WAVNet: Wide-Area Virtual Networks for Dynamic Provisioning of IaaS


Cloud computing offers a new resource-sharing paradigm, namely Infrastructure as a Service (IaaS), which provides a level of abstraction and isolation over underlying physical resources. IaaS over Wide-Area Networks (WANs) is a recent trend which enables sharing of Internet computing resources, including CPU cycles, storage resources and network devices. Computation is no longer restricted to a fixed site but can be re-located dynamically across geographical distance to improve manageability, performance and fault tolerance. Virtual machine (VM) live migration is the key technology that supports mobility of execution environment without incurring heterogeneity problem. However the current VM migration technology is designed for local-area networks (LANs), which prevents IaaS from being widely adopted for serving applications built across multiple security domains. Overlay-based virtual networking approaches shed some light on tackling the connectivity issues across NAT/firewalls. Yet current implementation does not provide a complete network infrastructure to adapt to IaaS, which requires close-to-native network performance and seamless support to VM live migration over WAN. We present WAVNet, a novel wide-area virtual network that supports IaaS on the Internet. Experiments show that WAVNet delivers advantageous performance and enables users to run parallel applications on virtual clusters with dynamic load balancing support.


Figure 1: WAVNet Overlay Network Figure 2: WAVnet System Architecture


WAVNet Design:

  • Replaces Xenís software bridge with a customized bridge that uses tap
  • Packet Assembler (PA): encapsulation and de-capsulation of data traffic and connection messages
  • WAV-Switch : keeps track of the connections, functions like a hardware Ethernet switch.
  • Host-to-host connections: built based on the combination of STUN protocol and UDP hole punching technique
Figure 3: Xen VM Migration (left) vs WAVNet VM migration (Right)

Performance Results


Figure 4: Migration Time Comparison (VM Image size = 128MB and 512MB)

Figure 5: Evaluation of Downtime, ICMP RTT, and ApacheBench HTTP throughput under WAVNet VM live migration (x : ICMP packet loss)

Copyright HKU CS Department 2009-2010