Energy Efficiency Focus in the SC12 Technical Program
by Natalie Bates, Co-chair Energy Efficient HPC Working Group (EE HPC WG)
Energy efficiency will again be a hot topic at SC12, with at least 38 Technical Program sessions focused on energy efficiency. A complete list of these sessions organized both chronologically and by topic can be found on the Energy Efficient HPC Working Group website. SC12, the annual International Conference for High Performance Computing, Networking, Storage and Analysis, will be held Nov. 10-16 in Salt Lake City, Utah. For more information, see the SC12 website.
BROAD SCOPE SESSIONS
“The Third Annual Workshop on Energy Efficient High Performance Computing – Redefining System Architecture and Data Centers” promises to be interesting to a broad audience. Some of the featured speakers include; Peter Kogge, University of Notre Dame who will look at the historical trends of power, energy and supercomputing; John Shalf, Lawrence Berkeley National Laboratory whose talk will focus on the energy requirements for applications; as well as Herbert Huber, Leibniz Supercomputing Center and Steve Hammond, National Renewable Energy Laboratory who will speak about energy efficient data centers.
There are four other technical programs that will cover the topic of energy efficiency at a high level. Kirk Cameron, Virginia Tech is on the slate to give two talks, both of which have clever and enticing titles with phrases about a “Growing Power Struggle” and “Energy Oddities.” Prohibitive energy costs motivated Thomas Ludwig, German Climate Computing Center to consider the cost and benefits of “HPC-Based Science in the Exascale Era.” Finally, there is a “Cool Supercomputing” Birds of Feather (BoF) organized by Pacific Northwest National Laboratory that covers tools and techniques for optimizing energy consumption at all levels.
“Setting Trends for Energy Efficiency” is a BoF representing a collaborative effort by the Top500, Green500, the Energy Efficient HPC Working Group and The Green Grid to standardize the power measurement methodology used when running system workloads for architectural comparison, such as High Performance Linpack. This is one of seven sessions that cover energy efficiency measures and metrics. The Green500, Top500 and now the Graph500 have their own BoFs and will report power consumption and energy efficiency as well as performance for their Lists. The High Performance Group at at the Standard Performance Evaluation Corporation (SPEC) has also organized a BoF that will discuss a new OpenMP benchmark suite with an optional energy metric that scales to 512 threads. From the home of the Green500 at Virginia Tech, Balaji Subramaniam will present his doctoral showcase on metrics for energy efficiency. Finally, an Intel team will present a paper on tuning for the Graph500 Traversal which includes both performance and energy efficiency results.
SESSIONS FOCUSSED ON SYSTEM HARDWARE
Thirteen of the sessions are exploring system hardware energy efficiency. Of these thirteen, seven of them focus on alternative processors like GPU and ARM that are continuing the trend towards aggregating low-power processors and using accelerators. There are three BoFs that explore alternative processors and all three are organized by Europeans. The Partnership for Advanced Computing in Europe (PRACE) explores a set of prototypes to test and evaluate promising new technologies for future multi- Petaflop/s systems that include GPUs, ARM processors, DSPs and FPGAs. The Barcelona Supercomputing Center is heading up an ARM-based exascale demonstration and will review their research results and plans at two BoFs; “Energy Efficient HPC” and “Exascale Research- The European Approach.” Besides these BoFs, there is a session as part of Broader Exchange where Calxeda, an ARM-based server provider, will present their products and roadmaps. NEC is presenting an exhibitor forum on “Hybrid Solutions with a Vector-Architecutre for Efficiency.” There is also a paper on “Multi-Core DSP” and a poster on modeling “Power-Performance Efficiency” for GPUs.
A new topic for SC this year is a focus on memory technologies, which was presaged by a keynote at the International Supercomputing Conference held in Hamburg, Germany last June when Dr. Byungse So, Samsung Senior Vice President gave a talk on “Advanced Memory Technology – #1 Factor for Energy Efficient HPC”. Two papers, RAMZzz and Mage, both explore novel memory system designs. Samsung and Micron, respectively are presenting exhibitor forums on “How Memory and SSDs can Optimize Data Center Operations” and “Hybrid Memory Cube (HMC)”.
Whereas memory is on the uptake, the focus on liquid cooling has waned with only two sessions this year compared to six last year at SC’11. Eurotech will present an exhibitor forum on “Differences Between Cold and Hot Water Cooling on CPU and Hybrid Supercomputers” and Green Revolution Cooling will present on “100% Server Heat Recapture in Data Centers is Now a Reality.”
DATA CENTER SESSIONS
Kimberly Cupps, Lawrence Livermore National Laboratory will present on “The Sequoia System and Facilities Integration Story”. It appears that she will be giving the same talk at two different sessions; on Monday during Broader Engagement as well as on Tuesday as an Invited Speaker. Also, the M+W Group will present an exhibitor forum on “Reducing First Costs and Improving Future Flexibility in the Construction of High Performance Computing Facilities.”
APPLICATION TUNING AND JOB SCHEDULING
There are nine sessions that describe research on tuning applications for energy efficiency and various aspects of energy efficient job scheduling. Seven of the nine sessions are doctoral showcases, papers or posters. There is a BoF on “Power and Energy Measurement Modeling”. In this BoF, members of the research community and industry will present current state-of-the-art and limitations in measuring and modeling power and energy consumption and their effect on HPC application performance. An open discussion about future directions for such work will follow, with the intention of creating a “wish list” of feature requests to HPC vendors. Another BoF of interest is the SLURM User Group Meeting, which provides an open source job scheduler. Also, Charles Lively, ORNL will give a talk during Broader Engagement on “Heading Towards Exascale – Techniques to Improve Application Performance and Energy Consumption Using Application-Level Tools”.
Following is a list of the titles for the doctoral showcases:
- High Performance Non-Blocking and Power-Aware Collective Communication for Next Generation InfiniBand Clusters
- Total Energy Optimization for High Performance Computing Data Centers
Following is a list of the titles for the papers:
- T* – A Data-Centric Cooling Energy Costs Reduction Approach for Big Data Analytics Cloud
- NUMA-Aware Graph Mining Techniques for Performance and Energy Efficiency
Following is a list of the titles for the posters:
- Statistical Power and Energy Modeling of Multi-GPU kernels
- Virtual Machine Packing Algorithms for Lower Power Consumption
Two other sessions that will cover energy efficiency include an all day workshop on “High Performance Computing, Networking and Analytics for the Power Grid” and a poster on “Pay as You Go in the Cloud: One Watt at a Time.”
Although this is a list of sessions with a specific focus on energy efficiency, many more sessions will include energy efficiency as part of a broader focus.