Date of Award

Fall 2017

Document Type


Degree Name

Doctor of Philosophy (PhD)


Computational Modeling & Simulation Engineering


Modeling and Simulation

Committee Director

Masha Sosonkina

Committee Member

Yuzhong Shen

Committee Member

Duc Nguyen

Committee Member

Manwo Ng

Committee Member

Tal Ezer


Achieving Exascale computing is one of the current leading challenges in High Performance Computing (HPC). Obtaining this next level of performance will allow more complex simulations to be run on larger datasets and offer researchers better tools for data processing and analysis. In the dawn of Big Data, the need for supercomputers will only increase. However, these systems are costly to maintain because power is expensive. Thus, a better understanding of power and energy consumption is required such that future hardware can benefit.

Available power models accurately capture the relationship to the number of cores and clock-rate, however the relationship between workload and power is less understood. Thus, investigation and analysis of power measurements has been a focal point in this work with the aim to improve the general understanding of energy consumption in the context of HPC.

This dissertation investigates power and energy consumption of many different parallel applications on several hardware platforms while varying a number of execution characteristics. Multicore and manycore hardware devices are investigated in homogeneous and heterogeneous computing environments. Further, common techniques for reducing power and energy consumption are employed to each of these devices.

Well-known power and performance models have been combined to form the Execution-Phase model, which may be used to quantify energy contributions based on execution phase and has been used to predict energy consumption to within 10%. However, due to limitations in the measurement procedure, a less intrusive approach is required.

The Empirical Mode Decomposition (EMD) and Hilbert-Huang Transform analysis technique has been applied in innovative ways to model, analyze, and visualize power and energy measurements. EMD is widely used in other research areas, including earthquake, brain-wave, speech recognition, and sea-level rise analysis and this is the first it has been applied to power traces to analyze the complex interactions occurring within HPC systems.

Probability distributions may be used to represent power and energy traces, thereby providing an alternative means of predicting energy consumption while retaining the fact that power is not constant over time. Further, these distributions may be used to define the cost of a workload for a given computing platform.