Fairness Issues in Software Virtual Routers

Size: px

Start display at page:

Download "Fairness Issues in Software Virtual Routers"

Cecily Watson
5 years ago
Views:

1 Fairness Issues in Software Virtual Routers Norbert Egi, Adam Greenhalgh, h Mark Handley, Mickael Hoerdt, Felipe Huici, Laurent Mathy Lancaster University PRESTO 2008 Presenter: Munhwan Choi

2 Virtual Router Table of contents Xen Virtual Router system System s baseline performance Evaluation of the forwarding planes Conclusion

3 Virtual Router A virtual router (VR) is an emulation of a physical router at the software and hardware levels VRs have independent IP routing and forwarding tables and they are isolated from each other Two main functions Constructing routing using any routing technology Forwarding packets to the next hops within the VPN domain From the VPN user point of view, a virtual router provides the same functionality as a physical router

4 Virtual Router Concept Customer Site(s) Provider s Network P P Customer Site(s) CE CECE PE PE CE CECE P P VPN Without VR P P CE CECE VR VRVR VR VRVR CE CECE P P VPN With VR

5 Virtual Routers: Why? One box can play the role of multiple independent routers. Resources sharing, management flexibility Multiple organizations sharing a single physical router. Small businesses within one building. Internet exchange Points. Entire physical network can be shared. Virtualize the routers. Tunnel between virtual routers over shared IP infrastructure. And then you have an Inter-domain overlay Great platform for experimentation (cf. VINI).

6 Virtual Router Issues Balance between performance, flexibility, isolation and fairness. Virtual router platform Ensures isolation and fairness between concurrent virtual routers without sacrificing performance

7 Xen virtualisation platform Xen virtualisation platform running Linux using the Click modular router package for forwarding Click s modular model Simple and flexible Close to optimal forwarding performance Support multi-threading, multi-core x86 CPUs

8 Xen Xen aims to be able to execute multiple operating systems on one physical x86 machine: Support popular OS (Linux, Windows XP, NetBSD), Scalable up to around 100 VMs, Securely, With close-to-native performance. Xen is a virtual machine monitor (VMM).

9 Xen Implementation CPU Virtualization of the CPU CPU Scheduling Time & Timers Memory Management Virtual Address Translation Physical Memory Device I/O Network Disk Disk Control Transfer

10 Para-virtualization Para-virtualization is a process where the guest operating system is modified to run in parallel with other modified systems Advantage: Improved performance Disadvantage: The hosted operating system must be modified before it can be hosted by the Xen Hypervisor (can be difficult)

11 Virtualization Overview Single OS image: Virtuozo, Vservers, Zones Group user processes into resource containers Hard to get strong isolation Full virtualization: VMware, VirtualPC, QEMU Run multiple unmodified guest OSes Hard to efficiently virtualize x86 Para-virtualization: ti UML, Xen Run multiple guest OSes ported to special arch Arch Xen/x86 is very close to normal x86

12 Hypervisor Control In Xen Domain0 is given greater access to the hardware and the hypervisor. It has: A guest OS running above the domain Hypervisor Manager software to manage elements within other existing domains. In VMWare Host OS acts underneath the domain.

13 Xen architecture Xen s Approach Xen 1.2 Architecture Xen 3.0 Architecture

14 Virtual Router system CPU Route Table Buffer Memory Line Interface MAC Line Interface MAC Line Interface MAC 14

15 Xen network internal overview

16 System s s baseline performance Dell PowerEdge 2950 servers Two 2.66 GHz quad-core Intel X5355 CPUs. Two dual-core CPUs in one package, each pair shares a 4MB L2 cache. A uniform memory architecture, with 8GB of 667MHz DDR2 memory Connected to the north bridge via four 5.3 GB/s channels. Networking Three Intel gigabit quad-port cards each using a PCIe x4 channel, for a total of 12 gigabit ports.

17 Forwarding Performance Maximum packet forwarding rate For minimum-sized 64-byte packets 7.1 million packets per second (Mp/s), (equivalent to 3.64 Gb/s). For larger, 1024-byte packets 1.43 Mp/s, or Gb/s, Theoretical maximum bidirectional rate for all Theoretical maximum bidirectional rate for all 12 gigabit ports.

18 Forwarding Performance

19 Sharing CPUs Virtual routers is to try to keep a packet on the same core throughout its processing to improve cache locality and thus reduce memory accesses.

20 Sharing Interfaces If a NIC is shared between more than one virtual router, then packets need to be classified on input so they are handled by the right router. Software classification Packets are polled from the NIC prior to classifying them into separate queues for each virtual router

21 Scheduling elements Assign all forwarding paths of each of the virtual routers to the available CPU cores need to know the cost of each schedulable element and the resource entitlement of each virtual router. calculate the most effective forwarding path-tocore allocation scheme. If there is more than one schedulable element assigned to a core, ensure fairness is maintained i within the boundaries of that t core.

22 Evaluation of the forwarding planes Scenario 1: all of the virtual routers on the platform have identical forwarding paths. Scenario 2: each virtual router can have a different forwarding plane, but each must be built from a set of trusted and provided Click elements. Scenario 3: allowing users to build forwarding planes from untrusted elements.

23 Virtual Router Architecture

24 Scenario 1: Static virtual forwarding plane Three identical forwarding paths sharing a CPU core. Round-Robin scheduler is sufficient to provide an equal share of the available CPU resources and the memory resources.

25 Scenario 1 Fair scheduling when packets are arriving at different rates per forwarding path Different weight scheduling of 6 forwarding paths on 3 cores Click s default CPU scheduler is based on proportional-share stride scheduling

26 Scenario 2: Configurable virtual forwarding plane Three FPs sharing a core. Two are identical whilst the third needs more resources to process a packet.

27 Scenario 3: Customisable virtual forwarding plane 3 FPs in dom0 and 3 domus forwarding via Xen s I/Ochannel. The first FP in dom0 and the first domu are sharing a core, the others use separate ones. 5 domains were directly mapped each to a pair of interfaces between which they forward packet and memory intensive task(s) were run in different number of other domains.

28 Conclusion Performance demonstrate A significant degree of isolation (protection, little perceived performance impact) Fairness between the virtual routers Without sacrificing Performance Future plan - Dynamical scheduling Scheduling the memory accesses globally and dynamically allocating the forwarding paths to the available cores based on their current resource usage for maximum performance will become an even more crucial issue.

29 Reference N. Egi, A. Greenhalgh, M. Handley, M. Hoerdt, L. Mathy, and T. Schooley, Evaluating xen for virtual routers, in PMECT07, August Xen and the Art of Virtualization

Using Xen For Router Virtualisation

July 2007 Using Xen For Router Virtualisation Adam Greenhalgh, Mark Handley,Tim Schooley University College of London Norbert Egi, Mickaël Hoerdt, Laurent Mathy University of Lancaster Evaluating Xen for