Predicting Power and Timing of Large-Scale Distributed Applications on Highly Heterogeneous Platforms


Applications in the High-Performance Computing (HPC) domain are often designed to run on cluster-like distributed platforms with hundreds of nodes. Due to the size of both, application and platform, predictions of application run time and energy consumption during application design are challenging.
I will develop a new approach to this challenge based on a simulation technique well-known in the embedded compu-ting domain. In order to apply such a methodology at HPC scale, I cannot execute actual applications during simula-tion. Instead, I run a symbolic simulation based on the Task Graph model of computation, which is popular in HPC. At the same time, I keep a high level of detail for communication, as this has much bigger impact on overall accuracy. The end result will be a time and energy prediction methodology that is fast enough for interactive feedback in the application design workflow while still maintaining a useful level of accuracy. It also allows designers to employ au-tomatic design space exploration across a wide range of high-level design choices like choice of algorithm, degree of parallelism, or task granularity.
Betreuer: Prof. Dr. Wolfgang Nebel


17. Juni 2019, 16:15


Jörg Walter, Universität Oldenburg


OFFIS, Escherweg 2, Raum F 02

Wegabm3x1goastsz0er ( (Stand: 10.09.2018)