MCS Division research featured in SUPER newsletter

May 30, 2014

SUPER – or the Institute for Sustained Performance, Energy, and Resilience – is a DOE SciDAC institute involving four national laboratories and eight universities. The goal is to ensure that DOE’s computational scientists can successfully exploit the emerging generation of high-performance computing systems.

The Argonne SUPER team, led by Paul Hovland, is developing multiobjective variants of the derivative-free nonlinear optimization algorithms used for search in the SUPER empirical autotuning systems. The April 2014 newsletter of SUPER highlights recent accomplishments in this area.

The newsletter describes the cutting-edge research bridging mathematical and performance optimization by SUPER team members Prasanna Balaprakash and Stefan Wild, from Argonne, and their collaborator Ananta Tiwari, from SDSC. Their findings, to appear in the Proceedings of the 4th International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems, show that whereas in some settings performance objectives (such as run time, power consumption, and energy consumption) are strictly correlated and there is a single, ideal decision point, in other settings significant tradeoffs exist. Emphasizing the importance of this work, Bob Lucas, the SUPER institute director, writes, “The results will allow users of leadership class systems to weigh the tradeoffs between alternative strategies so as to achieve the objectives most important to their applications.”

Also featured in the April newsletter is an interview with Balaprakash, who received his Ph.D. in engineering sciences from the Université Libre de Bruxelles, Belgium, and did his postdoctoral work at Argonne. He currently has a joint appointment with the MCS Division and the Leadership Computing Facility as an assistant computer scientist. Balaprakash states in the interview: “I am excited about utilizing my background in performance engineering and machine learning to form collaborations with other research groups. . . . I hope to advance the state of the art in performance modeling for future extreme-scale systems by increasing the applicability of mathematical modeling, machine learning, and optimization approaches.”

More details about this research can be found in the full paper “Multi-objective optimization of HPC kernels for performance, power, and energy,” at http://www.mcs.anl.gov/papers/P4069-0413.pdf. For further information about SUPER, see the SUPER website http://www.super-scidac.org/ and the newsletter at http://www.super-scidac.org/sites/super-scidac.org/files/SUPERNEWSLETTER... .