Seminar in Numerical Analysis: Fabio Baruffa (Leibniz-Rechenzentrum, München)

Towards code modernization and porting on Intel many-core architecture: the story of Gadget code

In the framework of the Intel Parallel Computing Centre at Leibniz Supercomputing Centre (LRZ), Fabio Baruffa will present recent results on the performance optimization of Gadget-3 on multi and many-core computer architectures, including the new Intel Xeon Phi processor of second generation, codenamed Knights Landing (KNL). An overview of results for node-level scalability, vector efficiency and performance are presented here. Our work is based on an isolated, representative code kernel, where threading parallelism, data locality and vectorization efficiency was improved. The node-level parallel efficiency improved by factors ranging from 5x to 16x on Haswell and KNL nodes, respectively. Moreover, a vectorization efficiency of 80% (6.6x) on a prototypical target loop of the code is obtained without programming using intrinsics instructions.

Export event as iCal