WP3 takes care of the continuous assessment and analysis of the parallel performance of the MAX flagship codes, pointing up the direction for the development aimed at the effective exploitation of the existing technology. To this aim, we make use of tools for code profiling and optimisation and of the most recent programming models. WP3 activity is functional to many other tasks in different work packages (WP1, WP2, WP4, WP5), in order to provide feedback on the progress obtained in terms of performance enhancement with respect to the relevant metrics and to the parallel efficiency. On the other hand, an important output of this WP is to discover and monitor code bottlenecks, to identify the code or architecture feature responsible for them (memory bandwidth, communication imbalance, latency, bandwidth to GPU, etc), and to propose dedicated solutions also through the engineering of ad-hoc proof-of-concepts. The solutions of these bottlenecks that require code refactoring or replacement of an algorithm will be implemented within WP1 by the code developers. Our activity will therefore be in continuous synergy with the developing teams of the MaX flagship codes. In the following we will describe how we have organised the benchmarking and profiling activity, introducing the tools adopted (Section 3) and presenting some preliminary results (Section 4, 5 and 6). The deployment status of flagship applications on EuroHPC systems will also be reported (Section 7).
Building similarity graph...
Analyzing shared references across papers
Loading...
Fabio Affinito
Laura Bellentani
Jan Jona Javoršek
Scuola Internazionale Superiore di Studi Avanzati
Jožef Stefan Institute
Istituto Nanoscienze
Building similarity graph...
Analyzing shared references across papers
Loading...
Affinito et al. (Sun,) studied this question.
www.synapsesocial.com/papers/69e9b95b85696592c86ec266 — DOI: https://doi.org/10.5281/zenodo.19678007