검색 상세

Programmable Network Virtualization for SDN-based Cloud Systems

Programmable Network Virtualization for SDN-based Cloud Systems

  • Yang, Gyeongsik (양경식, 대학원 컴퓨터학과 소프트웨어전공)

초록/요약

Cloud computing is an indispensable component of IT infrastructure. It relies on network virtualization to provide network connections to each tenant. Current network virtualization solutions in clouds automate the creation and operation of virtual networks to reduce the operational costs of network management and facilitate effective, efficient management of network infrastructures by the datacenter operator. However, these network virtualization solutions are not well-suited to meet the emerging demands to control virtual networks in a finer granularity, with tenants given the ability to directly program virtual switches and links in their virtual networks. To address these gaps, my study proposes a model called programmable Network Infrastructure-as-a-Service (p-NIaaS). In the proposed model, tenants can create virtual network infrastructures on an on-demand basis. The p-NIaaS model also overcomes other shortcomings of current network virtualization paradigms: it enables tenants to program their own packet processing logic and monitor network status from any virtual network infrastructures, which is not possible in current alternatives. This thesis also presents Libera network hypervisor to implement the p-NIaaS model. Libera provides for tenant-demanded virtual networks that can be programmed using software-defined networking concepts. Libera outperforms existing network hypervisors in the attributes of scalability, VM migration support, traffic engineering, and flexibility, where existing network hypervisors fall short in practical usability. Libera encompasses the following research objectives and methods to surmount the shortcomings of current alternatives: - Scalability: reducing the of resource consumption overheads resulting from the programmability provided to tenants, compared to the existing datacenter’s network virtualization solutions. - VM Migration Support: network reconfiguration techniques to efficiently allow seamless network services for tenants during VM migrations. - Traffic Engineering Support: enhancement of performance (e.g., throughput) of virtual networks through accurate and overhead-light network monitoring. - Flexibility: clean-state design that supports a range of desired physical switches (e.g., OpenFlow 1.0, OpenFlow 1.3, and white-box switches) and enables the dynamic mapping change of virtual network infrastructures while they are in operation to achieve VM migration and traffic engineering support. This thesis begins with a comprehensive introduction of Libera’s design, followed by demonstrations of proof-of-concept evaluations. The results show that Libera is highly scalable (with an enhancement of up to eightfold improvement) and exhibits desirable performances in supporting VM migration and traffic engineering (i.e., in achieving tenant-requested throughputs for their virtual network connections). In overhead comparisons, Libera incurs about 11 and 25 percent virtualization overheads, compared to a non-virtualized network. In conclusion, the results indicate that p-NIaaS is a promising prototype for an alternative to existing network virtualization solutions. The thesis continues by detailing schemes and evaluations for LiteVisor, which improves scalability and VM migration support based on the Locator, Identifier, and Tenant sEparating (LITE) forwarding scheme. The LITE scheme reduces the memory consumption of physical switches and supports the network reconfiguration to enable seamless VM migration, which are not feasible in existing network hypervisors. The thesis concludes by identifying areas for future research. One such focus of possible research is CLEO, a machine learning approach to enhanced traffic load balancing to improve the accuracy of throughput distribution in equal-cost multipath routing. The initial design and results of CLEO that guides research efforts in performance-assured programmable virtual networks are presented.

more

목차

Abstract

Contents i

List of Figures v

List of Tables viii
1 Introduction 1
1.1 Programmable Network Virtualization and Libera 1
1.2 Scope of the Study 3
2 Background and Motivation 6
2.1 Software-defined Networking 6
2.2 Datacenter Network Virtualization 7
2.3 SDN Network Virtualization 10
2.3.1 Scalability 13
2.3.2 VM migration support 15
2.3.3 Traffic engineering support 16
2.3.4 Flexibility 16
3 Libera for Programmable Network Virtualization 17
3.1 Programmable Network Infrastructure-as-a-Service Model 17
3.1.1 VN manager 18
3.1.2 VN controller 19
3.2 Libera network hypervisor components 21
3.2.1 VNM handler 21
3.2.2 Flexibility component 22
3.2.3 Scalability component 23
3.2.4 VM migration support component 26
3.2.5 TE support component 26
3.2.6 Control translation 27
3.3 Proof of Concept 28
3.3.1 Scalability 28
3.3.2 VM migration support 29
3.3.3 TE support 30
3.3.4 Overheads 30
4 LiteVisor: Scalability Improvement and VM Migration Support 32
4.1 Introduction 32
4.2 Motivation and Related Work 34
4.2.1 Shortage of flow rule memory 34
4.2.2 Network reconfiguration for VM migration 38
4.2.3 Motivation and goal 40
4.3 Design 41
4.3.1 LiteVisor architecture 41
4.3.2 LITE and LITE-based forwarding 42
4.3.3 Flow aggregation 46
4.3.4 Seamless network reconfiguration 48
4.3.5 Considerations 51
4.4 Evaluation 54
4.4.1 Flow aggregation 55
4.4.2 Seamless network reconfiguration 58
4.5 Discussion 62
4.5.1 Isolation of an aggregated flow 62
4.5.2 Tunneling overheads 63
4.5.3 Security 64
5 V-Sight: Network Monitoring for Programmable Virtual Network 65
5.1 Introduction 65
5.2 Background and Motivation 68
5.2.1 Existing schemes for network monitoring in SDN-based NV . 68
5.2.2 Requirements for network monitoring framework in SDN-based NV 71
5.3 V-Sight Framework Design 75
5.3.1 V-Sight framework architecture 75
5.3.2 Statistics virtualization 77
5.3.3 Transmission disaggregation 80
5.3.4 Pre-monitor scheduling 83
5.4 Evaluation 88
5.4.1 Settings 88
5.4.2 Transmission Delay 89
5.4.3 Scalability 93
5.4.4 Accuracy Overheads 95
5.5 Related Work 97
5.5.1 Monitoring in SDN 97
5.5.2 Monitoring in SDN-based NV 98
5.6 Discussion 98
5.6.1 Monitoring on programmable dataplane 98
5.6.2 Active monitoring virtualization 99
6 Possible Future Research Directions 100
6.1 Introduction 100
6.2 Motivation 102
6.3 Design and Evaluation 103
7 Conclusion 106
Bibliography 108

Acknowledgement

more
반출 Meta View 목록