Edge devices are embedded sensing or actuation devices accessible via wireless sensor networks in applications such as monitoring of structural health or environmental conditions in buildings. To avoid retrofitting costs and ease the deployment, these devices are often battery-powered, thus requiring manual battery replacement to maintain their operations over time. Yet, as the sensor network scales up to thousands of sensors, maintenance becomes a time consuming and labor-expensive task. Energy harvesting is often used to extend the lifetime of sensor nodes and avoid battery replacement.

In this dissertation, we present techniques that combine hardware, software, and artificial intelligence techniques to extend the lifetime of the sensor nodes devices for decades without sacrificing application performance even in low energy availability environments.

As a hardware solution, we present a sensing platform that can be deployed anywhere inside a building and monitor a wide range of parameters without needing periodic battery replacement that is typical of current solutions. Instead of a rechargeable battery, it uses a supercapacitor to store the energy harvested from the environment. To facilitate deployment and integration with existing buildings, the platform uses Bluetooth Low Energy (BLE) to relay data.

Since the amount of energy harvested can change between sensor node locations, and the applications can have different energy requirements over time, we present a learning-based method to extend the operating lifetime of network-connected edge devices while increasing the application performance with available energy. We describe design choices that enable an indoor environment sensing device to exploit reinforcement learning for periodic and event-driven sensing with ambient light energy harvesting. Using simulations and real deployments, we show that our sensor nodes adapt to ambient lighting conditions and send measurements and events continuously during nights and weekends without interruptions. We use real-world deployment data to continually adapt sensing to changing environmental patterns and use transfer learning to reduce the training time.

To be effective these techniques require prior knowledge of the environment in which the sensor nodes are deployed. In the absence of historical data, the application performance deteriorates. To address this problem, we present an approach that leverages meta reinforcement learning to increase the application performance of newly deployed batteryless sensor nodes without historical data. Our method exploits information from other sensor node locations to expedite the learning of newly deployed sensor nodes and improves the application performance after a few days of deployment.

Timed Decision Tables (TDTs) have been used earlier for modeling behavioral descriptions, applying presynthesis optimizations for efficient circuit synthesis and HDL restructuring. We describe here work that optimizes TDT models for generation of software in a high-level programming language. The optimization for software synthesis is targeted at reducing the numbers of conditionals and actions in the generated code. The TDT-based optimizations and software synthesis is implemented in C++. Experimental results on a set of examples show significant reduction in the number of conditionals checked.

Pre-2018 CSE ID: CS2008-0932

Proliferation of high resolution cameras on embedded devices along with the growing maturity of deep neural networks (DNNs) have enabled powerful mobile vision applications. To support these applications, resource-constrained mobile devices can be extended with server-class processors and GPU accelerators via mobile offloading techniques. However, this is challenging in latency-constrained applications because of large and unpredictable round-trip delays from mobile devices to the cloud computing resources. As a consequence, system designers routinely look for ways to offload to local servers at the network edge, known as the cloudlet.

This dissertation explores the potential of building emerging continuous mobile vision applications using the cloudlet. We identify two challenges in implementing a multi-tenant and real-time mobile offloading framework. First, an application may exploit DNNs for diverse tasks, e.g., object classification, detection, and tracking. We need methods to reduce delay and to improve throughput in DNN inference. Second, when the system hosts multiple clients and various applications concurrently, the contention on computing resources, e.g., CPUs and GPUs, leads to longer delays. The scheduling techniques that mitigate resource contention are thus essential for a low-latency serving system.

To address these challenges, we design new APIs, systems, and schedulers that enable a high-performance mobile offloading framework. The framework manages data as in-memory key-value pairs that can be cached on servers to avoid redundant data transfers. An application is programmed as a Directed Acyclic Graph (DAG) of queries that process mobile data. We investigate the infrastructure optimizations, such as batching and parallelization of DNN inference, for a variety of vision tasks, e.g., object detection, tracking, scene graph detection, and video description. To improve the level of resource utilization, our system co-locates delay-critical and delay-tolerant workloads on shared GPUs, using a new predictive approach. Adaptive batching algorithms that consider both accuracy and delay of DNN inference in a heterogeneous cluster of GPUs are studied as well. For CPU workloads, we present methods to mitigate resource contention by predicting future workloads, estimating contention, and adjusting task start times to remove the contention. We demonstrate the effectiveness of the framework through several evaluations with real world applications.

Pre-2018 CSE ID: CS2007-0881

Pre-2018 CSE ID: CS2006-0856

Pre-2018 CSE ID: CS2011-0960

Microelectronic scaling has entered into the nanoscale era with tremendous capacity and performance advantages that continue to drive new devices and systems from high-performance computing to ultra-low power Internet endpoints. This scaling, however, faces challenges due to serious effects of microelectronic variability that results in significant variation in individual device parameters. The most common manifestation of this variability is increased susceptibility to timing errors. Combating these errors usually results in increased guardbands in the circuit and architectural design, thus reducing the gains from process technology advances.

This dissertation focuses on methods to improve the timing error resilience of microelectronic computing systems by reducing the guardbands which also results in improved operational efficiency if microelectronic circuits. Timing errors can show at various abstraction levels --- the circuit layer, the architecture layer, and the software layer. Accordingly, we have proposed error tolerance methods that correspond to the layer where such errors manifest. Considering the interdependence of overall system performance or application quality on the design choices made at various abstraction levels, an integrated view of the overall effects of error tolerance strategies is necessary to evaluate the effects of these approaches in the system or application layer. Cross-layer optimizations are thus important in addressing the effects of timing errors. At the circuit layer, we examine the root cause of ``timing error'' via analysis of dynamic path sensitization of the circuit. We use machine learning methods to build a prediction model for the timing errors based on the useful features extracted from computation history, circuit workload, and circuit switching. Results show high prediction accuracy and fast computing performance that make the model useful in early circuit reliability evaluation. Second, at the architecture layer, by characterizing delay of various instructions, we dynamically adjust the clock frequency, that reduces timing errors and improve the operational efficiency. Finally, at the software layer, by utilizing the inherent ``error-tolerance'' of emerging applications such as neural networks, we reduce design margins under the premise that the application quality is acceptable. Specifically, we have investigated the vulnerability of emerging neural networks to timing errors and deliver an approximate computing hardware for neural networks that achieves significant energy savings with negligible accuracy loss. Stemming from this dissertation, our future research concerns building emerging high-performance, low-power, reliable, and intelligent non-conventional computing systems.

Frequency scaling of a processor based on slowdown factor, gives considerable power savings. The slowdown factors are computed from the task set executing on the processor. These factors are non-trivial for periodic tasks with deadlines not equal to their period. This paper describes how to compute static slowdown factors for a task set with an underlying earliest dealine first (EDF) schedule. We give a density slowdown method, which eficiently computes a slowdown factor. The constant slowdown factor not being optimal, we compute the optimal slowdown factors by looking at all the jobs up to the hyper-period of the task set. This is based on Yao's optimal algorithm for non-periodic tasks. Dynamic slowdown techniques further enhance the power savings. We show that dynamic slowdown factors can be easily incorporated over these static slowdown factors, thereby increasing the total energy savings.

Pre-2018 CSE ID: CS2007-0911