Performance analysis with a memory-bound Monte Carlo simulation on Xeon Phi

Physics simulations are known to be great resources exhausters (CPU, memory). Hardware acceleration can help reduce the need for CPU time and increase the available memory bandwidth. In this paper, we present the performance gain when running a memory-bound muon Monte Carlo simulation on an Intel Xeon Phi and an Intel Xeon CPU. We show how to increase performance on the Xeon Phi without modifying the Physics software frameworks we are using for our application. We investigate distributed simulations on multicore and manycore systems and also the impact of hyper-threading on performance. We extend this to a hybrid computing model, balancing the computing burden between both the manycore and multicore processors of a computing node. Finally, we improved memory usage on the Xeon Phi by sharing Kernel Memory pages using KSM, and we show that, using this approach, we can run 16% more simulation instances.