The IP over ATM Mailing List Archive by date

[Curtis Villamizar <curtis@ans.net>]

In message <9506281755.AA07973@netz.eng.sun.com>, Bill Zaumen writes:
> 
> In a previous message on this subject, Tim Salo asked the following:
> 
> > What, if any, support were you using for the TCP window scale option
> > under Solaris 2.4?
> > 
> > What choices do we have, beyond getting source code and implementing
> > it ourselves, for using the TCP window scale option under Solaris 2.4?
> 
> We used a script to set the window size for testing, and I've included
> it in case anyone will find it useful.  I also checked with the TCP
> implementors, who are aware of the option, but I really can't say
> anthing about when or whether it will be implemented because of company
> policy (and because I'm not in the group that handles the TCP
> implementation). 
> 
> Here is the (unsupported) script (there must be a space in the command
> between "size" and the value):


Bill,

We can all write trivial scripts like this, but apparently not
everyone is aware of the fundemental limitations of TCP without use of
the window scaling option in RFC-1323.  Without RFC-1323 extensions,
the maximum window size is 64 KB.  If you send a whole 64 KB without
any acknowledgement under ideal conditions and nothing is lost, and
then wait one round trip time (RTT) before sending another 64 KB, then
your maximum bandwidth is 64KB dived by the RTT.  For a US cross
continent circuit, the RTT is about 70 msec (68 on our backbone to be
exact).  If you approximate it at 64 msec, this yields 64 MB / 64 msec
* 8 bit/byte = 8 Mb/sec.  This is your theoretic upper bound, so
expect a bit less than that.  With RFC-1323, you can already do much
better than that (in practice, not just theoretically) on a DS3
circuit on the real Internet (at least you can on our backbone).

If you want to get 100+ Mb/s on an OC3 rather than 8Mb/s or less, then
you need a larger window than 64KB.  Therefore you need window scaling
and RFC-1323.  I can't overemphaisze this.  People at well connected
sites are eliminating Sun from consideration for this reason.

If you do your testing with no delay, you get very different results
than if you test on a real network that has a delay simply due to the
distances involved.  Get yourself a delay simulator and try it.

Curtis