Selecting an Optimal Interconnect Network for PARSEC

In this paper we investigate the communication performance of three high end supercomputers from the National Energy Research Scientific Computing Center (NERSC), and discover which machine allows for the optimal scaling for the large scale, scientific code PARSEC. Our first step is to discover the relevant message passing function calls that PARSEC makes use of so that we can test appropriate communication properties on each of the systems. We then discover the execution time for each of the relevant message passing calls by executing a small benchmark on each machine while increasing the number of processors between 32 and 256 by a factor of two in each step. These results are then used to hypothesize which machine should allow for the greatest speedup for PARSEC. This theory is tested by actually executing the PARSEC code on all three systems while increasing the number of processors. The specific machines (and interconnects) used are: Cray XT4 (Cray SeaStar2), IBM p575 (IBM Federation), and IBM SP RS/6000 (IBM Colony). Lastly, the setup of the communication backbone is investigated so that we can understand why the different architectures perform better/worse than others, and the specific protocols that the supercomputers use is presented, along with a discussion of how they work.