The International Conference for High Performance Computing, Networking, Storage and Analysis
Fast Prediction of Network Performance: k-packet Simulation.
Student: Nikhil Jain (University of Illinois at Urbana-Champaign)
Supervisor: Laxmikant Kale (University of Illinois at Urbana-Champaign)
Abstract: Prediction of application communication performance on supercomputer networks, without doing actual runs, is useful for a variety of what-if analyses: how does the performance change with different task mappings, studying the performance of future networks etc. A significant amount of research is conducted to study these aspects because the scalability of many applications is adversely affected by communication overheads. Flit-level and packet-level simulations have been shown to be useful for prediction, but they are inefficient and such simulations are very slow even for moderate-sized networks. We propose a new simulation methodology that relies on increasing the granularity of simulation to k-packets and use of simple heuristics for predicting the state of the network. Preliminary results show that the proposed approach accurately models network behavior, and is orders of magnitude faster than the previous methods: up to two minutes per prediction for 16,384 cores of Blue Gene/Q using the given benchmarks.