Cell Relay Archive

["ronald h. davis" <ronaldd@lucent.com>]

Patrick Hurley wrote:
> 
> 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?
> 

there are a few papers on the atm forum site which discribe tcp over atm.
most are simulations.  prof. raj jain at ohio state has studied this area
as well.  kent englund is another name that comes to mind so you can do a
search on those names and possibly find some papers.

since i don't remenber the numbers off the top of my head the qualitative
answer is that buffer size defends upon whether you are using abr or ubr.
within ubr it also depends upon whether you are using ppd or epd.  abr
places the least buffer demands upon the atm network but it pushes congestion
out of the network to the edge devices which must have sufficient buffers
to handle bursts.  the amount of buffer depends upon the number of tcp sources
are handled by each edge device.

ubr with no cell drop strategy requires the largest buffers in the atm switch
as you might expect.  for a given buffer size within an atm switch whether
epd provided better tcp performance than ppd depends upon the epd threshold:
too aggressive a threshold results in packets getting discard that might have
actually been successfully transmitted.

specific numbers for buffer size depend upon a number of assumptions such as
tcp window size, rto, &c.

-- 
  __  ______  __  / __/ | 	lucent technologies, naperville il, usa
_/ (_(_) / (_(_/_/_(_/  . 	ronald.h.davis@lucent.com
for there are many great deeds done in the small struggles of life.
life, misfortune, isolation, abandonment, poverty, are battlefields
which have their heros; obscure heroes, sometimes greater than the
illustrious heros.
				-victor hugo