Author: Felix Mössbauer
Title: DASH Benchmarking and Performance Assessment
Abstract:
The ever growing computational power in HPC systems is bought by
a steadily rising number of compute cores and nodes. Hence,
efficient utilization of all resources requires massive
parallelism and high demands on the programming skills. To meet
this challenges, the DASH C++ library provides an implementation
of the PGAS programming model that features a virtual global
address space. Therewith DASH provides distributed data
structures as well as algorithms to work on the containing
data. This thesis evaluates the performance of DASH compared to
UPC++ and "native" implementations using micro benchmarks.  As
benchmarking in the area of HPC depends strongly on the system
topology, various methods for topology discovery are evaluated.
Furthermore, conceptual shortcomings and their solution
possibilities of the PGAS model are presented. Finally the
performance of DASH is demonstrated by a matrix multiplication
that outperforms state of the art linear algebra libraries.