Network Virtualization - Computer Networks Group

Chair of Future Communication Prof. Dr. K. Tutschku Institute for Multimedia and Distributed Systems Faculty of Computer Science

Network Virtualization: Implementation Steps Towards the Future Internet K. Tutschku (Uni Würzburg) , T. Zinner (Uni Würzburg), A. Nakao (NICT/UT), P. Tran-Gia (Uni Würzburg)

The Internet is Broken?

Typical deficiencies: Addressing, mgmt complexity, protocol ossification

Clean slate approach? UC Berkey, Intel: Declarative Networking; Parc (V. Jacobson): Content Centric Networking

Evolutionary approach? IRTF RRG/ ALTO BOF: Location / Identifier Split, (ISP-Supported) Path Selection Service

Brief analysis of deficiencies and achievements from application point of view

What Features are Currently Missing?

Performance (End-to-end Quality), but: QoS-Islands are available (depending on technology and provider)

Reliability, but: Sophisticated Resilience concepts available at experienced ISPs

Competition/Business models, but: J. Crowcroft: “… i can go on the web and get my gas, electricity, … changed , why is it not possible to get a SPOT price for broadband internet?” (E2E-interest mailing list on April 26th, 2008)

Achievements of todays‘s Internet

P2P, 67,3% eMail, 1,2%

Efficient P2P CDNs Networks Typical traffic distribution in residential access systems

FTP, 0,3% Web, 7,9%

other, 23,3% Source: Telefonica (2003)

Internet Video Traffic Benchmarks

Terrabytes per month YouTube − worldwide (Cisco est., May 2008)

100.000

P2P Video Streaming in China (Jan. 2008)

33.000

YouTube − United States (May 2008)

30.500

US. Internet backbone at year end 2000

25.000

US. Internet backbone at year end 1998

6.000 Source: CISCO (2008)

Multi-Source Download (eDonkey, BT) Offers file X

Peer Offers file X

Transfer of segment B

Offers file X Index server

Transfer of segment A Looking for X

Two overlays (virtual structures) with different application layer functions (searching / content exchange); each with different topology, addressing, and routing Symmetric roles (consumer and provider) and integrative (multi-network services) Selection of providing peer (routing of content) based on resource quality (throughput)

Diversity I: multi-provider environment West coast

East coast


High diversity wrt. paths: § Three North-american nationwide ISPs Tier1 (AS 3967 Exodus, AS3356 Level3, AS6467 Abovenet; M. Liljenstam et al., 2003)

Multiple routes for increased resilience and competition are (theoretically ) readily available! Autonomous identification of available resources needed

(Thanks to Michael Menth für vsualization)

Terremark: NAP of the Americas


carrier-neutral Network Access Point
Located in Miami
Connectivity from 160 carriers and 148 countries
Available technologies: § Optical, Ethernet, MPLS, Voice and legacy Internet traffic

Networks under change: transport systems Management plane

Service request (FAX, Web)

„semi-manual“ provisioning

E3 Remote office

ATM Headquarter

Networks under change: transport systems Management Plane Control Plane auto. Signaling auto. provisioning IP layer EPON Remote office

100GE layer

DWDM layer

Headquarter

MultiLayerNetworks


State-of-the-art optical transport systems: § Ultra-high transmission capacities; embedding of different transport network into one physical network (multi-layer networks) § Decay of CAPEX per Bit Increased automation self-* features (self-operation, self-organization)

However: higher complexity („numerous overlays“?)

Diversity II: multi-quality environment


25% of paths violate the triangle inequality (wrt. packet delay) § Measurements in PlanetLab by S. Banerjee et al. (2004)

Using an intermediate

A

direct connection

➞ Internet routing is far from optimal B C ➞ Better paths exist; capacity is readily available Triangle Inequality (TI): D(A,C) ≤ D(A,B) + D(B,C) ➞ Can be offered (competition) ➞ Again: autonomous identification of available resources needed ! „Multi-homing“ not really available current IP protocols

