“If you can’t measure it, you can’t improve it” -- Software Improvements from Power/Energy Measurement Capabilities

Moderator: Daniel Reed (University of Iowa)

Panelists: Satoshi Matsuoka (Tokyo Institute of Technology, RIKEN), Sadaf Alam (Swiss National Supercomputing Centre), Jack Deslippe (Lawrence Berkeley National Laboratory), Tapasya Patki (Lawrence Livermore National Laboratory)

Abstract: "If you can't measure it, you can't improve it." We now have some sites that are extremely well instrumented for measuring power and energy. Lawrence Berkeley National Laboratory’s supercomputing center is a prime example, with sensors, data collection and analysis capabilities that span the facility and computing equipment. We have made major gains in improving the energy efficiency of the facility as well as computing hardware, but there are still large gains to be had with software- particularly application software. Just tuning code for performance isn’t enough; the same time to solution can have very different power profiles. We are at the point where measurement capabilities are allowing us to see cross-cutting issues- such as the cost of spin waits. These new measurement capabilities should provide a wealth of information to see the tall poles in the tent. This panel will explore how we can identify these emerging tall poles.

Presentation: File
Presentation: File
Presentation: File
Presentation: File


Back to the Panel Archive Listing