Re: Problems on windows

[Maarten Koopmans <maarten.koopmans --at-- surfnet.nl>]

Nevermind, I find out how to tune iperf to get (more) reliable results:

On XP, 2.02 (Cygwin version) works fine, as long as:
- you set the window size to 96K
- use 4-12 threads
- uset t > 30 to get rid of startup effects.

Your mileage may vary with the window size depending on the latency in
your WAN.

So adding -t 30 -P 4 -w 96K will work for most end users and will test
up to 300 mbit on XP here on SURFnet6. -P 12 will take that further, but
creates more startup effects for ADSL users, so t will need to grow then if you want to cover the complete range of network possibilities for end-users.

I also learnt that the order of the parameters matters(!). -P 4 -w 96K
in the beginning doesn't work, on the end it does.

--Maarten

Maarten Koopmans wrote:

> Hi,
>
> We are using iperf as one of the underlying tools in the SURFnet 
> Detective (detective.surfnet.nl) and I'm seeing some really strange 
> behaviour. In our setup we have a bunch of linux servers on 1 GBit, 
> and when testing from linux clients over the WAN we get between 
> 300-800mbit depending on how many threads, window size etc (using TCP).
>
> Using iperf 2.01 on windows I don't get more than 20 mbit. Using the 
> -P option crashes that version. Nevermind, I downloaded version 2.02 
> based on cygwin, and decided to test some more from home. Home is an 
> ADSL2+ connection 12mbit down, 1mbit up.
>
> Surprisingly, when running iperf 2.02 with multiple threads, my 
> upstream bandwidth grows as I add more threads. With 1 thread it's 
> 0.70Mbit, but with 4 threads and a window size of 512k I get 2 mbit, 
> this climbs to 3mbit upstream with 20 threads. With my ADSL2+  this 
> simply isn't possible (1mbit upstream max), so I suspect "something is 
> wrong".
>
> This puts me between a rock and a hard place:
> 1) multithreaded client (TCP) on WinXp looks unreliable, at least in 
> the ADSL range
> 2) single threaded (TCP) on high bandwidth networks is at least an 
> order of 10 lower on WinXP then the Linux box sitting next to it on 
> high-bandwidth networks.
>
>
> Has anybody else observed this behaviour, or even better, has anybody 
> a solution for this?
>
> Thanks!
>
> Maarten Koopmans
> Manager Middleware Services
> SURFnet