Making a Cheaper Bridge: Using Intel IXP2400 Network Processors

1
Making a Cheaper Bridge Using Intel IXP2400
Network Processors

• Paul Royal paul.royal_at_cc
• Joe Uhl gte398u_at_cc

2
Problem Context Delta TPF

• What is a TPF?
• Transaction Processing Facility OS used on
  legacy IBM Mainframes (circa 1970s)
• OK, who cares?
• Despite age, reliable and represents significant
  investment
• Delta still uses a TPF for its operations
• Obviously computationally limited when compared
  to modern hardware
• Amazing new features like TCP Stack introduced in
  1999
• So they cant replace it how does it interact
  with the rest of the world?
• For Delta, some TPF-only operations offloaded to
  modern equipment and software
• Communication with TPF arbitrated through custom
  ASICs (Application Specific Integrated Circuits)
  that mimic things like an IBM 3490 tape device
• Combination of tape drive mimicry and event
  repetition used to maintain correct state

3
Problem Description

• So whats wrong with how things are?
• Integrating new services with TPF is difficult
• Need another way to perform integration besides
  making new ASICs
• ASICs are expensive
• More ASICs will be needed as additional
  resources/services are offloaded from the TPF
• Current ASICs have limited adaptability
• Which brings us to the alternative Use a cheaper
  and more flexible, but as scalable bridge in
  place of the ASICs

4
The Bridge Intel IXP2400

• What is an IXP 2400?
• Programmable Network Processor
• Gigabit Interface
• Connected via PCI slot in host machine
• Important Note Most processing should occur on
  IXP, not host machine
• Capable of parallel execution
• XScale core and 8 MicroEngines (2x4 Clusters) at
  600 Mhz
• Architecture resembles IBM cell processor
• Complex memory hierarchy to assist parallel
  execution
• Microengine registers, Next-to-neighbor
  registers, Core scratch memory, QDR SRAM, DDR
  DRAM
• Complex enough to run embedded Linux

5
IXP2400 Architecture Diagram
6
Our Focus

• Temporal Joins/Splits
• Situations where
• Single request to IXP must be split to ltTPF, Web
  Resourcegt as appropriate
• Two responses/incoming packets need to be joined
  into one and sent to one or more systems
• Considerations
• Some splits/joins may take priority over others
• Request(s) may be atomic (one packet) or
  fragmented
• A time window on response time (response/request
  must be sent within x time units) may exist
• Goal Program IXP2400 to handle these
  splits/joins in an efficient and scalable manner

7
References

• Y. Kim, Next Generation Network Processor
  Development Tools, Intel Communications Group,
  2002, www.uniquest.co.kr/kor/addFiles/download/
  presentation_for_IXP_series.ppt
• V. Oleson, G. Eisenhaur, C. Pu, K. Schwan, B.
  Plale, D. Amin, Operational Information Systems
  An Example From the Airline Industry, Georgia
  Institute of Technology, 2000.
• J. Thanigachalam, Working on the IXP2400,
  University of Wisconsin, http//wail.cs.wisc.edu/
  ixpdev/ixp.html.
• H. Vin, Network Processors Intels IXP-family
  of Processors, University of Texas at Austin,
  http// www.cs.utexas.edu/users/vin/Classes/CS395T
  -Fall03/Slides/11.pdf