|Name:||(A03) SCAMP: A “Pseudo” Trace Driven Simulation toward Scalable Network Evaluation|
|Time:||Wednesday, June 22, 2016
10:00 am - 11:00 am
|Breaks:||10:00 am - 11:00 am Coffee Break|
|Presenter:||Miwako Tsuji, RIKEN/AICS|
|Abstract:||Co-design for HPC is a bidirectional approach: where a system would be designed on demand from applications, and applications must be optimized to the system. In order to co-design the network of large scale systems, it is important to evaluate the communication performance of applications. While the trance driven network simulator is widely used to estimate the communication performance, it is difficult to evaluate large future systems if the current system is smaller than the target ones.
We propose SCAMP (SCAlable Mpi Profiler) to tackle the scaling-problem in the trace driven network simulator. Firstly, user runs an application on a current system, which may be smaller than the target system, to obtain trace-files. Then, the trace-files are copied and modified to create "pseudo" trace-files. The number of "pseudo" trace-files is equal to the target number of nodes and larger than the current system. After that, an existing network simulator is driven using the pseudo files.
We perform experiment to compare the run-time estimated by SCAMP and real run-time. SCAMP performs well for very simple benchmarks. However, it sometimes underestimates the elapsed times because (1) we have not taken into account the change of computation time due to the change of the system size and (2) other effects such as OS-jitter are ignored.
We consider applications with simple communication patterns since it is easy to create "pseudo"trace-files. While real-world applications have complex communication patterns, HPC-applications would have relatively simple communication patterns. We'll investigate the applicability of SCAMP as a future work.
Miwako Tsuji, RIKEN AICS; Jinpil Lee, RIKEN AICS; Taisuke Boku, University of Tsukuba; Mitsuhisa Sato, RIKEN AICS