Michelle M. Zhu

Enhanced First-Fit Decreasing Algorithm for Energy-Aware Job Scheduling in Cloud

With the emerging of many data centers around the globe, heavy loads of large-scale commercial and scientific applications executed in the cloud call for efficient cloud resource management strategies to save energy without compromising the performance and system throughput. According to the statistics from the Data Centre Dynamic (DCD) organization, the expected energy consumption by computer servers would increase by 19% in 2013 compared with the previous year. Such trend may continue for many years. Moreover, the estimated energy consumption of computers in the U.S. was about 2% out of the total electricity consumption in 2010, which makes IT industry the second pollution contributor after aviation. In this paper, a novel approach for scheduling, sharing and migrating Virtual Machines (VMs) for a bag of cloud tasks is designed and developed to reduce energy consumption with guaranteed certain execution time and high system throughput. This approach is derived from an Enhanced First Fit Decreasing (EFFD) algorithm combined with our VM reuse strategy. Furthermore, virtual machine migration method is introduced to dynamically monitor the cloud situation for necessary migration. Our simulation results using Cloud Report show that EFFD with our VM reuse strategy gains higher resource utilization rate and lower energy consumption than Greedy, Round Robin (RR) and FDD without VM reuse.

Energy-Aware Workflow Job Scheduling for Green Clouds

With the increasing deployment of many data centers and computer servers around the globe, the energy cost on running the computing, communication and cooling together with the amount of CO2 emissions have increased dramatically. In order to maintain sustainable Cloud computing with ever-increasing problem scale, we design and develop energy-aware scientific workflow scheduling algorithm to minimize energy consumption and CO2 emission without sacrificing Quality of Service (QoS) such as response time specified in Service Level Agreement (SLA). The underlying available computing capacity and network bandwidth is represented as time-dependent because of the dual operation modes of on-demand and reservation instances supported by many commercial Cloud data centers. The Dynamic Voltage and Frequency Scaling (DVFS) is utilized to lower the CPU frequencies of virtual machines as long as the finishing time is still before the specified deadline. Our resource provision and allocation algorithm aims to meet the response time requirement and minimize the Virtual Machine (VM) overhead for reduced energy consumption. The consolidated VM reuse can lead to higher resource utilization rate for higher system throughput. The effectiveness of our algorithm is evaluated under various performance metrics and experimental scenarios using software adapted from open source CloudSim simulator. The simulation results show that our algorithm is able to achieve an average up to 30% of energy savings.

Concurrent Bandwidth Reservation Strategies for Big Data Transfers in High-Performance Networks

Because of the deployment of large-scale experimental and computational scientific applications, big data is being generated on a daily basis. Such large volumes of data usually need to be transferred from the data generating center to remotely located scientific sites for collaborative data analysis in a timely manner. Bandwidth reservation along paths provisioned by dedicated high-performance networks (HPNs) has proved to be a fast, reliable, and predictable way to satisfy the transfer requirements of massive time-sensitive data. In this paper, we study the problem of scheduling multiple bandwidth reservation requests (BRRs) concurrently within an HPN while achieving their best average transfer performance. Two common data transfer performance parameters are considered: the Earliest Completion Time (ECT) and the Shortest Duration (SD). Since not all BRRs in one batch can oftentimes be successfully scheduled, the problem of scheduling all BRRs in one batch while achieving their best average ECT and SD are converted into the problem of scheduling as many BRRs as possible while achieving the average ECT and SD of scheduled BRRs, respectively. The aforementioned two problems are proved to be NP-complete problems. Two fast and efficient heuristic algorithms with polynomial-time complexity are proposed. Extensive simulation experiments are conducted to compare their performance with two proposed naive algorithms in various performance metrics. Performance superiority of these two fast and efficient algorithms is verified.

An Innovative Energy-Aware Cloud Task Scheduling Framework

With the increased popularity of cloud computing, the number and scales of cloud data centers have kept growing at unprecedented speeds. In the meanwhile, the energy consumption by the data centers has kept commensurately increasing as well. Therefore, the focus of cloud resource management and scheduling has relatively shifted from mere performance to also energy efficiency. In this paper, we present a novel, Energy-Aware Task Scheduling framework that makes integrated exploitation of the two well-known energy saving techniques, DVFS and VM Reuse, on cloud task scheduling in a data center. We present our scheduling approach and framework via a specific algorithm, called EATS-FFD, that assumes FFD as its base scheduling policy. With minor modification, the presented framework can be made to work with a different base scheduling policy, resulting in a correspondingly different scheduling algorithm. Our approach achieves better energy-efficiency without sacrificing system QoS. The effectiveness of our approach is evaluated under various experimental scenarios using the Cloud Report tool running on the open source CloudSim platform.

Energy-Efficient Resource Management for Scientific Workflows in Clouds

The elastic resource provision, non-interfering resource sharing and flexible customized configuration provided by the Cloud infrastructure has shed light on efficient execution of many scientific applications. Due to the increasing deployment of data centers and computer servers around the globe escalated by the higher electricity price, the energy cost on running the computing, communication and cooling together with the amount of CO2 emissions have skyrocketed. In order to maintain sustainable Cloud computing facing with ever-increasing problem complexity and big data size in the next decades, we design and develop energy-aware scientific workflow scheduling algorithm to minimize energy consumption and CO2 emission while still satisfying certain Quality of Service (QoS) such as response time specified in Service Level Agreement (SLA). We also apply Dynamic Voltage and Frequency Scaling (DVFS) and DNS scheme to further reduce energy consumption within acceptable performance bounds. Our multiple-step resource provision and allocation algorithm achieves the response time requirement in the step of forwarding task scheduling and minimizes the VM overhead for reduced energy consumption and higher resource utilization rate in the backward task scheduling step. The effectiveness of our algorithm is evaluated under various performance metrics and experimental scenarios using software adapted from open source CloudSim simulator.

