Evacuation during natural disasters is an important preventive measure. Previous works on evacuation traffic analysis and evacuation path planning problems are mainly based on small size road networks and human mobility datasets. In this paper, we use a large-scale human mobility dataset of the North Carolina (NC) state during hurricane Florence in 2018 in designing an emergency and safe evacuation path planning system (ESEP). ESEP consists of three main parts: mobility dataset analysis, latent variable neural network model based evacuation prediction model, and evacuation path planning system to evacuate people within the shortest possible time while ensuring maximum connectivity to safe cities along the routing path at minimum possible risk of being affected by the hurricane and also avoiding congestion. We additionally propose a map partitioning based method for dividing the problem into several smaller equivalent sub-problems to reduce the computation complexity. Our trace-driven evaluations show that the ESEP-suggested evacuation paths achieve better performance in terms of all design factors. The experimental result also shows that the map partitioning based method reduces the computation time of finding evacuation paths by 5 0% while maintaining the most objective values within 10% of the global optimal solution.

With the increased usage of edge devices having local computation capabilities, deep neural network (DNN) training in a network of edge devices becomes promising. Several recent works have proposed fully edge-based distributed training systems for situations when the communication to cloud is unstable or intermittent. However, such distributed systems become vulnerable when there are untrusted devices that launch data and model poisoning attacks during training, deteriorating the accuracy of the DNN model. To handle this challenge, we propose a Trustworthy distributed system for Machine learning training in an edge device network (TrustMe). TrustMe realizes both data and model parallelisms. It detects the untrusted devices producing illegitimate outputs. Next, it reassigns the training tasks of the untrusted devices to other trusted devices in such a way that the reassignment and the training that is restarted after the reassignment require minimal time. Our container-based emulation and real device experiments demonstrate that TrustMe achieves up to 12% higher accuracy and 45% less training time compared to existing methods in the presence of untrusted devices.

Epidemic simulation traditionally serves as one of the important methods to forecast how an epidemic may spread among a population. However, there are two key limitations that restrict the scope of such methods. The first limitation is that the existing tools rely on different sets of static parameters (e.g., infection probability, recovering probability) for simulating an epidemic spread that may fail to capture the dynamic nature of population interactions that acts as a dominant factor in an epidemic spread scenario such as COVID-19 pandemic. To handle this challenge, we propose a machine learning based model that combines a Graph Convolutional Neural Network (GCN) and a Recurrent Neural Network (RNN). It integrates the ability of the GCN to capture spatial dependency in human interaction and the ability of the RNN to incorporate temporal effects of the virus spread. The second limitation is that these methods do not address the computation overhead problem when dealing with time-dynamic graphs. Training a GCN on a very large graph suffers from the communication overhead from different graph partitions and the computation overheads stemming from partitioning dynamic graphs. This limitation impacts the scalability of the existing systems. To solve this challenge, we partition the graph in a computationally less expensive manner by partitioning the graph using the min-cut principle. We conducted comprehensive large scale real-world human mobility data driven experiments. Our experimental result shows that the proposed machine learning based forecasting model achieves overall 84% classification accuracy with greater than 72% precision and 62% recall. Also, the proposed graph partitioning approach reduces computation time and commutation overhead by a significant margin.

The COVID-19 pandemic is an ongoing pandemic of coronavirus disease since 2019. Millions of cases and deaths attributed to it have been confirmed in the world. So far the detection of COVID-19 heavily relies on the specialized tests (e.g., based on saliva or respiratory swabs). Some approaches use smart devices (e.g., Whoop) for coronavirus infection detection using respiratory rate. Machine learning (ML) techniques have become a promising approach for the coronavirus infection detection. Therefore, in this paper, we introduce a machine learning based COVID infection predictor. We measure the prediction accuracy of five ML models. We use Chi-square test and knowledge-based manual feature selection to select important features for prediction to reduce prediction time overhead without compromising prediction accuracy. We also study the accuracy with different input features (those that can be measured by medical devices and by smart devices) and find that removing some features has no or slight influence on the prediction accuracy. Since insufficient or unbalanced training data decreases the prediction accuracy, we further propose a Generative Adversarial Network (GAN) ML based predictor that produces synthetic data (close to real data) for ML training. Our extensive experiments show the effectiveness of our methods in improving the detection accuracy. Our study results can provide guidance on developing the coronavirus infection predictors based on different data sources and devices. We open sourced our code in GitHub.

A large-scale cloud data center needs to provide high service reliability and availability with low failure occurrence probability. However, current large-scale cloud data centers still face high failure rates due to many reasons such as hardware and software failures, which often result in task and job failures. Such failures can severely reduce the reliability of cloud services and also occupy huge amount of resources to recover the service from failures. Therefore, it is important to predict task or job failures before occurrence with high accuracy to avoid unexpected wastage. Many machine learning and deep learning based methods have been proposed for the task or job failure prediction by analyzing past system message logs and identifying the relationship between the data and the failures. In order to further improve the failure prediction accuracy of the previous machine learning and deep learning based methods, in this paper, we propose a failure prediction algorithm based on multi-layer Bidirectional Long Short Term Memory (Bi-LSTM) to identify task and job failures in the cloud. The goal of Bi-LSTM prediction algorithm is to predict whether the tasks and jobs are failed or completed. The trace-driven experiments show that our algorithm outperforms other state-of-art prediction methods with 93% accuracy and 87% for task failure and job failures respectively.

