Towards An Efficient Cloud Computing System: Data Management, Resource Allocation and Job Scheduling
Date of Award
8-2016
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Electrical and Computer Engineering (Holcomb Dept. of)
Committee Member
Professor Haiying Shen, Committee Chair
Committee Member
Professor Kuang-Ching Wang
Committee Member
Professor Adam Hoover
Committee Member
Professor James (Zijun) Wang
Abstract
Cloud computing is an emerging technology in distributed computing, and it has proved to be an effective infrastructure to provide services to users. Cloud is developing day by day and faces many challenges. One of challenges is to build cost-effective data management system that can ensure high data availability while maintaining consistency. Another challenge in cloud is efficient resource allocation which ensures high resource utilization and high SLO availability. Scheduling, referring to a set of policies to control the order of the work to be performed by a computer system, for high throughput is another challenge. In this dissertation, we study how to manage data and improve data availability while reducing cost (i.e., consistency maintenance cost and storage cost); how to efficiently manage the resource for processing jobs and increase the resource utilization with high SLO availability; how to design an efficient scheduling algorithm which provides high throughput, low overhead while satisfying the demands on completion time of jobs.
Replication is a common approach to enhance data availability in cloud storage systems. Previously proposed replication schemes cannot effectively handle both correlated and non-correlated machine failures while increasing the data availability with the limited resource. The schemes for correlated machine failures must create a constant number of replicas for each data object, which neglects diverse data popularities and cannot utilize the resource to maximize the expected data availability. Also, the previous schemes neglect the consistency maintenance cost and the storage cost caused by replication. It is critical for cloud providers to maximize data availability hence minimize SLA (Service Level Agreement) violations while minimize cost caused by replication in order to maximize the revenue. In this dissertation, we build a nonlinear programming model to maximize data availability in both types of failures and minimize the cost caused by replication. Based on the model's solution for the replication degree of each data object, we propose a low-cost multi-failure resilient replication scheme (MRR). MRR can effectively handle both correlated and non-correlated machine failures, considers data popularities to enhance data availability, and also tries to minimize consistency maintenance and storage cost.
In current cloud, providers still need to reserve resources to allow users to scale on demand. The capacity offered by cloud offerings is in the form of pre-defined virtual machine (VM) configurations. This incurs resource wastage and results in low resource utilization when the users actually consume much less resource than the VM capacity. Existing works either reallocate the unused resources with no Service Level Objectives (SLOs) for availability\footnote{Availability refers to the probability of an allocated resource being remain operational and accessible during the validity of the contract~\cite{CarvalhoCirne14}.} or consider SLOs to reallocate the unused resources for long-running service jobs. This approach increases the allocated resource whenever it detects that SLO is violated in order to achieve SLO in the long term, neglecting the frequent fluctuations of jobs' resource requirements in real-time application especially for short-term jobs that require fast responses and decision making for resource allocation. Thus, this approach cannot fully utilize the resources to process data because they cannot quickly adjust the resource allocation strategy dealing with the fluctuations of jobs' resource requirements. What's more, the previous opportunistic based resource allocation approach aims at providing long-term availability SLOs with good QoS for long-running jobs, which ensures that the jobs can be finished within weeks or months by providing slighted degraded resources with moderate availability guarantees, but it ignores deadline constraints in defining Quality of Service (QoS) for short-lived jobs requiring online responses in real-time application, thus it cannot truly guarantee the QoS and long-term availability SLOs. To overcome the drawbacks of previous works, we adequately consider the fluctuations of unused resource caused by bursts of jobs' resource demands, and present a cooperative opportunistic resource provisioning (CORP) scheme to dynamically allocate the resource to jobs. CORP leverages complementarity of jobs' requirements on different resource types and utilizes the job packing to reduce the resource wastage and increase the resource utilization.
An increasing number of large-scale data analytics frameworks move towards larger degrees of parallelism aiming at high throughput. Scheduling that assigns tasks to workers and preemption that suspends low-priority tasks and runs high-priority tasks are two important functions in such frameworks. There are many existing works on scheduling and preemption in literature to provide high throughput. However, previous works do not substantially consider dependency in increasing throughput in scheduling or preemption. Considering dependency is crucial to increase the overall throughput. Besides, extensive task evictions for preemption increase context switches, which may decrease the throughput. To address the above problems, we propose an efficient scheduling system Dependency-aware Scheduling and Preemption (DSP) to achieve high throughput in scheduling and preemption. First, we build a mathematical model to minimize the makespan with the consideration of task dependency, and derive the target workers for tasks which can minimize the makespan; second, we utilize task dependency information to determine tasks' priorities for preemption; finally, we present a probabilistic based preemption to reduce the numerous preemptions, while satisfying the demands on completion time of jobs.
We conduct trace driven simulations on a real-cluster and real-world experiments on Amazon S3/EC2 to demonstrate the efficiency and effectiveness of our proposed system in comparison with other systems. The experimental results show the superior performance of our proposed system. In the future, we will further consider data update frequency to reduce consistency maintenance cost, and we will consider the effects of node joining and node leaving. Also we will consider energy consumption of machines and design an optimal replication scheme to improve data availability while saving power. For resource allocation, we will consider using the greedy approach for deep learning to reduce the computation overhead caused by the deep neural network. Also, we will additionally consider the heterogeneity of jobs (i.e., short jobs and long jobs), and use a hybrid resource allocation strategy to provide SLO availability customization for different job types while increasing the resource utilization. For scheduling, we will aim to handle scheduling tasks with partial dependency, worker failures in scheduling and make our DSP fully distributed to increase its scalability. Finally, we plan to use different workloads and real-world experiment to fully test the performance of our methods and make our preliminary system design more mature.
Recommended Citation
Liu, Jinwei, "Towards An Efficient Cloud Computing System: Data Management, Resource Allocation and Job Scheduling" (2016). All Dissertations. 2292.
https://tigerprints.clemson.edu/all_dissertations/2292