Dynamic Load balancing and CFD Simulations on the MecaGRID and GRID5000 - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Reports (Research Report) Year : 2006

Dynamic Load balancing and CFD Simulations on the MecaGRID and GRID5000


CFD simulations on Clusters and GRIDS having mixed processor speeds present several challenges to achieve efficient load balancing. If both the fast and slow processors are given the same amount of work, the faster processors will finish their computations first and wait for the slower processors to finish. To achieve load balancing more work must be given to the faster processors so that all the processors finish their computations at the same time (work is proportional to the processor mesh size). Another complication is that for current Clusters and GRIDS in the near future, the user will not know in advance the mixture of fast and slow processors that will be assigned to their computation, thus the user cannot partition the mesh in advance of the CFD simulation. This difficulty is doubly complicated as the mesh partitioning step is usually performed on a workstation thus not directly linked to the parallel CFD code. For mesh partitioners executing on parallel computers, the complication arises in that the mesh partitioning code and the CFD code are separate MPI codes designed to be run independently of each other. As a result the two codes cannot simply be run back-to-back as each code may be assigned different mixtures of fast and slow processors resulting in a partitioned mesh not optimal for the CFD run. In this study, in order to overcome the problems related to computing with arbitrary mixtures of fast and slow processors, the mesh generator has been integrated into the CFD code. Thus optimal size partitions are automatically created for different mixtures of fast and slow processors. The efficiency of this approach is demonstrated for Clusters, the MecaGRID and the GRID5000. Validation tests using the GRID5000, the MecaGRID, and the INRIA nina-pf cluster produced speedups on the order of 1.32 to 1.52 relative to the same run using the homogeneous partitioning which compares well with the theoretical speedup of 1.5. Finally, we use the dynamic computing capability of the new CFD code to compute a 20 million vertices mesh using 256 processors at five different GRID5000 sites.
Fichier principal
Vignette du fichier
RR-5884.pdf (284.7 Ko) Télécharger le fichier

Dates and versions

inria-00071384 , version 1 (23-05-2006)


  • HAL Id : inria-00071384 , version 1


Stephen F. Wornom. Dynamic Load balancing and CFD Simulations on the MecaGRID and GRID5000. [Research Report] RR-5884, INRIA. 2006. ⟨inria-00071384⟩
147 View
122 Download


Gmail Facebook Twitter LinkedIn More