ECSS Symposium February 2015
February 17, 2015
Migration to Phis and Enhancing Vectorization
Presenter(s): Jim Browne (University of Texas)
This talk will cover two topics: selection of execution modes for use of heterogeneous compute nodes which include Intel Xeon Phis and enhancing vectorization in application codes. Each lecture will be about 20 minutes. The lecture on selection of execution modes for heterogeneous compute nodes describes a systematic process for choosing among CPU only, Phi only, symmetric node and offload execution modes. The lecture is based on the Quick Start Guide developed by the Stampede Technology Insertion project. The Quick Start Guide is available from the TACC web site. Enhancement of vectorization is accomplished by combining compiler static analysis with runtime measurements of execution behavior to generate recommendations for adding directives, pragmas and source code changes to increase vectorization. The average gain in performance obtained in the case studies was in excess of 40% for execution on the Xeon Phis. The process for enhanced vectorization is supported by a tool called MACVEC which will soon be available for general use on Stampede.