The increasing frequency of natural disasters has heightened the demand for UAV (Unmanned Aerial Vehicle) technologies. UAVs, especially drones, can monitor remote disaster areas and provide situational awareness to emergency responders. Equipped with cameras and onboard computers, drones can detect survivors in real-time, enhancing the efficiency of Search and Rescue (SAR) operations. Due to limited battery capacity, the drones must be deployed along a path of the shortest possible length to avoid delays in detecting the survivors in a given disaster area. Traditional path-planning algorithms struggle to address the dynamic conditions in disaster areas. We propose an adaptive drone path planning framework for real-time survivor detection in disaster areas to address this. This framework aims to improve survivor detection by guiding UAVs along routes with higher probabilities of the presence of survivors. Adopting a ”Learn-As-You-Go” strategy, it trains a Potential Survivor Location (PSL) prediction model to identify way-points for drone sorties. Next, it leverages a novel computationally efficient path planning approach called Prediction-Based Priority-Aware Path Planning (PB-PAPP) to navigate towards the identified PSLs. Also, we present a Weight Synthesis module that enhances the prediction quality over time by aggregating the weights of the models trained by the drones, allowing continuous adaptation in changing environments. Finally, we present a prototype lightweight decentralized machine learning system that combines the above modules to facilitate real-time survivor detection. Compared to existing algorithms, our framework demonstrates an 84-97% reduction in overhead for adaptive path-planning.

In federated learning, model aggregation is performed at a common server leading to a single point of failure, which aggravates with increasing number of clients causing large network delays that may slow down the entire system. In such cases, hierarchical aggregation is performed with edge devices deployed at different geographical locations. Typically, edge devices employ online learning algorithms for processing the incoming data stream and so, one-way aggregation strategies might end up overwriting the local model updates after the clients receive the aggregated model. Hence, we propose a two-way aggregation strategy wherein models are aggregated both on the client and server. Also, since the incoming distribution may be temporal in nature, we need to dynamically update the aggregation tree. We quantify the non IIDness among different geo-spatially distributed dataspouts without compromising data privacy. We propose a novel dynamic re-clustering algorithm which is triggered by the intracluster and inter-cluster divergence. We implement the above algorithm in an end-to-end FL system that can perform two-step aggregation for aggregation of parent-side and child-side models in the aggregation tree to preserve the model updates in asynchronous execution of the system. Using benchmark datasets, we demonstrate that our proposed framework achieves 95% accuracy, which we plan to further improve.

We consider use cases that necessitate the reliable transmission of critical communication to aid mobile users, such as providing emergency network services to survivors in disaster areas, supporting victims in a traffic accident, or monitoring public events occurring in remote places. In such cases, it has become a practice for network service providers to temporarily deploy a swarm of UAVs to complement the traditional unreliable mode of communication in such regions that may become intermittently unavailable. Given that the battery capacity limits the flight time of UAVs, it is important to design an optimal deployment plan such that ground devices in the area of interest (for example, users in disaster areas) are always connected to one or more UAVs, providing them with network connectivity. We consider that Battery Swapping Stations (BSS) are deployed in the area of interest (AoI) such that the UAVs can promptly replenish their battery to maintain uninterrupted services. Further, the relative position of the UAVs acting as network access points must be adjusted concerning the users’ current locations of the said network service to maintain uninterrupted connectivity to the User Equipments (UEs) in the given AoI. To that end, we propose a Deep Reinforcement Learning (DRL) algorithm to optimally adjust the positions of the UAVs concerning the UEs as well as the BSSs to maintain network connection despite dynamic changes in the position of UEs. Additionally, our algorithm also triggers the UAVs to swap their batteries whenever a certain optimal threshold for the available battery level has been reached. Since the UAVs are maneuvered from one optimal position to the next one based on the mobility of the UEs, there is an inherent dynamic sequential decision-making process involved in the given problem, making it a good fit for DRL. With extensive simulation experiments, we demonstrate that our approach outperforms the state-of-the-art in this field.

For real-time edge computing applications working under stringent deadlines, communication delay between IoT devices and edge devices needs to be minimized. Since the generalized assignment problem being NP-Hard, an optimal assignment of IoT devices to the edge cluster is hard. We propose the application RL based heuristics to obtain a near-optimal assignment of IoT devices to the edge cluster while ensuring that none of the edge devices are overloaded. We demonstrate that our algorithm outperforms the state-of-the-art.

