Making Parallel Applications Communication-Aware: From Models to Implementations
Invited talk: Thomas Lambert presented "Making Parallel Applications Communication-Aware: From Models to Implementations" in visio the 27/3/2020 at 11h00.
Parallel applications like scientific applications and MapReduce applications transfer data between tasks, subsequent processes, nodes, etc. These data transfers (communications) have a significant impact on the performance of parallel applications. Moreover, replication, which is usually used to exploit parallelism and improve resilience in parallel applications, can be utilized to optimize communication. In this talk, I will demonstrate the importance of communication-awareness using three widely used applications including: MapReduce and stream data applications, container provisioning and dense matrix-matrix multiplication. More precisely, I will show how to characterize and model the communication patterns of parallel applications and use these models to build and implement algorithmic solutions that can help to improve the performance of parallel applications.
Thomas Lambert received a master degree from ENS of Lyon, and a PhD degree from the University of Bordeaux in 2017. He is currently a post-doctoral researcher at Inria Rennes in the STACK team. His main interests in research are scheduling aspects of problems linked to High-Performance Computing and Big Data applications, in particular issues related to replication of data in parallel computing.