Traditionally, HPC systems have been designed with bulk synchronous parallelism in mind, where no dynamic features were welcome. However, processors are becoming more dynamic due to thermal issues. For example, modern processors take advantage of the thermal headroom to improve performance by opportunistically adjusting its voltage and frequency based on thermal and energy constraints (e.g., Intel's Turbo Boost technology). A question for us is how these dynamic features affect HPC applications and how to tackle with such dynamic behaviour. To answer the question, I evaluated the effectiveness of several strategies for reducing uneven heat accumulation and performance variation. In my talk I will go through: 1) an overview of how modern processors achieve power and performance proportionality through different mechanisms, and which interfaces are provided to system software to read and interact with the hardware, 2) the anatomy of the Linux kernel related to frequency/idle selection and thermal states, and 3) the effectiveness of various techniques to reduce thermal imbalance and the performance variation.

Bio:

Pablo Llopis Sanmillán is a Computer Science PhD student in the University Carlos III of Madrid, Spain. His research interests include system-level energy efficiency, particularly within the I/O stack, storage systems, and virtualization. He also enjoys constantly keeping up to date with new technology, and operates an OpenStack-based cloud in the university's computing center. During his PhD program he was awarded the 2013 IBM PhD Fellowship, interned at IBM Zurich Research Laboratory, and is currently a summer student visiting Argonne National Laboratory for the first time.