Internet In A Box

1
Internet In A Box
Presented by Zhangxi Tan with Prof. David
Patterson RAD Lab, UC Berkeley
2
Outline

• Goal of this talk
• Introduction
• Comparison to existing testbeds
• Programmable Data Path
• Comments Discussion

3
Goal (1/2)

• Bring out the discussion
• What should Internet In A Box look like?
• Architecture
• Application
• Survey related projects
• Gather comments on existing projects (Pros/Cons)
• Collect suggestions for our Internet In A Box
  project

4
Goal (2/2)

• Questions
• What would be good experiments for IIAB ?
• Where IIAB is better and where worse than
  alternatives ?
• What data needed for such experiments?
• What measurements needed for such experiments ?

5
Outline

• Goal of this talk
• Introduction
• Comparison to existing testbeds
• Programmable Data Path
• Comments Discussion

6
Introduction

• Internet In A BOX
• Build distributed system testbed with over 1,000
  nodes using multi modular FPGA system
• Research Accelerator for MultiProcessing (RAMP)
• Motivation FPGA as New Research Platform
• FPGA generations every 1.5 yrs 2X CPUs, 2X clock
  rate
• 25 64-bit simple soft core RISC CPU (100Mhz in
  2004) can fit in FPGA (Virtex-II), 1000-CPU
  system from 40 FPGAs

7
RAMP

• 5 Xilinx Virtex II Pro, 4GB DDR2/ FPGA
• Multi Gigabit Transceivers (MGTs) connections
• 2.5 Gbps/Ch, up to 40 Gbps in total
• 1/10 Gb Ethernet, InfiniBand, 3GIO compatible
• Connecting heterogeneous routers

8
Pros of IIAB

• Advantages
• Large emulation scale O(1000) and great
  scalability
• Low cost, power consumption and space
  requirement
• Generate repeatable and credible results
• Hardware and software programmability
  (customized router data plane, OS and etc.)
• Interoperability with unmodified nodes
  (commercial routers, switches and PCs)

9
Outline

• Goal of this talk
• Introduction
• Comparison to existing testbeds
• Programmable Data Path
• Comments Discussion

10
Comparison to existing testbeds

• 3 basic network experimental techniques
• Emulation
• SMP/Cluster based network emulator (Duke, Utah)
• Live testing
• Overlay based testbed (NSF, USC, MIT)
• Simulation
• Traditional network simulators (ns2, OPNET)

11
Cluster Based Network Emulator (1/2)

• ModelNet _at_ Duke University
• http//issg.cs.duke.edu/modelnet.html

• Cluster based emulation
• Run unmodified live codes on end host
• ModelNet core is software Router Core
• Scale to O(100) nodes
• Use real IP traffic

"Scalability and Accuracy in a Large-Scale
Network Emulator," OSDI 2002.
12
Cluster Based Network Emulator (2/2)

• Emulab _at_ University of Utah www.emulab.net

• 328 Nodes connected by high speed Cisco switches
• Based on ns2 emulation engine (nse)
• Combined emulation/simulation
• Real-time execution delay
• (clock sync. problem)
• Extension
• Using network processor to accelerate emulation
• Interface to PlanetLab/RON

K. Fall. Network Emulation in the Vint/NS
Simulator. IEEE ISCC 99
13
Comparison to existing testbeds

• 3 basic network experimental techniques
• Emulation
• SMP/Cluster based network emulator (Duke, Utah)
• Live testing
• Overlay based testbed (NSF, USC, MIT)
• Simulation
• Traditional network simulators (ns2, OPNET)

14
Overlay based testbed (1/2)

• PlanetLab/GENI
  http//www.planet-lab.org

• Aims to deploy up to 1000 sites 637 nodes over
  302 sites (Jan 2006)
• Multi-user shared testbed (Slice)
• Based on Internet
• GENI is an upgraded version (www.geni.net)
• Limitations
• Poor performance
• Experiments not repeatable
• Difficult to debug
• Poor observability
• Typical nodes usage lt100

L. Peterson, T. Anderson, D. Culler, T. Roscoe A
Blueprint for Introducing Disruptive Technology
into the Internet. HotNets-I, October 2002.
15
Overlay based testbed (2/2)

• Other overlay based testbeds
• Resilient Overlay Networks (RON) MIT
• Active Network Backbone (ABone) USC/ISI
• X-Bone USC/ISI
• Similar problems with PlanetLab

16
Comparison to existing testbeds

• 3 basic network experimental techniques
• Emulation
• SMP/Cluster based network emulator (Duke, Utah)
• Live testing
• Overlay based testbed (NSF, USC, MIT)
• Simulation
• Traditional network simulators (ns2, OPNET)

17
Simulators

• The Network Simulator - ns2 http//www.isi.edu/ns
  nam/ns/
• Famous event-driven simulator
• Supports various OS
• Problems
• Abstract network and modeled traffic (less
  credible results)
• Does not model actual packet processing
• Poor performance

