Supercomputers are used to solve some of the world's most computationally demanding problems. Exascale systems, to be comprised of over one million cores and capable of $10^{18}$ floating point operations per second, will probably exist in the next five years, and will provide unprecedented computational power for parallel computing workloads. Unfortunately, while these machines hold tremendous promise and opportunity for applications in High Performance Computing (HPC), graph processing, and machine learning, it will be a major challenge to fully realize their potential, because to do so requires balanced execution across the entire system and its millions of processing elements. When different processors take different amounts of time to perform the same amount of work, performance imbalance arises, large portions of the system sit idle, and time and energy are wasted. Larger systems incorporate more processors and thus greater opportunity for imbalance to arise, as well as larger performance/energy penalties when it does.

In this talk, we explain how to design system software to mitigate variability on large scale parallel machines. Our approaches span (1) the design, implementation, and evaluation of a new high performance operating system to reduce some classes of performance variability, (2) a new performance evaluation framework to holistically characterize key features of variability on new and emerging architectures, and (3) a distributed modeling framework that derives predictions of how and where imbalance is manifesting in order to drive reactive operations such as load balancing and speed scaling. Collectively, these efforts provide a holistic set of tools to promote scalability through the mitigation of variability.