IFT Proposal

1
IFT Proposal

• Changes Since Atlanta

2
Feedback During and after Atlanta

• Adaption indication for TCP is a requirement
  (SCTP is example).
• It is desirable to enable software
  implementations over existing TCP implementations
  with best possible efficiency.

3
Major Changes to IFT

• Add adaption indication.
• Make CRC optional and negotiated.
• Allow speculative placement.
• Allow negotiated markers if both ends consent.

4
Optional CRC

• Although CRC has low cost for custom hardware
  implementations, it is high cost for software
  implementations.
• Many users will not require added protections of
  CRC. Examples exist of widely used protocols
  over TCP without added CRC.

5
Speculative Placement

• Pros
• Allows direct out of order placement by server
  side without requiring client side to insert
  markers or enforce alignment.
• Allows software client on existing TCP
  implementation to talk to optimized hardware
  server in a way that is optimal for both ends.

6
Speculative Placement

• Cons doing out of order placement requires the
  following
• NIC must track portions of buffer that have been
  written so as not to allow valid data to be
  overwritten by invalid speculative data.
• Application interlock required so NIC is notified
  when a buffer may be rewritten.
• Application must ignore unwritten portions of
  buffers as they may be trashed.

7
Speculative Placement

• May or may not be a good idea on balance.
• Interesting to consider, though.

8
Negotiated Markers

• May not make sense to do both Speculative
  Placement and markers. Markers are not needed if
  SP is used.
• Are markers really a win without guaranteed
  alignment?
• If not, then markers should be used only when
  BOTH ends of a connections are RDDP custom
  hardware. Otherwise not negotiated.