Being comprised of resource-constrained edge devices, live migration is a necessary feature in edge clusters for migrating state of an entire machine in case of machine failures and network partitions without disrupting continued availability of services. While most of the prior work in this area has provided solutions for live migration on clusters comprised of resource-rich servers or fog servers with high computing power, there is a general lack of research in live migration on the low-end ARM devices comprised in edge clusters. To that end, we propose a lightweight algorithm for performing live migration on resource constrained edge clusters. We provide an open source implementation of the above algorithm for migrating containers in Linux. We demonstrate that our algorithm outperforms state-of-the-art live migration algorithms on resource constrained edge clusters with network partitions and device failures.

Despite the emergence of edge computing as a key technology paradigm, there is a general lack of auto-scaling techniques specifically designed for edge computing applications. Further, existing auto-scaling solutions tailor-made for the cloud cannot be readily applied to an application running on an edge cluster. In this paper we present AMAS - a novel auto-scaling algorithm, designed specifically for an edge cluster, which allows edge devices to be automatically and seamlessly added or deleted from an edge cluster according to dynamic variations in the workload. By design, AMAS can help users and enterprises minimize the cost of infrastructure while maintaining necessary SLO (i.e., Service Level Objective) deadlines for different edge applications. We demonstrate that AMAS outperforms the state-of-the-art auto-scalers in failure-prone conditions.

We present OptEx, a closed-form model of job execution on Apache Spark, a popular parallel processing engine. To the best of our knowledge, OptEx is the first work that analytically models job completion time on Spark. The model can be used to estimate the completion time of a given Spark job on a cloud, with respect to the size of the input dataset, the number of iterations, and the number of nodes comprising the underlying cluster. Experimental results demonstrate that OptEx yields a mean relative error of 6 percent in estimating the job completion time. Furthermore, the model can be applied for estimating the cost-optimal cluster composition for running a given Spark job on a cloud under a completion deadline specified in theSLO(i.e., Service Level Objective). We show experimentally that OptEx is able to correctly estimate the required cluster composition for running a given Spark job under a given SLO deadline with an accuracy of 98 percent. We also provide a tool which can classify Spark jobs into job categories based on bisimilarity analysis on lineage graphs collected from the given jobs.

We demonstrate Dyn-YCSB, a tool that builds upon YCSB (Yahoo Cloud Serving Benchmark suite) to assist users in simulating dynamic variations in workloads. Dyn-YCSB automatically varies the parameters in YCSB workloads over time according to user-specified time series functions, without requiring users to manually change the workload configuration in individual nodes each time the workload parameters needs to be modified. The dynamic workload variations simulated with Dyn-YCSB can be used to evaluate the adaptability of such frameworks to changing workload characteristics. We demonstrate the ability of Dyn-YCSB to evaluate the adaptability of OptCon, an automated framework, that tunes the consistency settings of Cassandra with respect to the latency and staleness thresholds in an SLA.

CRDTs are distributed data types that make eventual consistency of a distributed object possible and non ad-hoc. Geo-distributed systems are spread across multiple data centers at different geographic locations to ensure availability and performance despite network partitions. These systems must accept updates at any replica and propagate these updates asynchronously to every other replica. Conflict-Free Replicated Data Types (CRDTs) ensures eventual consistency in the replicas despite asynchronous delivery of updates. Extending this idea to fog computing servers where connection reliability is low, eventual consistency amongst the servers is required. We configure Kubernetes, an open-source container orchestration system used for automating deployment, scaling, and management of containerized applications, and use it for cluster deployment of CRDT based low resource intensive AntidoteDB can be used for deployment on fog servers to ensure eventual consistency amongst these servers. We have developed an automated benchmarking tool for benchmarking edge computing applications.

