lowpri - RTO for unidirectional reciever

lowpri - RTO for unidirectional reciever

[Rick Jones]

Folks -

I was going through old email to do a status report and was reminded of 
a discussion in netperf-talk that might be of interest to people here 
since it involved a Linux receiver:

http://www.netperf.org/pipermail/netperf-talk/2008-November/000478.html

A Mac OSX bug was found/presumed, but there was an interesting behaviour 
on the linux receiver side as well - the rtx of the ignored FIN happened 
after what IIRC is the initial RTO.  It suggested that the ACK of the 
SYN|ACK didn't update the rtt estimate or if it did, didn't so so very 
much.  Bug? Feature?  No idea, but food for thought.

rick jones

Re: lowpri - RTO for unidirectional reciever

[David Miller]

From: Rick Jones <rick.jones2@hp.com>
Date: Mon, 08 Dec 2008 18:02:01 -0800

> A Mac OSX bug was found/presumed, but there was an interesting
> behaviour on the linux receiver side as well - the rtx of the
> ignored FIN happened after what IIRC is the initial RTO.  It
> suggested that the ACK of the SYN|ACK didn't update the rtt estimate
> or if it did, didn't so so very much.  Bug? Feature?  No idea, but
> food for thought.

We should not use the SYN|ACK for RTT estimates.

It's a bug, and in fact we used to, but I fixed that long
ago.

Since there is no data in the packet, and it's therefore very small
compared to real data bearing packets, the RTT measured might be
artificially small and not really representative.

Re: lowpri - RTO for unidirectional reciever

[Rick Jones]

David Miller wrote:
> We should not use the SYN|ACK for RTT estimates.
> 
> It's a bug, and in fact we used to, but I fixed that long
> ago.
> 
> Since there is no data in the packet, and it's therefore very small
> compared to real data bearing packets, the RTT measured might be
> artificially small and not really representative.

A miracle occurs going from zero bytes of payload but several bytes of 
TCP options (scale and MSS) to having one byte of payload and no bytes 
of options? (Timestamp options remaining the same in both cases)

rick jones

Re: lowpri - RTO for unidirectional reciever

[David Miller]

From: Rick Jones <rick.jones2@hp.com>
Date: Tue, 09 Dec 2008 11:25:19 -0800

> A miracle occurs going from zero bytes of payload but several bytes
> of TCP options (scale and MSS) to having one byte of payload and no
> bytes of options? (Timestamp options remaining the same in both
> cases)

All it takes is ~40 bytes on an extremely slow link.