A high-efficient inter-domain data transferring system for virtual machines.
Dingding Li,Hai Jin,Yingzhe Shao,Xiaofei Liao
DOI: https://doi.org/10.1145/1516241.1516308
2009-01-01
Abstract:ABSTRACTVirtual machine (VM) technologies are becoming more and more important among industrial and academic institutions with the decreasing cost of computer hardware. It can offer a lot of benefits including performance isolation, server consolidation, and live migration. However, the development of virtualization technology is preferred to application isolation as yet, thus the performance overheads of communicate between virtual machines which are resident on the same physical machine are relatively high. In network-intensive applications, such as Internet servers, are consolidated in a physical machine using VM technology, the inter-domain shared memory communication mechanisms is necessary. In other words, nearly all the virtual machine monitors are not very efficient in I/O performance to meet these applications' requirements. As a result, some scholars resort to inter-domain shared memory mechanisms to improve data transferring performance, but some of these solutions only provide one-way tunnel for domains to communicate with each other, and others are not network-bypass, which use the existing AF_INET protocol family, and incur unnecessary overhead brought by TCP/IP stack. In this paper, we present an inter-domain data transferring system, called IDTS, with dual-way tunnels and fully network-bypass features in Xen virtual machine environment. The design of IDTS is different from virtual network scheme in that IDTS uses static circular memory buffer shared between domains instead of the Xen page-flipping mechanism. In order to enable dual-way communication, IDTS provides two shared memory tunnels for co-resident domains to communicate with each other. Our evaluation demonstrates that IDTS can reduce inter-domain round trip latency by nearly a factor of 4, and increase bandwidth by approximately 5 times, and also keep high scalability, compared with traditional TCP/IP approach.