Questions about Internet Packet Losses

John Hawkinson jhawk at bbnplanet.com
Tue Jan 14 13:19:21 UTC 1997


[wow, I clearly went to sleep to early]

> (2) is a random distribution plus some spikes, plus some modality due to
> TCP's that can only generate packets of <= N bytes for different popular
> MSS's.  So I think it still works (but will save more detailed analysis
> till later ....).

When Tony first posted his assertion I sent him a (!) message and then
did a sanity check myself, with 1000 SYNs off a transit FDDI:

I'll repeat it here; 935/1000 SYNs had MSS options, here they are. I just
used tcpdump and didn't bother to pull off the wscale stuff since it 
seems like it might be interesting.

Certainly my data shows a lot more 1460s than I would have expected,
which makes me happy.

 594 1460
 128 512
  90 536
  23 4312
  18 512,nop,wscale
  15 1460,wscale
  14 536,nop,wscale
  10 1460,nop,wscale
   8 966
   6 4096
   5 256
   4 472
   3 1024
   2 212
   2 1500
   2 1464
   2 1440
   1 768
   1 660
   1 460
   1 4378
   1 216
   1 1500,nop,wscale
   1 1450
   1 1420
   1 1410

Certainly these numbers shouldn't be asssumed to be representative,
however 512 is unfortunately a serious contender.

> > 2) What OS is using a 512 MSS?
> 
> From a bunch of (~ 20,000) traces in my study, looks like
> 
>         Irix 4.0/5.3, BSDI 1.1/2.0, OSF/1 3.0/3.2, and
> 	often but not always SunOS
> 
> use either 512 or 536, or sometimes 500, 524, or 548 (BSDI).  Never
> anything higher.  This doesn't mean they *can't* use anything higher,
> just that they never found an opportunity to do so.

Well, my SunOS machine offers an mss of 1460 because I decided that it
as long as it wasn't going to use Path MTU Discovery, and the vast
majority of links out there have MTU's >=1500. Under SunOS, this is
controlled by tcp_default_mss, just change it in
/sys/netinet/in_proto.c or adb your kernel. Of course, this means that
you will may cause fragmentation at links with MTU's less than 1500.
Hopefully that is isolated to high-latency dialup links with adequate
buffering.

This is not a problem/solution unique to SunOS, but it probably is one
that hasn't been encouraged very much...

Perhaps it should be?

--jhawk





More information about the NANOG mailing list