Excerpts from Riyaz Shiraguppi's message of Wed Aug 04 09:02:43 -0500 2010:
> Hi,
> 
> I want to get estimation of bandwidth even if packet drops are active.
> 
> I am using network delay box across two machines which can be configured to
> introduce packet drops.
> 
> I am dropping 1 packet out of 1000. Window size is set to 32MB. Line rate is
> 600 Mbps.
> 
> However, I observed that iperf calculated bandwidth goes down as delay
> increases (delay = round trip time between server and client).
> 
> 
> 
> For 0 delay, it is showing 580 Mbps ( no effect of packet drops) which is
> pretty accurate with 600 Mbps line
> 
> For 20 delay, calculated bandwidth drops till  128 Mbps.
> 
> For 30 delay, calculated bandwidth drops till  106 Mbps.
> 
> For 50 delay, calculated bandwidth drops till  50 Mbps.
> 
> 
> 
> Is there any way using iperf to get better bandwidth estimation even at
> active packet drop case?

Iperf is not a bandwidth estimation tool, it's a available throughput
measurement.  That is, it simply uses the TCP implementation on the end
systems and tries to push traffic accross the link.  If you'd like a bandwidth
estimation tool there are several out there, generally based on measuring
inter packet gaps of packet trains.  

For more information about bandwidth esitmation tools do some looking into
pathchar, pathload, pchar, etc.  I also included some information in my
measurement tools overview talk at NANOG43:

http://www.nanog.org/meetings/nanog43/abstracts.php?pt=MjkmbmFub2c0Mw==&nm=nanog43

You can't really do this kind of test in vanilla TCP.  (You might be able to
do some kind of analysis with a specially instrumented TCP, but that's getting
pretty far off the beaten path).

It's is actually expected behavior (if not always intuitive) that TCP will
have lower throughput given the same loss but a longer round trip time (RTT).
This is due to exponential backoff in TCP Reno.  I'll not go into the gory
details here, there's plenty of good information on the net about TCP protocol
dynamics.  

ESnet (my employer) keeps a fairly extensive web site dedicated to high
throughput data transfer in the wide area including information about TCP
tuning and dynamics.  You can find this site at:

http://fasterdata.es.net/

Cheers,

Jon
-- 
Jon M. Dugan <[email protected]>
ESnet Network Engineering Group
Lawrence Berkeley National Laboratory

------------------------------------------------------------------------------
This SF.net email is sponsored by 

Make an app they can't live without
Enter the BlackBerry Developer Challenge
http://p.sf.net/sfu/RIM-dev2dev 
_______________________________________________
Iperf-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/iperf-users

Reply via email to