University of Houston

In this work, we use energy instrumentation to study the health and workloads of a computing system. Analysis of energy consumption with the goal of understanding the computing system in an uncontrolled environment is an open research area. The main challenge is to infer the system state only from discrete time series energy data.

We have analyzed power consumption data on computing systems. Our focus is on how to distinguish various events and how to reveal the health condition of the system. In addition to studying the data collected in a laboratory environment, we have analyzed a 3-year continuous energy measurements of a large enterprise computing environment. We can infer system health, failures, activities, and trends from energy data.

We have investigated power consumption data of networking systems, especially the low power wireless networks. We designed two novel features called Radio Awake Length Counter and Radio Awake Overlap Counter. We evaluated our approaches on three real-world testbeds andvarious network scenarios. We found that these features reveal network protocols, application workloads, and routing topology from energy data alone. This was not possible only from energy data prior to this work.

The contributions of this work are: (a) Techniques to analyze and reveal health information of computing system. The energy profiling during boot up, idle and failure exposes operating states of the system. (b) Design of two novel features that use fine-grained energy instrumentation data on networking systems, to identify routing protocol, infer network topology, and determine application workloads. Our proposed features can achieve 97% accuracy when used to identify the routing protocols, and infer the network topology with 98% accuracy. (c) Identification of sources of waste in computing systems. At least 60% of energy consumed per day was wasted when the collection of computers we studied were left in idle state in a computer lab environment.