Ray Characterization

Last week, I characterized how parallel CP and RAY were by examining the CUDA files for both the benchmarks. I found that CP was more parallel than RAY -- which mande it an interesting benchmark to analyze despite its small size.

This week, we wanted to finish our characterization of RAY by making it more parallel in software. This entailed changing the image size that the Ray Tracer would operate on and the number of thread blocks (and consequently the number of threads in the queue at a given time). The following are the calculations related to the tweaked version of RAY which serves as an indicator of the degree of parallelism with which RAY executes on the GPU:

Number of threads in each block = 16 x 8 = 128 threads.

Number of thread blocks issued = (512/16, 512/8) = (32, 64)

Thus the number of blocks in the queue at a given time = 32 x 64 = 2048.

This ensured that the RAY benchmark was executing as parallel in software as the Coulombic Potential benchmark in the number of work elements that it was issuing. Given that the RAY CUDA file issues more instructions in its core computation that CP does, and that a lot of its instructions (CUDA instructions) seem to be computationally more expensive, RAY can now be treated as both a long program which would occupy a significant amount of space in the instruction cache (and hence would not be entirely resident in the instruction cache) AND would issue those instructions with a high degree of parallelism.

In order to examine the affect of changing the software parallelism issued by Ray, we repeated the methodology that Fabiha and DH had employed when they were examining the performance of the different benchmarks as a function of the number of SM cores.