A scalable multi-datacenter layer-2 network architecture

Chen Chen, Changbin Liu, Pingkai Liu, Boon Thau Loo, Ling Ding
2015 Proceedings of the 1st ACM SIGCOMM Symposium on Software Defined Networking Research - SOSR '15  
Cloud today is evolving towards multi-datacenter deployment, with each datacenter serving customers in different geographical areas. The independence between datacenters, however, prohibits effective inter-datacenter resource sharing and flexible management of the infrastructure. In this paper, we propose WL2, a Software-Defined Networking (SDN) solution to an Internet-scale Layer-2 network across multiple datacenters. In WL2, a logically centralized controller handles control-plane
more » ... n and configuration in each datacenter. We achieve scalability in three ways: (1) eliminating Layer-2 broadcast by rerouting control-plane traffic to the controller; (2) introducing a layered addressing scheme for aggregate Layer-2 routing; and (3) creating an overlay abstraction on top of physical topology for fast flow setup. WL2 is fault-tolerant against controller and gateway failures. We deployed and evaluated WL2 in a 2,250-VM testbed across three datacenters. The results indicate high performance and robustness of the system. However, current cloud infrastructure is more like multiple clouds rather than a multi-datacenter cloud, in that different datacenters usually operate independently except for data replication. This setup has several drawbacks. First, it is hard for network administrators to manage the infrastructure as a whole, prohibiting interdatacenter optimizations such as load balancing and virtual machine (VM) migration. Second, customers, especially those that have business presence at different locations, may demand an interdatacenter sub-network to achieve low-latency and seamless access from each location. This requirement cannot be easily satisfied by today's cloud. A datacenter is often built as an IP network for scalability. However, this increases the complexity of network management by introducing substantial configuration overhead (e.g., IP division, BGP routing). Moreover, since IP adopts the location-aware addressing scheme, it is difficult, if not possible, to perform VM live migration [17] across IP prefix boundaries. IP address change during VM migration would inevitably invalidate existing network sessions. In contrast, Layer-2 (L2) network provides a simple abstraction for both users and administrators. Central to L2's ease of management is its plug-and-play semantics, which allows administrators to add or remove devices with minimal disruptions. Moreover, L2 network uses flat addressing scheme, which supports VM live migration in the network without disrupting ongoing network sessions [14, 16] . Public clouds including Amazon EC2 [13], Google Compute Engine [6], and Microsoft Azure [28] all present users with L2 network abstraction. OpenStack [31], the de facto standard of open cloud platforms, also adopts L2 network. Despite its ease of configuration and management, L2 network does not scale, particularly under the scenario of wide-area networks (WANs). The major scalability bottleneck lies at the broadcast nature of control traffic in a L2 network, such as Spanning Tree Protocol (STP), Dynamic Host Configuration Protocol (DHCP) and Address Resolution Protocol (ARP). A number of unconventional L2 designs [18, 30, 34] have been proposed to address the scalability issues of L2 within a single datacenter. It remains challenging to design and implement a scalable L2 solution in the wide-area scenario, where multiple datacenters of potentially heterogeneous topologies are interconnected via the Internet. In light of this, we present WL2 (Wide-area Layer-2), which provides a L2 network abstraction that spans multiple datacenters across diverse geographical regions, and yet maintains good scalability and high availability. Specifically, WL2 provides the following features: • Software-defined networking (SDN). WL2 uses an SDNbased architecture [23] to achieve scalability in a wide-area L2 network. In each datacenter, WL2 deploys a centralized SDN controller cluster, and creates full-mesh virtual overlay
doi:10.1145/2774993.2775008 dblp:conf/sosr/ChenLLLD15 fatcat:xif7clzbxrb7vfnykbn6q3ohte