As the architecture of modern supercomputers becomes increasingly heterogeneous and diverse, traditional programming approaches (e.g., MPI + OpenMP + some dedicated GPU API) hit some fundamental limits that make it hard to develop mathematical software that is efficient and, at the same time, portable and simple to maintain and extend. In this lecture we will show how task-based parallelism can be used to overcome these limitations. This parallel programming paradigm consists in describing an algorithm as a graph of tasks whose execution is delegated to a runtime system. This paradigm allows for very asynchronous execution where communications and computations can be easily overlapped and for very high productivity because algorithms can be expressed at a very high level with all the complex handling of parallelism and data transfer being delegated to the runtime. We will show how this programming paradigm can be used to implement some scalable linear algebra algorithms.
Speaker
Alfredo Buttari – CNRS, IRIT, Université de Toulouse
Alfredo Buttari, is currently a CNRS Research Director at the IRIT laboratory of Toulouse where he previously held a
Researcher position from 2008 to 2021. He is a member of the APO (Algorithmes Parallèles et Optimisation) team. In 2006 he received the PhD degree in computer science of the University of Rome “Tor Vergata”; successively, in 2006-2007 he held a post-doctoral position at the Innovative Computing Laboratory of the University of Tennessee Knoxville under the direction of Jack Dongarra and, in 2008 a post-doctoral position at the LIP laboratory, ENS-Lyon. He is an expert of high performance, parallel computing, mathematical software and parallel algorithms for dense and sparse linear algebra.
Event Timeslots (1)
Fri 20 – Applications & Performance Evaluation
-
A. Buttari (CNRS, IRIT, Université de Toulouse)