Recent storage systems trade strong consistency for performance, availability, and scalability. However, this makes it hard to understand the semantics that the storage system provides, and also makes the design and implementation of the storage system itself more error-prone. This paper proposes a comprehensive solution to these problems. In particular, we propose a specification language named ConSpec, which enables the formalization of different consistency semantics that a storage system may provide, using a uniform syntax that is independent of the design and implementation of the target storage system. We use ConSpec to revisit several existing models in light of a common way to define and compare them. Furthermore, we generalize the CAP theorem, whose original formulation only considered linearizability, to precisely define the class of consistency definitions that can and cannot be implemented in a highly-available, partition-tolerant way. Finally, we present the design and implementation of a new consistency checker that takes a trace from a storage system (e.g., the output of a test suite) and validates whether it meets any consistency semantics defined using ConSpec. The evaluation of our consistency checker shows that it is able to verify the correctness of long traces in a reasonable time.

Users of distributed datastores that employ quorum-based replication are burdened with the choice of a suitable client-centric consistency setting for each storage operation. The above matching choice is difficult to reason about as it requires deliberating about the tradeoff between the latency and staleness, i.e., how stale (old) the result is. The latency and staleness for a given operation depend on the client-centric consistency setting applied, as well as dynamic parameters such as the current workload and network condition. We present OptCon, a machine learning-based predictive framework, that can automate the choice of client-centric consistency setting under user-specified latency and staleness thresholds given in the service level agreement (SLA). Under a given SLA, OptCon predicts a client-centric consistency setting that is matching, i.e., it is weak enough to satisfy the latency threshold, while being strong enough to satisfy the staleness threshold. While manually tuned consistency settings remain fixed unless explicitly reconfigured, OptCon tunes consistency settings on a per-operation basis with respect to changing workload and network state. Using decision tree learning, OptCon yields 0.14 cross validation error in predicting matching consistency settings under latency and staleness thresholds given in the SLA. We demonstrate experimentally that OptCon is at least as effective as any manually chosen consistency settings in adapting to the SLA thresholds for different use cases. We also demonstrate that OptCon adapts to variations in workload, whereas a given manually chosen fixed consistency setting satisfies the SLA only for a characteristic workload.

We present OptEx, a closed-form model of job execution on Apache Spark, a popular parallel processing engine. To the best of our knowledge, OptEx is the first work that analytically models job completion time on Spark. The model can be used to estimate the completion time of a given Spark job on a cloud, with respect to the size of the input dataset, the number of iterations, the number of nodes comprising the underlying cluster. Experimental results demonstrate that OptEx yields a mean relative error of 6% in estimating the job completion time. Furthermore, the model can be applied for estimating the cost optimal cluster composition for running a given Spark job on a cloud under a completion deadline specified in the SLO (i.e.,Service Level Objective). We show experimentally that OptEx is able to correctly estimate the cost optimal cluster composition for running a given Spark job under an SLO deadline with an accuracy of 98%.

Users of distributed datastores that employquorum-based replication are burdened with the choice of asuitable client-centric consistency setting for each storage operation. The above matching choice is difficult to reason about asit requires deliberating about the tradeoff between the latencyand staleness, i.e., how stale (old) the result is. The latencyand staleness for a given operation depend on the client-centricconsistency setting applied, as well as dynamic parameters such asthe current workload and network condition. We present OptCon, a novel machine learning-based predictive framework, that canautomate the choice of client-centric consistency setting underuser-specified latency and staleness thresholds given in the servicelevel agreement (SLA). Under a given SLA, OptCon predictsa client-centric consistency setting that is matching, i.e., it isweak enough to satisfy the latency threshold, while being strongenough to satisfy the staleness threshold. While manually tunedconsistency settings remain fixed unless explicitly reconfigured, OptCon tunes consistency settings on a per-operation basis withrespect to changing workload and network state. Using decisiontree learning, OptCon yields 0.14 cross validation error in predictingmatching consistency settings under latency and stalenessthresholds given in the SLA. We demonstrate experimentally thatOptCon is at least as effective as any manually chosen consistencysettings in adapting to the SLA thresholds for different usecases. We also demonstrate that OptCon adapts to variationsin workload, whereas a given manually chosen fixed consistencysetting satisfies the SLA only for a characteristic workload.

We present a novel execution environment for multi-agent systems building on concepts from cloud computing and peer-to-peer networks. The novel environment can provide the computing power of a cloud for multi-agent systems in intermittently connected networks. We present the design and implementation of a prototype operating system for managing the environment. The operating system provides the user with a consistent view of a single machine, a single file system, and a unified programming model while providing elasticity and availability.