Dynamic parameters tuning for HPC clusters exploitation

Abstract

The High Performance Computing community faces a problem of enormous power and energy consumption, which is the major obstacle in building supercomputers exceeding exaflop performance. The only solution is a hardware-software codesign, introducing more power-efficient hardware with power management and monitoring features, as well as a software stack for power- and energy-efficient computing.

Over the last years, more and more heterogeneous hardware has been accommodated to meet power-related goals. However, it makes software development complicated. Implementing software that fully utilizes the available hardware is not easy. Underutilized hardware creates a space for optimization from an energy consumption point of view using a power knob management and, therefore, energy savings without application performance degradation.

This thesis presents a methodology of Fine-grain dynamic tuning of power knobs during a parallel application execution on heterogeneous hardware to achieve energy savings. Thanks to complex execution time coverage by regions of interest, high tuning granularity starting at the level of ten milliseconds and a large set of controlled power knobs, it pushes the achievable energy savings to the limit. The automatic identification of the optimal configuration is designed to control application performance degradation caused by the tuning.

The developed MERIC runtime system and related tools implement this methodology, providing support to tune various tuning parameters, automatic optimal configuration identification, resource consumption measurement, and measurement data visualization for application behaviour understanding.