Recently I attended ISC 2016 in Frankfurt. Here are a few of the talks that grabbed my attention.
One of the highlights was the Workshop on Energy-Aware High Performance Computing (EnA-HPC). I particularly enjoyed Dimitrios S. Nikolopoulos’s keynote on exploiting approximate computing and resilience to reduce energy consumption. The talk focused on two approaches: firstly, balancing between accurate and approximate (more energy efficient) versions of functions to achieve a target accuracy — and secondly, investigating the extent to which RAM refresh rates could be lowered to save energy at the expense of accuracy.
Later in the workshop, Joseph Schuchart discussed a shift from power variation to performance variation on Haswell clusters and also showed a strong correlation between RAPL measurements and an AC reference. This was followed by a talk from Kashif Nizam Khan on estimating wall socket power consumption using RAPL data. On a related note, in the main conference, Andrea Borghesi presented his award winning paper on Predictive Modeling for Job Power Consumption in HPC Systems.
The conference featured several talks on auto-tuning, performance analysis, modeling and prediction. Highlights included Sandra Wienke’s presentation on auto-tuning for locality-aware loop scheduling, Boyana Norris on fine-level measurement of application arithmetic intensity and Torsten Hoefler on static and dynamic techniques for automatic performance modeling. I also attended a tutorial on performance analysis and auto-tuning tools including a brief introduction to the Periscope Tuning Framework.
I can’t write about ISC 2016 without mentioning the exciting announcement of a new number one supercomputer in the Top500, the Sunway TaihuLight. It has over 10 million cores and achieved 93 PFLOPS on HPL (nearly three times the performance of the now second place Tianhe-2). For more details on the new machine see Jack Dongarra’s report. The Green500 (which ranks the Top500 by performance-per-watt) was announced too with the current number one, Shoubu, achieving 6.67 GFLOPS/Watt.
All in all this was successful trip with many interesting talks for me to explore further.
Craig Blackmore is currently completing a Ph.D and with Embecosm over the summer to work on the TSERO project.