Re: [ns] TCP traffic generation rate

[Brian Lee Bowers <bowersb@lanl.gov>]

On Tuesday 31 July 2001 14:24, Sunil Gowda wrote:
> Hello everyone,
>     My simulations require TCP traffic in terms of GBps. I was able
> to achieve just 200KBps of traffic per agent when i attached a FTP
> application to the agent. The window size was 1000.

What are the characteristics of your network (bandwidth and delay of 
each link)?  Which version of TCP are you using (Tahoe, Reno, New 
Reno, Vegas)?  Is a window size of 1000 large enough (I routinely run 
sims where a window size of 5183 would only just keep the pipe full).

I just ran a 150 second sim that managed to deliver ~7.6M packets of 
1500 bytes (or 12000 bits) each.  That is an average of over 600Mbps 
on a 622Mbps bottleneck link.  I needed 62.2Mb to keep the pipe full 
(622Mbps bottleneck link with a 100ms RTT) so a window size of 1000 
packets would not have been enough for my sim (your actual mileage 
may vary based on bottleneck bandwidth and RTT for your sim).

> How can i generate such large scale traffic?

I'm not sure the problem is with the traffic generator.  Without 
knowing more about your network, I can't say if your network setup is 
to blame.  I'm getting more than a few hundred Kbps from my traffic 
generators (using FTP).

> What is the maximum rate that can be achieved?

Well, I've seen a 622Mbps bottleneck link reach saturation from a 
single sender.  Never tried to go much higher.

> What is the largest window size that we can have?

Well, the variable maxcwnd_ is an int, so I would have to guess about 
2,000,000,000 (give or take several hundred million).  That number is 
in packets, so it really is quite large.  There is also a variable, 
wnd_, that is a double.  I've forgotten the maximum value an IEEE 
compliant double can hold but it is pretty doggone big - I should 
probably wander off to google and find the value, but I'm being lazy 
AND snobbish today :)

> Also is the window scaling option implemented in ns?

It really isn't.  ns uses sequence numbers to count packets, not 
bytes.  I've yet to run a sim that got close to 2,000,000,000.

> Connecting thousands of agents is not a viable option as the
> simulation will be very slow.

I normally run with 100 senders on one side of the bottleneck and 100 
receivers on the other side of the bottleneck, or I run with 200 
nodes, half on one side and half on the other side of the bottleneck, 
with every node being both sender and receiver.  It really isn't that 
much slower than running with one sender and one receiver.

> Any help is highly appreciated.

More information please.  You haven't given enough for anyone to make 
more than a wild guess about the problem.

> Thanks in advance

You're welcome (for what, I have no idea).

> Sunil

-- 
Brian Lee Bowers	|	RADIANT Team (Summer Intern)
bowersb@lanl.gov	|	Los Alamos National Laboratory