The Center for Computational Science and Engineering is conducting research on the computational fluid dynamics of combustion. The application under development will help improve the understanding of the dynamics of such systems as diesel engines and industrial burners. Currently, we can model 3D systems using relatively simple reaction kinetics models. Our codes run effectively on the T3E using a library of software written to support Adaptive Mesh Refinement (AMR) algorithms for fluid systems on distributed memory parallel systems. The present algorithm is well suited to the existing T3E architecture in terms of CPU use, but we under-utilize the available system memory. Presently, each production simulation results in a considerable amount of output, corresponding to roughly one tenth of the available RAM per simulation phase. We find that we can easily use 25GB per production job. Each production job is one part of an overall series of jobs that model one physical experiment. For that reason we make heavy use of mass storage facilities. Currently our I/O Bandwidths scale well for processor to disk storage up to 256 processors. Some system configuration or AMR library software changes will be required to use 512 processors. Our software environment is heavily dependent on C++ and Fortran compilation systems. We are very disappointed with the current debugging and performance analysis tools available on the Cray T3E.
The next generation version of our software is likely to require upwards of 10 times the floating performance of the current code, mainly because of the use of a more realistic reaction kinetics model. In addition, the memory requirements will increase by roughly a factor of two. One implication is that our Disk requirements will likely increase by a factor corresponding to the CPU requirement, a factor of 10. Disk usage per production run will increase to 250GB, with corresponding increases need in the requirements for mass storage.