Search for: All records

Cold start delays are a main pain point for today’s FaaS (Function-as-a-Service) platforms. A widely used mitigation strategy is keeping recently invoked function containers alive in memory to enable warm starts with minimal overhead. This paper identifies new challenges that state-of-the-art FaaS keep-alive policies neglect. These challenges are caused by concurrent function invocations, a common FaaS workload behavior. First, concurrent requests present a trade-off between reusing busy containers (delayed warm starts) versus cold-starting containers. Second, concurrent requests cause imbalanced evictions of containers that will be reused shortly thereafter. To tackle the challenges, we propose a novel serverless function container orchestration algorithm called CIDRE. CIDRE makes informed decisions to speculatively choose between a delayed warm start and a cold start under concurrency-driven function scaling. CIDRE uses both fine-grained container-level and coarse-grained concurrency information to make balanced eviction decisions. We evaluate CIDRE extensively using two production FaaS workloads. Results show that CIDRE reduces the cold start ratio and the average invocation overhead by up to 75.1% and 39.3% compared to state-of-the-art function keep-alive policies. 

Real-time cyber-physical systems (CPS) rely on Perception-Cognition-Actuation (PCA) pipelines to enable autonomous observation, decisionmaking, and action execution. Closed-loop PCA systems utilize feedback-driven control to iteratively adapt actions in response to real-time environmental changes whereas open-loop PCA systems execute single actions without iterative feedback. The overall performance of these systems is inherently tied to the models selected for each pipeline component. Recent advancements in neural networks, particularly for perception tasks, have substantially enhanced CPS capabilities but have introduced significant complexity into the PCA pipeline. While traditional research [1] often evaluates perception models in static, controlled settings, it fails to account for the cascading latency and accuracy trade-offs that manifest across interconnected PCA modules in dynamic, real-time applications. Additionally, the proliferation of distributed device-edge-cloud architectures [2] has expanded computational possibilities but introduced new challenges in balancing latency and accuracy with resource constraints. The holistic impact of model selection, deployment platforms, and network conditions on application performance in real-time scenarios remains under-explored. 

The impact of human activity on the climate is a major global challenge that affects human well-being. Buildings are a major source of energy consumption and carbon emissions worldwide, especially in advanced economies such as the United States. As a result, making grids and buildings sustainable by reducing their carbon emissions is emerging as an important step toward societal decarbonization and improving overall human well-being. While prior work on demand response methods in power grids and buildings has targeted peak shaving and price arbitrage in response to price signals, it has not explicitly targeted carbon emission reductions. In this paper, we analyze the flexibility of building loads to quantify the upper limit on their potential to reduce carbon emissions, assuming perfect knowledge of future demand and carbon intensity. Our analysis leverages real-world demand patterns from 1000+ buildings and carbon-intensity traces from multiple regions. It shows that by manipulating the demand patterns of electric vehicles, heating, ventilation, and cooling (HVAC) systems, and battery storage, we can reduce carbon emissions by 26.93% on average and by 54.90% at maximum. Our work advances the understanding of sustainable infrastructure by highlighting the potential for infrastructure design and interventions to significantly reduce carbon footprints, benefiting human well-being. 

Existing approaches for autonomous control of pan-tilt-zoom (PTZ) cameras use multiple stages where object detection and localization are performed separately from the control of the PTZ mechanisms. These approaches require manual labels and suffer from performance bottlenecks due to error propagation across the multi-stage flow of information. The large size of object detection neural networks also makes prior solutions infeasible for real-time deployment in resource-constrained devices. We present an end-to-end deep reinforcement learning (RL) solution called Eagle1 to train a neural network policy that directly takes images as input to control the PTZ camera. Training reinforcement learning is cumbersome in the real world due to labeling effort, runtime environment stochasticity, and fragile experimental setups. We introduce a photo-realistic simulation framework for training and evaluation of PTZ camera control policies. Eagle achieves superior camera control performance by maintaining the object of interest close to the center of captured images at high resolution and has up to 17% more tracking duration than the state-of-the-art. Eagle policies are lightweight (90x fewer parameters than Yolo5s) and can run on embedded camera platforms such as Raspberry PI (33 FPS) and Jetson Nano (38 FPS), facilitating real-time PTZ tracking for resource-constrained environments. With domain randomization, Eagle policies trained in our simulator can be transferred directly to real-world scenarios2. 

Edge devices rely extensively on machine learning for intelligent inferences and pattern matching. However, edge devices use a multitude of sensing modalities and are exposed to wide ranging contexts. It is difficult to develop separate machine learning models for each scenario as manual labeling is not scalable. To reduce the amount of labeled data and to speed up the training process, we propose to transfer knowledge between edge devices by using unlabeled data. Our approach, called RecycleML, uses cross modal transfer to accelerate the learning of edge devices across different sensing modalities. Using human activity recognition as a case study, over our collected CMActivity dataset, we observe that RecycleML reduces the amount of required labeled data by at least 90% and speeds up the training process by up to 50 times in comparison to training the edge device from scratch. 

User interaction is an essential part of many mobile devices such as smartphones and wrist bands. Only by interacting with the user can these devices deliver services, enable proper configurations, and learn user preferences. Push notifications are the primary method used to attract user attention in modern devices. However, these notifications can be ineffective and even irritating if they prompt the user at an inappropriate time. The discontent is exacerbated by the large number of applications that target limited user attention. We propose a reinforcement learning-based personalization technique, called Nurture, which automatically identifies the appropriate time to send notifications for a given user context. Through simulations with the crowd-sourcing platform Amazon Mechanical Turk, we show that our approach successfully learns user preferences and significantly improves the rate of notification responses. 

Voice controlled interactive smart speakers, such as Google Home, Amazon Echo, and Apple HomePod are becoming commonplace in today's homes. These devices listen continually for the user commands, that are triggered by special keywords, such as "Alexa" and "Hey Siri". Recent research has shown that these devices are vulnerable to attacks through malicious voice commands from nearby devices. The commands can be sent easily during unoccupied periods, so that the user may be unaware of such attacks. We present EchoSafe, a user-friendly sonar-based defense against these attacks. When the user sends a critical command to the smart speaker, EchoSafe sends an audio pulse followed by post processing to determine if the user is present in the room. We can detect the user's presence during critical commands with 93.13% accuracy, and our solution can be extended to defend against other attack scenarios, as well. 

Mobile technologies that drive just-in-time ecological momentary assessments and interventions provide an unprecedented view into user behaviors and opportunities to manage chronic conditions. The success of these methods rely on engaging the user at the appropriate moment, so as to maximize questionnaire and task completion rates. However, mobile operating systems provide little support to precisely specify the contextual conditions in which to notify and engage the user, and study designers often lack the expertise to build context-aware software themselves. To address this problem, we have developed Emu, a framework that eases the development of context-aware study applications by providing a concise and powerful interface for specifying temporal- and contextual-constraints for task notifications. In this paper we present the design of the Emu API and demonstrate its use in capturing a range of scenarios common to smartphone-based study applications.