Datacenter Network Visualization in Multi-Tenant Environments

Outsourcing of computation and storage infrastructure into the cloud entails new challenges for the architecture and design of multi-tenant datacenters. With the evolution of virtualization techniques, tenant applications do not need to be operated on dedicated servers. Software switches hereby play an important role by mediating between physical infrastructure and virtualized applications. In this paper, we discuss the application of Open vSwitch, a software implementation of a network switch that is particularly designed for use in cloud environments. We analyse its features and performance in comparison to traditional concepts. Software-defined datacenters provision pools of compute, storage and networking resources and distribute them to their customers [BBE13b]. Today, primarily networking is seen a barrier to software-defined datacenters [FRZ13]. Address spaces that are tied from guest operating systems down to features that are installed on physical network devices, i.e. physical load balancers, result in slower provisioning and limited placement options. VMs (Virtual machines) are geographically dependent on hardware features, vendors and the deployment on underlying fabrics. Programmatic provisioning, control and visibility, allow instant deployment of complex Layer 2 to Layer 7 topologies. From a provisioning perspective maintaining those services is operationally intensive and expensive [GHMP09]. Network Virtualization introduces a layer between physical networking equipment and guest OS. The virtual overlay network can e.g. be created with a tunnelling technology called VXLAN (Virtual Extensible Local Area Network). We discuss a novel approach to a software-defined datacenters as dealt with in the LRZ (Leibniz Super Computing Centre in Munich) infrastructure. We describe the state-of-the-art in datacenter network virtualization in Section 1. Section 2 provides an overview to Open vSwitch. We look at selected features and the overall performance compared to cutting edge hardware switches and experimental ports like the DPDK vSwitch. Section 3 summarizes this paper. 1 Network Virtualization An optimal resource allocation by the virtualized networks directly relates to financial savings [GHMP09]. Existing methods for network virtualization consist of plenty of primitives and proposals, but are nothing more than point solutions, in the sense that only single aspects of networking are being virtualized [CKRS10]. VLANs (Virtual Local Area Networks) can virtualize the IP space by applying NAT (Network Address Translation) and therefore, share overlapping IP addresses with multiple tenants. With MPLS (Multi Protocol Label Switching), paths across multiple physical networks can be virtualized. VRFs (Virtual Routing and Forwarding) enables infrastructure operators to have virtual FIBs (Forwarding Information Bases) and therefore moves the routing processes into VMs. On behalf of OpenFlow the Data and Control Plane split paves the way for SDN (Software Defined Networking) [sdn].