Operating System Virtualization Manager

Apps

apps

Apps

Apps

MngOS

GuestOS

GuestOS

GuestOS

GuestOS

VM

VM

VM

VM

VM

Apps

VMM

HostOS

Hardware

Hardware

Hypervisor-based

Host-based

Consolidation of multiple server into a single machine operational expenditures (OPEX)

reduction of

Private machines with strong resource isolation, e.g. for testing or for personal configuration Use computer center as a PC with user-specific environment Memory/space invariance: location doesn’t matter Small snapshots: SBUML (Scarp-Book User Mode Linux; Potter et al., 1999) need only 10% of real size fast transition of complete router configuration ! However: avoid “multi-layer trap” (1000s of VMs)

Smart resource mgmt

Virtual networks for convergent services Stellt X zur Verfügung

Diversity § Exploit diversity of resources § Locate optimal resources

Peer Stellt X zur Verfügung

Transfer von Segment B

Stellt X zur Verfügung Index server

Transfer von Segment A Sucht X

Overlays § Overlays: application-oriented topology, addressing, and routing § Multi-Network Services § Self-operation of functions

Network






OS virtualization § Consolidation and efficient operation § Strong isolation of resources

virtualization

Build a „personal network (PN)” for an application (PN  PC) Integration of different technologies and administrative domains Push application-layer mechanisms safely down the stack Re-use of generic infrastructure on small time scale Resource management? Avoid “multi-layer” trap by autonomic/self-* operation

Implementation: Xen network virtualization architecture Driver domain

Virtual machine 1

Back-end driver Ethernet bridge

Frond-end driver

Currently: local actions only! Virtual machine 2 Frond-end However: How to control driver resource isolation across Hypervisor controls multiple entities? Virtual machine n fair access and Frond-end resource isolation driver Virtual interrupts

NIC driver interrupts Driver control

Packet data

Interrupt dispatch

Hypervisor Control + data

NIC

Self-Organization? CPU/memory/disk/other devices Scheduling? From: S. Rixner “Network Virtualization – Breaking the Performance Barrier”, ACM Queue, Jan./Feb. 2008.

Transport Virtualization (TV)
Example: Virtual Memory § OS integrates disconnected physical memory, even disk space, into continuous memory § location of physical memory doesn’t matter


Transport Virtualization (Tutschku, Nakao, 2008): abstraction concept for data transport resources
Physical location of transport resource doesn't matter (as long resource is accessible)
Achieved by: abstract data transport resources § combined from one or more physical/overlay transport resources, e.g. leased line, wave length path, an overlay link, MPLS path, or an IP forwarding capability § physical resources can be used preclusive or concurrently § basic resources can be located in even different physical networks or administrative domains

Routing Overlays: One-hop Source Routing (OSR) Routing Overlay

1 2

Divert selected endhost packets

Path oracle Request Paths for Diverted Packets

3

Encapsulated, send using path

4

Decapsulate, egress to destination

One-hop Source Router (OSR)

Gummadi et al (2004): Increased reliability while being scalable Nakao, Khor, Lane (2007/8): Multipath aware application to resource mgmt Control/data plane separation implemented (path selection and forwarding) ! May be inefficient

Reduction of overhead (since edge-based) Spread network-virtualization capable

routers almost everywhere (in particular in core)

Virtual Resource Management: Concurrent Multipath Transfer in Multi-homed Overlays (aka Striping)

overall transport pipe

Different overlays

(First) aim: Solution:

Individual Overlay paths

Physical topologies of different providers

Obtain high throughput (i.e. performance) Combine multiple paths (provided by disjoint overlays) into an overall transport pipe

Virtual Resource Management: Concurrent Multipath Transfer in Multi-homed Overlays (aka Striping)

overall transport pipe

Different overlays

Individual Overlay paths

Physical topologies of different providers

Features: Increased reliability and very high capacity Interdomain traffic management and edgebased performance control Avoid congested areas

Conceal triangle inequality violations Exploit high access bandwidth/ discover available bandwidth in core

Mechanism and Evaluation

Mean value analysis is insufficient and second moment

consider first

Conclusion
Expected Features of the Future Internet § Faster, more reliable, more business cases, increased interaction with users: symmetric rolls, „Architecture for Participation“ § Forming of applications-specific overlays

 Integration/convergence of different transport systems und operator domains by overlays ? However: How to manage a bunch of overlays? Cross talk vs. Isolation? Local vs. Distributed Optimization!  Design networks for applications (rather than designing applications for networks)
Network virtualization: § Consolidation of multiple (virtual) network into one physical infrastructure § Making networks independent from resource locations transport virtualization

Thank you! Questions?