Characterizing and Understanding PDES Behavior on Tilera Architecture

The emergence of many core architectures with shifting balance between computation and communication overhead can have a tremendous impact on performance and scalability of fine-grained parallel applications such as PDES. It may also be necessary to rethink the design philosophy of key PDES subsystems, that were traditionally focussed on hiding long communication delays. In this paper, we perform extensive evaluation of PDES on Tile64Pro - a new 64-core chip from Tilera. For our studies, we use the recently developed multithreaded version of the popular ROSS simulator and show that the performance of this simulator (with many optimizations proposed) scales by a factor of 27X when it is executed on 56 cores of the Tilera chip for Phold benchmark with 20% remote communication. We also evaluate the impact of performance optimizations that we propose on both conservative and optimistic versions of the simulator and also analyze the sensitivity to various simulation parameters. Finally, we explore the issues of object placement and model partitioning on Tilera architecture.