Cell Relay Archive

[murray@pa.dec.com (Hal Murray)]

In article <6hqidv$e08$1@debris.uits.indiana.edu>, Patrick Hurley <patrick.hurley@wco.com> writes:
> Has anyone done research or know of research that exists that studies
> the number of cells that a queue must tolerated before it drops IP
> packets?  
> 
> I have seen distributions where it shows that most (average) bursts are
> for 5 cells and the maximum burst is for 93 or 100 cells.  
> 
> I know of the self-similar paper, that is not what I am looking for.


I suspect you can find burst size data to backup almost any number
you want.

What sort of traffic do you want to support?  CBR/video?  Bursty?
LAN or WAN?

The MTU for IP over ATM is 9K.  That's >180 cells.  If you are going
to send full size packets at full wire speed you will see bursts at
least that long.

If you want to send bulk data (say FTP) at full wire speed, then
you will see bursts of many packets.  Modern hardware/software can
easily do this at 155 megabits.  (Maybe even 622 by now.)

If your traffic is CBR and only using a small fraction of the link
then all your bursts should be short, mostly 1 cell.


I'm not sure if there is a simple way to translate burst size
into buffer size requirements.  Are you interested in buffers
per switch, per link, or per VC?  What sort of switch architecture?

With simple CBR traffic it only takes a few cells per VC.

With simple credit based flow control it takes ~10 cells per VC
for short links plus 3 or 4 per km per vc.  (That's at 155 megabits.)
That's FLOWmaster as implemented in GIGAswitch/ATM, not a standard.

For TCP over UBR, everything will work great (no dropped packets) if
you have enough buffer space for the max TCP window.


-- 
These are my opinions, not necessarily my employers.