DAN Distributed Code Caching for Active Networks Dan Decasper, - - PowerPoint PPT Presentation

INFOCOM’98 980401 1

Distributed Code Caching for Active Networks Dan Decasper, Bernhard Plattner

dan@arl.wustl.edu, plattner@tik.ee.ethz.ch

Applied Research Laboratory (ARL), Washington University, St.Louis Computer Engineering and Network Laboratory (TIK), ETH Zurich

DAN

INFOCOM’98 980401 2

Active Networking

• Is active networking at gigabit rates possible?
• How does an architecture supporting such rates

look like?

• How does the platform to implement such a

system look like?

INFOCOM’98 980401 3

Programmable switches

• Very promising real world applications

demonstrated

• No major performance problems
• out-of-band “learning”

INFOCOM’98 980401 4

Capsules

• Very flexible but suffer from Performance and

Security problems

• assuming 10 Gb/s, 1 KByte packets

– Switch must forward 1.3 million packets/s on every port

• Using 300 MHz processor, that leaves 231

cycles to receive, process, and forward a packet

• no time for interpretation, virtual machines,

context switches

• No capsules in a multi-gigabit environments in

the near term

INFOCOM’98 980401 5

Do we need ALL the flexibility?

• Potential active networking functionality is more

application specific than user specific

• Number of active networking functions grows

with the number of new applications and communication standards

• Automatic installation and upgrading of such

functions is very desirable

INFOCOM’98 980401 6

The challenge

• Active networking should allow applications to

control networking nodes and how their packets are processed and forwarded

• Requirement should not considerable degrade

the performance of each network node

• Fundamental challenge:

– Allow relocating part of the processing from the end- systems into the network – minimize the amount of processing on a single node – make the processing as efficient as possible – keep the necessary flexibility and customizability typical to Capsules

INFOCOM’98 980401 7

DAN

• Our Architecture: Distributed Code Caching
• Hardware and Software Platform
• Applications
• Conclusions and Status

INFOCOM’98 980401 8

Function identifiers

• Ethernet/IPv4/TCP packet

– Functions identified by Protocol numbers/Port numbers or hardware

Ethernet Vers HLen TOS ID Fragment Offset Source Address Destination Address Total length Flags TTL

Protocol

Header Checksum Options (if any)

32 bits Protocol

Source Address Destination Address

Source Port

...

Destination Port

IPv4 TCP

INFOCOM’98 980401 9

Distributed Code Caching

• Abstract view:
• Today:

– Function identifiers commonly identify known functions

• r packet is dropped by the router.
• New:

– Let router look for the implementation of the identified function on a Code Server!

• Devices featuring that property are called

Active Network Nodes (ANN)

fi2 P1 fi3 P2 ... ... PN

INFOCOM’98 980401 10

Distributed Code Caching (Cont.)

ANN ANN Code Server Workstation ANN Workstation Workstation Workstation Video server

INFOCOM’98 980401 11

Important properties

• Active modules in machine code, no per packet

processing overhead -> HIGH PERFORMANCE

• Security addressed by usage of simple and well

known cryptology techniques

– Active modules signed and authenticated. – Keys distributed using DNS security extensions (requires only one key to be installed initially on an ANN)

• Policies

– Acceptance policies – Caching policies

INFOCOM’98 980401 12

Potential problems

• Delay: Download time minimization required and

possible

– Probe packets – Code server arrangement optimization -> Minimizing distance between ANN and code server – Cache misses potentially rare

• Not as flexible as typical “capsule” approaches

– Active modules have to be installed on Code Servers – Code servers can automatically exchange active modules – Best case: Active module manually installed on only

• ne code server

INFOCOM’98 980401 13

Platform

• Distributed Code Caching is only a first step

towards gigabit active networking

• Optimized hardware and software platform

required

• Over past years, we prototyped technological

hardware and software components

– APIC (1.2 Gbit/s ATM host adapter) – WUGS (Washington University Gigabit Switch) – Crossbow (modular, extensible router platform based

• n NetBSD)

INFOCOM’98 980401 14

ANN Hardware

to other ANN

WUGS ATM "Backplane"

. . .

BI - Bus Interface ANPE Cache CPU Memory BI APIC ANPE Cache CPU Memory BI APIC ANPE Cache CPU Memory BI APIC ANPE Cache CPU Memory BI APIC

. . .

A B C D

to other ANN to other ANN to other ANN ANPE - Active Network Processing Element

Active Network Node (ANN)

Load Balancing

INFOCOM’98 980401 15

ANPE Software Architecture

• Implemented on top of NetBSD/Crossbow
• Control-Path components in User Space, Data-

Path components in Kernel

• Kernel:

– Function Dispatcher – Resource Controller

• User Space:

– Active Module Loader – Module Database Controller – Security Gateway – Policy Controller

INFOCOM’98 980401 16

Applications

• Automatic Network Protocol Deployment /

Revision

– especially well suited for IPv6 options

• Large-Scale reliable multicast

– Faster recovery through topology knowledge – Application-specific multicast

• Congestion control for real-time video and audio
• High-performance media gateways for real-time

multicast audio/video sessions

INFOCOM’98 980401 17

Conclusion and Status

• We believe Active Networking is a fascinating

idea worth pursuing!

• Existing systems do not focus on multi-gigabit

environments

• Some restrictions regarding flexibility for the

benefit of performance is a valid compromise

• Submitted proposal to DARPA for a

“Scalable, High Performance Active Network Node”

• Currently negotiating contract with DARPA
• Web site:

http://www.arl.wustl.edu/arl/projects/ann/ann.html