth and the transfers between the simulation and the ParaView servers hosting the DSM. This work showed real benefits, particularly on one of our Cray XE6 testing machines using a block cyclic redistribution. On these large HPC machines that do not support the dynamic MPI process management set of functions, we improved our connection system so that simulation and post-processing can be coupled within an MPMD job. Taking also advantage of one-sided communication models and of the Cray Gemini interconnect communication performance, our framework has been sensibly improved and should be optimal in the coming months.
The interface has also been enhanced with a steering interface that allows us to control the simulation work-flow and send back not only parameters, but also complete meshes in parallel, which can then be read by the simulation using either our steering interface or HDF5 calls. This has been demonstrated with SPH-flow, a CFD code developed by Ecole Centrale de Nantes and HydrOcean, replacing dynamically and in parallel a falling wedge with a deforming sphere.
This work has been realized and is currently carried on at CSCS - Swiss National Supercomputing Centre in the framework of Jérôme Soumagne PhD thesis (under the co-supervision of Mr. John Biddiscombe) and within the NextMuSE European project 7th FWP/ICT-2007.8.0 ( ,  ,  ).