Enhanced Weighted Round Robin (EWRR) with DVFS Technology in Cloud Energy-Aware

In recent years, the rapid evolving Cloud Computing technologies multiply challenges including minimizing power consumption and meeting Quality-of- Services (QoS) requirements in the presence of heavy workloads from a large number of users using shared computing resources. Powering a middle-sized data center normally consumed 80,000kW power every year and computer servers consume around 5% of the global power [1]. In order to address the skyrocket energy cost from the high level resource management aspect, we propose an energy efficient job scheduling approach based on a modified version of Weighted Round Robin scheduler that incorporates VMs reuse and live VM migration without compromising the Service Level Agreement (SLA). Enhanced Weighted Round Robin (EWRR) algorithm enhanced scheduler can monitor the running VMs status for possible VM consolidation or Migration. In addition, VM Manager observes the VMs utilization rate to start live migration from the over-utilizing Processing Element (PE) to under-utilized PEs or to the hibernated PEs by sending WOL (Wake-On-LAN) signal to activate them. Moreover, we have integrated our Dynamic Voltage and Frequency Scaling (DVFS) algorithm in CPU utilization model to specify the required frequency for each task depending on the task complexity and the deadline.

High-Throughput Scientific Workflow Scheduling under Deadline Constraint in Clouds

Cloud computing is a paradigm shift in service delivery that promises a leap in efficiency and flexibility in using computing resources. As cloud infrastructures are widely deployed around the globe, many data- and computeintensive scientific workflows have been moved from traditional high- performance computing platforms and grids to clouds. With the rapidly increasing number of cloud users in various science domains, it has become a critical task for the cloud service provider to perform efficient job scheduling while still guaranteeing the workflow completion time as specified in the Service Level Agreement (SLA). Based on practical models for cloud utilization, we formulate a delay-constrained workflow optimization problem to maximize resource utilization for high system throughput and propose a two-step scheduling algorithm to minimize the cloud overhead under a user-specified execution time bound. Extensive simulation results illustrate that the proposed algorithm achieves lower computing overhead or higher resource utilization than existing methods under the execution time bound, and also significantly reduces the total workflow execution time by strategically selecting appropriate mapping nodes for prioritized modules. Index Terms— computing

An ELU Network with Total Variation for Image Denoising

In this paper, we propose a novel convolutional neural network (CNN) for image denoising, which uses exponential linear unit (ELU) as the activation function. We investigate the suitability by analyzing ELUs connection with trainable nonlinear reaction diffusion model (TNRD) and residual denoising. On the other hand, batch normalization (BN) is indispensable for residual denoising and convergence purpose. However, direct stacking of BN and ELU degrades the performance of CNN. To mitigate this issue, we design an innovative combination of activation layer and normalization layer to exploit and leverage the ELU network, and discuss the corresponding rationale. Moreover, inspired by the fact that minimizing total variation (TV) can be applied to image denoising, we propose a TV regularized L2 loss to evaluate the training effect during the iterations. Finally, we conduct extensive experiments, showing that our model outperforms some recent and popular approaches on Gaussian denoising with specific or randomized noise levels for both gray and color images.

Improved Scheduling Algorithms for Single-Path Multiple Bandwidth Reservation Requests

Colossal amounts of data are being generated in extreme-scale e-Sciences with the advent of new computation tools and experimental infrastructures. Such extremely large and complex data sets normally need to be transferred remotely for data storage and analysis. Reserving bandwidth as needed along selected paths in high-performance networks (HPNs) has proved to be an effective way to satisfy the high-demanding requirements of such data transfer. The most common data transfer requirement from users is the data transfer deadline. However, users oftentimes want to achieve other data transfer performance parameters, such as the earliest completion time (ECT) and the shortest duration (SD). For the bandwidth reservation service provider, all bandwidth reservation requests (BRRs) in one batch should be scheduled for high scheduling efficiency and system throughput. In this paper, we study the problem of scheduling all BRRs in one batch while achieving their best average transfer performance on one reservation path in an HPN. Two data transfer performance parameters, ECT and SD, are specifically considered. Because of the limited bandwidth resources of the reservation path, the problems of scheduling all BRRs in one batch on one reservation path while achieving their best average ECT and SD are converted into the problems of scheduling as many BRRs as possible while achieving the average ECT and SD of scheduled BRRs, respectively. We prove these two converted problems as NP-complete problems, and improve two existing heuristic algorithms proposed previously for similar problems. Extensive simulation experiments show the superior scheduling performance of these improved algorithms in terms of several performance metrics.

Fault-tolerant bandwidth reservation strategies for data transfers in high-performance networks

Many next-generation e-science applications require fast and reliable transfer of large volumes of data with guaranteed performance, which is typically enabled by the bandwidth reservation service in high-performance networks. One prominent issue in such network environments with large footprints is that node and link failures are inevitable, hence potentially degrading the quality of data transfer. We consider two generic types of bandwidth reservation requests (BRRs) concerning data transfer reliability: (i) to achieve the highest data transfer reliability under a given data transfer deadline, and (ii) to achieve the earliest data transfer completion time while satisfying a given data transfer reliability requirement. We propose two periodic bandwidth reservation algorithms with rigorous optimality proofs to optimize the scheduling of individual BRRs within BRR batches. The efficacy of the proposed algorithms is illustrated through extensive simulations in comparison with scheduling algorithms widely adopted in production networks in terms of various performance metrics.