Re: [tcpm] Increasing the Initial Window - Notes

[rick jones <[e-mail]>]

On Nov 10, 2010, at 9:00 AM, Hagen Paul Pfeifer wrote:

> * rick jones | 2010-11-10 08:33:30 [-0800]:
>
>> Do your customers run systems with support for TSO and/or LRO
>> enabled, or do they disable that today?  TSO can/may/will cause
>> 10*MSS or larger blobs of data to be injected into the network at one
>> time no?
>
> No, TSO does not change the packet behavior on the wire. At least no  
> CC
> related mechanisms. Packet spacing and stuff like that ignored here.

Earlier you wrote "New data are always pushed into the network in this  
quantum - a way to
much for the standard TCP CC mechanism to tune into a steady state for  
network
with a low BDP."  I read that as implying a concern about how much  
data any one connection injects into the network at any moment in time.

TSO does not change what any one packet looks like on the wire, but it  
certainly does change how many of them rush the wire at one time, yes  
bounded by cwnd - something those effectively infinite data sources  
like FTP will be pushing as far as it will go.  And depending on  
whether the stack receiving the LRO-stretched ACKs is counting bytes  
ACKed or number of ACKs the LRO can induce similar bursty behaviour.

>> And LRO can/may induce stretch ACK behaviour not unlike the
>> ACK avoidance heuristics of stacks such as those in HP-UX or Solaris.
>> What I've seen so far has suggested the goal is to get IW10 "blessed"
>> not to make IW10 mandatory.
>
> If the IETF "blessed" IW10 every vendor will enable it ...

And I presume, make it configurable.  It is already configurable on a  
per-route basis in the Linux stack.

rick jones
http://homepage.mac.com/perfgeek