guest ::
| Home > Publications database > Trace-based performance simulation of large-scale applications |
| Home > Publications database > Trace-based performance simulation of large-scale applications |
| Master Thesis | PreJuSER-5144 |
2009
Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag
Jülich
Please use a persistent id in citations: http://hdl.handle.net/2128/3665
Report No.: Juel-4297
Abstract: The massively parallel computer architectures emerged in the last years create the platform to redefine the limits of todays scientific simulations. To exploit these platforms efficiently, applications need a yet unprecedented scaling behavior to several thousands of processes. The complexity of systems of this magnitude presents a challenge to performance prediction in general. Sometimes it is feasible to extrapolate from small-scale execution behavior to larger scales, however, this is not always the case. Applications and algorithms that change their behavior significantly when executed with a high number of processes will then have to be investigated on the corresponding scale. Several performance analysis tools exist to help the developer of large-scale simulations to identify performance critical phenomena in their application execution. Choosing the best optimization strategy is still highly depending on the experience of the performance investigator with the application, its algorithms, the analysis tools, and the computer architecture as well as its software. Cost-benefit ratios of specific code optimizations are often hard to estimate precisely. Prediction of a modified application’s execution behavior on large scales can help with estimating the cost-benefit ratio of a specific application modification. This thesis introduces the application performance simulator Silas, which uses an event-trace-based approach to model and predict application execution behavior. Its focus lies on the simulation of hypothetical code optimizations, based on the modification of an existing execution trace. It is embedded in the performance analysis tool Scalasca, which is a scalable set of tools supporting performance investigators in optimizing large-scale applications. Simulated optimizations may include the scaling of region instances, the balancing of parallel region instances, and the deletion of region instances and message transfers from the event trace. A model for trace-based performance simulation as well as the needed event-trace manipulation to simulate optimizations are presented. The implementation of specific parts of the simulator is discussed and test results of Silas investigating synthetic and real-world applications to demonstrate its effectiveness are presented.
|
The record appears in these collections: |
PDF
Fulltext by OpenAccess repository