18
Summary Why IIAB is good
19
Outline

• Goal of this talk
• Introduction
• Comparison to existing testbeds
• Programmable Data Path
• Comments Discussion

20
Existing Attempts

• Survey of existing projects
• FPGA based routers and packet processors
  (Stanford/Washington)
• Network processor based routers
  (Princeton/Columbia)
• Software routers (MIT, Berkeley)

21
FPGA based routers/packet processors (1/2)

• Field Programmable Port Extender (FPX) -
  Washington University http//www.arl.wustl.edu/pro
  jects/fpx/

• OC48 (2.5 Gbps) link speed
• All logic on Xilinx FPGA Virtex II/Virtex
• Dynamically loadable modules into FPX
• Projects based on FPX
• Intrusion/worm detection
• Fast routing
• Advanced protocol processing

22
FPGA based routers/packet processors (2/2)

• NetFPGA _at_ Stanford (Prof. Nick McKeown)
  (http//klamath.stanford.edu/nf2/)
• Goal Undergrad/graduate education
• Building edge/border router/Ethernet switch on
  FPGA board
• 2 Xilinx Virtex II Pro FPGAs
• Current Scale ? (Hot Interconnects 13, 05) 8,
  10-Mbps Ethernet ports per board
• Small scale network emulation (multiboard)

23
Existing Attempts

• Survey of existing projects
• FPGA based routers and packet processors
  (Stanford/Washington)
• Network processor based routers
  (Princeton/Columbia)
• Software routers (MIT, Berkeley)

24
Network Processor

• Network Processor An application specific
  processor for network applications
• Performance and Flexibility
• Tradeoff between ASIC and GPP
• Multicore, multithread architecture
• Many commercial products
• Intel IXP 2800/1200 (active)
• IBM PowerNP (discontinued?)
• Motorola C-5

Intel IXP2800
25
Network Processor Based Router (1/2)
Extensible Router VERA _at_ Princeton (Prof. Larry
Peterson)http//www.cs.princeton.edu/nsg/router

• Goals
• Easily extended to support new network services
  (overlay, media gateway, etc.)
• Exploits commercially available hardware
  components (commodity processors, network
  processors)

26
Network Processor Based Router (2/3)
PC
Data Plane (Network Processor) -Common case
traffic -Traditionally only IP Control Plane
(PC) - Exceptional traffic (Fancy traffic
QoS, Monitor, etc) -Traditionally routing
protocols
Control Plane (packet flows)
Data Plane (packet flows)
Building a Robust Software-Based Router Using
Network Processors, SOSP01 A Comparative Study
of Extensible Routers, OPENARCH 02
IXP1200
VERA Architecture
27
Network Processor Based Router (3/3)

• Limitation
• Hard to program network processor
• Related project
• Columbia Gensis Kernel on IXP1200
  (http//comet.columbia.edu/genesis/)
• Intel IXA university program (http//www.ixaedu.co
  m)
• Research projects
• Papers

28
Existing Attempts

• Survey of existing projects
• FPGA based routers and packet processors
  (Stanford/Washington)
• Network processor based routers
  (Princeton/Columbia)
• Software routers (MIT, Berkeley)

29
Software Based Router (1/2)

• MIT Click Modular Router
• Also a programming language for router data plane
  programming
• Modular programming concepts benefit other
  projects
• Berkeleys NP-Click for Intel IXP1200
• Building FPGA-Click for IIAB?

http//pdos.csail.mit.edu/click/
30
Software Based Router (2/2)

• eXtensible Open Router Platform (XORP) (ICSI,
  UC Berkeley) http//www.xorp.org

• Recent attempt on open source software router
  (NSDI05)
• Flexible, extensible and standardized interface
  for network research
• Two layered architecture (High/Low), but mainly
  focus on higher layer (control plane)
• Lower plane using Click, but will support
  customized hardware (e.g. Intel IXP1200)
• Will IIAB support XORP API?

31

• Comments Discussion

32
Discussions

• What about hosts and links?
• Any related projects?
• How to build the world?
• What to build and what not to? (level of details)
• Topology, host/router ratio
• Resource consumption
• Any other Application?
• Starting from LAN in a box?
• Host/Router/Link virtualizations?
• Tens of hosts and several switches/routers
• B2EE platform

33
Hardware/FPGA based router (3/4)
NetFPGA in VNS

• Separate virtual topologies using VLAN (encoding
  MAC address)
• Scale?
• Emulate/simulate complex topologies, like
  Internet?

• Reconfigurable Networking Hardware A Classroom
  Tool Martin Casado, Gregory Watson, Nick
  McKeown, Hot Interconnects 13, Stanford, August
  2005

34
Introduction

• FPGA as New Research Platform
• FPGA generations every 1.5 yrs 2X CPUs, 2X clock
  rate
• 25 64-bit simple RISC CPU (100Mhz in 2004) can
  fit in FPGA (Virtex-II), 1000-CPU system from 40
  FPGAs
• Internet In A BOX
• Build a distributed system testbed with over
  1,000 nodes in a controlled environment