Image storage systems are designed to support images sharing and retrieving applications, especially for smartphone and other mobile devices. Existing image storage systems store images in random data servers without considering their similarities. When a front-end server sends out an image query, it receives similar images from a large number of data servers, causing possible network incast congestion and long query latency. To solve this problem, we propose CStorage, an efficient classification-based image storage system. In CStorage, similar images are stored in the same data server. Thus, a front-end server receives query results from a single data server, which reduces the occurrence of incast congestion as well as the image retrieval latency. CStorage also leverages the deep learning technique to cluster images and provides a high precision and recall rates. Experimental results show the effectiveness of CStorage in reducing image retrieval latency and improving precision and recall rates of searching results.

Nowadays, a large amount of services are deployed on the edge of the network from the cloud since processing data at the edge can reduce response time and lower bandwidth cost for applications such as healthcare in smart homes. Resource management is very important in the edge computing since it is able to increase the system efficiency and improve the quality of service. A common approach for resource management in edge computing is to assign tasks to the remote cloud or edge devices just according to several factors such as energy, bandwidth consumption, and latency. However, the approach is insufficiently efficient and falls short in meeting the requirements of handling health emergency when being applied in smart homes for healthcare. In this paper, we propose a task scheduling approach called HealthEdge that sets different processing priorities for different tasks based on the collected data on human health status and determines whether a task should run in a local device or a remote cloud in order to reduce its total processing time as much as possible. Based on a real trace from five patients, we conduct a trace-driven experiment to evaluate the performance of HealthEdge in comparison with other methods. The results show that HealthEdge can optimally assign tasks between the network edge and cloud, which can reduce the task processing time, reduce bandwidth consumption and increase local edge workstation utilization.

Demand-prediction based resource provisioning schemes help assure service level objectives (SLO) in cloud systems. We notice that if a provisioning scheme does not exclude bursts from historical resource demands in normal demand prediction or always uses a large padding to correct under-prediction, it will lead to resource over-provisioning and low resource utilization. To improve the previous schemes, in this paper, we present a Resource-efficient Predictive Resource Provisioning system in clouds (RPRP) that excludes bursts in demand prediction and has algorithms to specifically handle bursts to avoid resource over-provisioning. Rather than setting padding to a possibly high value, RPRP has a load-dependent padding algorithm that adaptively determines padding based on predicted demands. To handle bursts, RPRP embodies a responsive padding algorithm that adaptively adjusts padding to recover from both under-provisioning and over-provisioning. We implemented RPRP on top of Xen and conducted both trace-driven simulation and real-world testbed experiments. The experimental results show that RPRP achieves higher resource utilization, more accurate demand predictions, and fewer SLO violations than previous schemes.

Many high-performance computing (HPC) sites extend their clusters to support Hadoop MapReduce for a variety of applications. However, HPC cluster differs from Hadoop cluster on the configurations of storage resources. In the Hadoop Distributed File System (HDFS), data resides on the compute nodes, while in the HPC cluster, data is stored on separate nodes dedicated to storage. Dedicated storage offloads I/O load from the compute nodes and provides more powerful storage. Local storage provides better locality and avoids contention for shared storage resources. To gain an insight of the two platforms, in this paper, we investigate the performance and resource utilization of different types (i.e., I/O-intensive, data-intensive and CPU-intensive) of applications on the HPC-based Hadoop platforms with local storage and dedicated storage. We find that the I/O-intensive and data-intensive applications with large input file size can benefit more from the dedicated storage, while these applications with small input file size can benefit more from the local storage. CPU-intensive applications with a large number of small-size input files benefit more from the local storage, while these applications with large-size input files benefit approximately equally from the two platforms. We verify our findings by trace-driven experiments on different types of jobs from the Facebook synthesized trace. This work provides guidance on choosing the best platform to optimize the performance of different types of applications and reduce system overhead.

Cloud storage system usually experiences data loss, hindering data durability. Three-way random replication is commonly used to prevent data loss in cloud storage systems. However, it cannot effectively handle correlated machine failures. Although Copyset Replication and Tiered Replication can reduce data loss in correlated and independent failures and enhance data durability, they fail to leverage different data popularities to substantially reduce the storage cost and bandwidth cost caused by replication. To address these issues, we present a popularity-aware multi-failure resilient and cost-effective replication (PM-CR) scheme for high data durability in cloud storage. PMCR splits the cloud storage system into primary tier and backup tier, and classifies data into hot data, warm data and cold data based on data popularities. To handle both correlated and independent failures, PMCR stores the three replicas of the same data into one Copyset formed by two servers in the primary tier and one server in the backup tier. For the third replicas of warm data and cold data in the backup tier, PMCR uses the Similar Compression method for read-intensive data and uses the Delta Compression method for write-intensive data to reduce storage cost and bandwidth cost. As a result, these costs are reduced and data durability and availability are enhanced without compromising data request delay greatly. Extensive experiment results based on trace parameters show that PMCR achieves high data durability, low probability of data loss, and low storage cost and bandwidth cost compared to previous replication schemes.

