Progress has been achieved to realizing the goal of creating a large MPI computer to use hundreds of processors by pooling the cluster resources of the MecaGRID members. This report summarizes experiences in developing and performing computations on the MecaGRID project for 8-64 processors. The MecaGRID connects the INRIA clusters (70 CPUs) with the IUSTI (155 km, 30 CPUs) and CEMEF Ecole des Mines de Paris at Sophia Antipolis (1 km, 24 CPUs) 15. The AERO-F (mono phase) and AEDIF (two-phase flow) codes have been successfully adapted to GRID Computing. Implicit and explicit solvers were evaluated using the AERO-F mono-phase code and two-phase flows using the AEDIF code.
The largest test case contained more than 1,000,000 million vertices with 7 million unknowns. The MecaGRID runs with 64 CPUs were 4-5 times slower than the same application on the INRIA clusters. The loss in efficiency using the MecaGRID is due to increased communication times needed to transfer data between the three MecaGRID clusters. Local communication times increased by factors of 100-200 over the same run on the INRIA clusters.
Several experiments suggest that the domain decomposition may have an impact on the efficiency of the Grid application16The role of mesh decomposition on optimizing commnications needs to be examined and will be an important subject to creating more efficient Grid Computing.
Advances in network transfer technology are needed before an efficient MecaGRID is possible. In 3-5 years high-speed optical networks should be a reality and Grid Computing will become as efficient as cluster computing today.
An interesting option available to create an efficient MecaGRID is to take advantage of the existing 2.5 Gbps VTHD highspeed network connecting INRIA Sophia Antipolis and INRIA Rhone Alpes. This would require a partnership between the two INRIAs sites as INRIA Rhone Alpes is not at present a member of the MecaGRID. Other options would be to create high-speed networks between the existing MecaGRID member sites.