Design and Implementation of a QoS Supported Gateway VNF for InfiniBand and Ethernet Networks

Abstract

High —performance platforms such as Datacenters or High —Performance Computing (HPC) systems rely on high —speed interconnection networks to cope with the increasing communication requirements of modern applications. Most Datacenters or HPC systems uses InfiniBand(IP over InfiniBand (IPoIB) to support IP-based applications) or Gigabit Ethernet or both as the interconnection technology. However, these two networks have their inherent semantic differences. Therefore, it is imperative to introduce a gateway inorder to translate InfiniBand and Gigabit Ethernet protocols so as to make nodes following these protocols able to communicate with each other.

Existing hardware gateway solutions have speed advantages, but it is expensive at the same time in terms of scalability and manageability. In this thesis, we present a Gateway Virtual Network Function (VNF) that can be plugged in an NFV infrastructure and seamlessly translate Ethernet and InfiniBand (IPoIB) frames.

Primarily proposed gateway VNF uses the hardware addresses of both the InfiniBand and Ethernet nodes to learn the respective network segments. Once probing of network segments is completed, it then uses PKey —VLAN, SL —PCP mapping policies to create secure fabrics and support QoS across the heterogeneous fabric. Lastly, to translate the InfiniBand(IPoIB) and Ethernet frames it uses the learned hardware addresses, Pkey —VLAN mappings and SL —PCP mappings.

An OMNeT++ based IPoIB simulation model was also designed and calibrated with that of real InfiniBand IPoIB devices. This IPoIB simulation model together and a Gigabit Ethernet simulation model called Core4Inet (designed and maintained by) were then used to simulate our proposed gateway VNF. The results obtained shows that the performance achieved regarding end —end latency and priority application completion order by our proposed gateway VNF is at par with existing hardware gateway solutions with little expected variations of about 8.51%CV. With little or no variations in performance, scaling and management of our proposed gateway VNF will be less expensive than existing hardware solutions.