In cloud systems, efficient resource provisioning is needed to maximize the resource utilization while reducing the Service Level Objective (SLO) violation rate, which is important to cloud providers for high profit. Several methods have been proposed to provide efficient provisioning. However, the previous methods do not consider leveraging the complementary of jobs' requirements on different resource types and job size concurrently to increase the resource utilization. Also, by simply packing complementary jobs without considering job size in the job packing, it can decrease the resource utilization. Therefore, in this paper, we consider both jobs' demands on different resource types (in the spatial space) and jobs' execution time (in the temporal space); we pack the complementary jobs (whose demands on multiple resource types are complementary to each other) belonging to the same type and assign them to a Virtual Machine (VM) to increase the resource utilization. Moreover, the previous methods do not provide efficient resource allocation for heterogeneous jobs in current cloud systems and do not offer different SLO degrees for different job types to achieve higher resource utilization and lower SLO violation rate. Therefore, we propose a Customized Cooperative Resource Provisioning (CCRP) scheme for the heterogeneous jobs in clouds. CCRP uses the hybrid resource allocation and provides SLO availability customization for different job types. To test the performance of CCRP, we compared CCRP with existing methods under various scenarios. Extensive experimental results based on a real cluster and Amazon EC2 show that CCRP achieves 50% higher or more resource utilization and 50% lower or less SLO violation rate compared to the previous resource provisioning strategies.

In large-scale computing clusters, when the server storing a task's input data does not have sufficient computing capacity, current job schedulers either schedule the task and transmit the input data to the closest server or let the task wait until the server has sufficient computing capacity, which generates network load or task delay. To handle this problem, load balancing methods are needed to reduce the number of overloaded servers due to computing workloads. However, current load balancing methods either do not consider the computing workload or assume that it is proportional to the number of data blocks in a server. Through trace analysis, we demonstrate the diversity of computing workloads of different tasks and the necessity of balancing the computing workloads among servers. Then, we propose a cost-efficient Computing load Aware and Long-View load balancing approach (CALV). In addition to the computing load awareness, CALV is also novel in that it achieves long-term load balance by migrating out data blocks from an overloaded server that contribute more computing workloads when the server is more overloaded and contribute less computing workloads when the server is more underloaded at different epochs during a time period. CALV also has a lazy data block transmission method to improve the load balanced state and avoid network load peak. Trace-driven experiments in simulation and a real computing cluster show that CALV outperforms other methods in terms of balancing the computing workloads and cost efficiency.

Functioning as an intermediary between tenants and cloud providers, cloud service brokerages (CSBs) can bring about great benefits to the cloud market. CSBs buy the cloud resources, i.e., servers, with lower prices from cloud providers and sell the resources to the tenants with higher prices. To maximize its own profit, a CSB may distribute tenants' requests to the clouds that waste energy resources. However, as energy costs of cloud computing have been increasing rapidly, there is a need for cloud providers to optimize energy efficiency while maintain high service level performance to tenants, not only for their own benefit but also for social welfares (e.g., protecting environment). Thus, for green cloud companies, two questions have arisen: 1) under what pricing policies from the cloud providers to the CSB, a profit-driven CSB is willing to minimize the total cloud energy cost while satisfy tenant demands and 2) how should a CSB distribute tenants' demands to achieve this objective? To address question 1), we find a pricing policy for cloud providers such that maximizing CSB's profit is equivalent to minimizing cloud providers' energy cost. To address question 2), we first devise a greedy solution, and then propose an approximation algorithm with a constant approximation ratio. Both simulation and real-world Amazon EC2 experimental results demonstrate the effectiveness of our pricing policy to incentivize CSBs to save energy for cloud providers and the superior performance of our algorithms in energy efficiency and resource utilizations in comparison with the previous algorithms.

Question and Answer (Q&A) websites such as Yahoo!Answers provide a platform where users can post questions and receive answers. These systems take advantage of the collective intelligence of users to find information. In this paper, we analyze the online social network (OSN) in Yahoo!Answers. Based on a large amount of our collected data, we studied the OSN's structural properties, which reveals strikingly distinct properties such as low link symmetry and weak correlation between indegree and outdegree. After studying the knowledge base and behaviors of the users, we find that a small number of top contributors answer most of the questions in the system. Also, each top contributor focuses on only a few knowledge categories. In addition, the knowledge categories of the users are highly clustered. We also study the knowledge base in a user's social network, which reveals that the members in a user's social network share only a few knowledge categories. Based on the findings, we provide guidance in the design of spammer detection algorithms and distributed Q&A systems. We also propose a friendship-knowledge oriented Q&A framework that synergically combines current OSN-based Q&A and web Q&A. We believe that the results presented in this paper are crucial in understanding the collective intelligence in the web Q&A OSNs and lay a cornerstone for the evolution of next-generation Q&A systems.