Parallel Tree Partitions and their Role in Scalable Computing

Decanato - Facoltà di scienze informatiche

Data: 12 Novembre 2018 / 10:30 - 11:30

USI Lugano Campus, room SI-006, Informatics building (Via G. Buffi 13)

Speaker:	Carsten Burstedde
	Universität Bonn, Germany
Date:	Monday, November 12, 2018
Place:	USI Lugano Campus, room SI-006, Informatics building (Via G. Buffi 13)
Time:	10:30

Abstract:

In this talk, we illustrate a specific encoding of forest-of-octrees meshes that enables particularly elegant and highly scalable recursive algorithms for many common tasks. These algorithms solve local/near problems, which are often symmetric between sender and receiver processes, as well as non-local/far and asymmetric problems. Assembling the ghost layer of a mesh, for example, classifies as near since direct neighbor processes communicate. Reassigning moving objects between processes, on the other hand, links processes whose partitions may be far from each other in space, and a sender to some process does not necessarily receive from that process. Another prominent non-local/far problem arises in parallel visualization, when we match the subdomain of a data process with the subimage of a compositing process. We are inspired by large-scale applications and exemplify how the above algorithms translate into compelling parallel scalability.

Biography:

Carsten Burstedde is Professor at the Institute for Numerical Simulation at University Bonn, Germany. He is developing numerical techniques for accurate and efficient computer simulations. A particular focus is on scalable algorithms that run on the largest supercomputers worldwide. A cornerstone of this research is the p4est software that he is lead-authoring.

Host:

Prof. Rolf Krause