Implementation Report on Experiences with Various TCP RFCs

1
Implementation Report on Experiences with Various
TCP RFCs

• Murari Sridharan muraris_at_microsoft.com
• Deepak Bansal dbansal_at_microsoft.com
• Dave Thaler dthaler_at_microsoft.com

2
We did a new implementation of a bunch of RFCs

• Started new TCP/IP stack in 2001, shipped in
  Windows Vista.
• Including
• Window Scaling (RFC 1323, PS)
• ECN (RFC 3168, PS)
• DSACK (RFC 3708, EXPERIMENTAL)
• F-RTO (RFC 4138, EXPERIMENTAL)
• Then we did lots of testing of the RFCs above

3
Window Scaling (RFC 1323), 1/3

• Previously in Windows, but off by default
• We ran into a number of problems testing it for
  Vista
• IGD problems multiple popular vendors, popular
  models
• Popular vendor 1, several models having SPI
  block TCP handshake with WS automatic detection
  and recovery at end system
• Popular vendor 2, models with SPI drop packets
  on connections after a large scaling factor (gt2)
  successfully negotiated during initial handshake
  no way to recover at TCP layer
• Firewall problems multiple vendors have almost
  identical bug in old versions
• Multiple popular vendors, multiple old models
  (one only if a non-default feature enabled) drop
  packets if scaling factor gt 2 has been negotiated
• Core Issue Broken firewall logic that does not
  scale window in TCP headers correctly while
  determining how much data to let through

4
Window Scaling, 2/3

• More problems
• Broken websites websites running old versions of
  firewalls in front of their web servers are
  broken when scaling factor gt 2 is negotiated
• A more thorough testing (of over 10 million
  websites) has revealed 2-4 websites impacted by
  this issue
• Proxy problems
• Popular web proxy TCP splicing option (has been
  disabled by default since 2006) has problems
  understanding the scale factor.

5
Window Scaling, 3/3

• Bottom Line
• Problems are fairly likely if used by a consumer
  (not behind a proxy)
• Much less problematic if behind a proxy since
  proxy may not negotiate Window Scaling
• Result Limited to scale factor of 2 in Windows
  Vista for HTTP connections

6
ECN (RFC 3168), 1/2

• New for Windows Vista
• Straightforward to implement
• We ran into a number of problems
• IGD problem 1 one of the most popular versions
  from one of the most popular vendors
• When a data packet arrives with either ECT(0) or
  ECT(1) (indicating successful ECN capability
  negotiation) indicated, router crashed
• Cannot be recovered at TCP layer
• IGD problem 2 multiple popular vendors with
  popular models
• When the 3-way handshake completes, prior to any
  data being sent, router hit a bug causing further
  packet loss
• Easily detected and recovered at TCP layer

7
ECN, 2/2

• Bottom Line
• Problems are fairly likely if used by a consumer
  (bad IGDs exist)
• Much less problematic in campuses and enterprises
  (IGDs less likely to be present)
• Result Turned off by default in Windows Vista

8
DSACK (RFC 3708)

• New for Windows Vista
• Straightforward to implement
• No issues yet
• Now enabled by default
• Actual benefits hard to quantify to a specific
  improvement in real world

9
F-RTO (RFC 4138)

• New for Windows Vista
• Straightforward to implement
• No issues yet
• Now enabled by default
• We have seen improvements in wireless
  environments but we have not isolated the
  benefits to specific enhancements
• We agree with advancing F-RTO to PS

10
Questions?

• Details will be at http//support.microsoft.com/kb
  /934430(not published yet)