Turbulent combustion research is carried out under the direction of Suresh Menon of the Computational Combustion Laboratory, Georgia Tech School of Aerospace Engineering.
LESLIE2D Computation Combustion Code
Performance profile of LESLIE2D was undertaken to compare the quad Pentium Pro 200 MHz cluster against the Cray-T3E.
The first graph shows a speedup comparison as a function of processor count on the two machines for three different problem sizes. Clearly, for the problem sizes chosen, the speedup is much better on the Cray-T3E than on the IHPC cluster. The main noticeable feature is the large drop in performance from 16 to 32 processors. The reason for this lies in the large amount of data that is exchanged between tasks at each time step and the single network interface card in the IHPC machines that results in significant contention for the network interface. A look at the scaleup of the computational costs alone, as is shown in the next figure, shows excellent scalability with increasing granularity.
The second graph compares the pure computational power of two processors (SGI and Pentium Pro 200 MHz) per timestep as a function of the number of grid points per processor. On an average, the SGI processor is over 4 times faster than the Pentium Pro processor and scales better with increase in problem size.
The final graph shows timings for writing checkpointing (i.e., restart) data on the IHPC cluster for four different problem sizes. The methodology used by the model is to gather all checkpointing data at one processor then access the disk from the single processor. The methodology results in large network interface contention for all problem sizes as is seen by the insensitivity of the I/O times to increases in problem sizes.
