Re: Iperf or Iperf like test points?

[Jared Mauch <jared () puck nether net>]

On Wed, Aug 24, 2005 at 02:15:08PM -0400, Walter C. Ames wrote:
> With that said, the problem that I am facing is that there are no 
> consistently reliable tools that NetOps (or end users for that 
> matter) can use to truly evaluate bandwidth performance on large pipes.
>
> Ex: All of the test sites that I have tried from a 100M/FD attached 
> Linux box, riding a GigE backbone to multiple GigE transit lines 
> typically yields BW test results in the 3-7Mbps range.  Yet when I 
> Iperf across the backbone I get more reasonable results of between 
> 80-90Mbps TCP.

        So your limiting factor is likely not the network.  I typically
use the 'udp' iperf test when we've been involved with customers
that don't believe they're getting the right bw.

        Here's why:

        1) Some host TCP stacks are old/broken
        2) If there is loss, TCP will be butt-slow, the UDP results
immediately show you the loss.
        3) It removes the system memory/disk from the list of things
to be concerned with.  Many people try to test with FTP or HTTP and
get poor performance, we've been able to consistenly show them our
network performs correctly.
        4) bandwidth test sites don't have a farm of fe/ge connected hosts
lying around waiting to be hammered to death.

> The extent of the problem is that I hand off 10M - GigE connections 
> to my end users and they want a way to test it that is 'Off-Net'.  My 
> on-net test platforms give them great results, however since they are 
> on-net the end users dismiss the results (thinking they are fixed I guess).

        We've not had trouble with customers understanding that we can
only control our network.  

> To date I have not found a reasonable method of accomplishing this.
>
> That being said, is anyone on this list aware of such a formation of 
> Iperf nodes across the net connected at GigE or better to accomplish 
> this goal?  If not I would be willing to start one and give up a 
> server or two and some of my bandwidth to help others out who are 
> probably experiencing (or have experienced) this type of problem in the 
> past.

        We have hosts scattered around our network that we use for
iperf testing with customers when there are troubles.  Most are
fe connected, but some are ge.  We'd rather not see the short
bursty 100m flows across our network unless we're aware of them as
it can easily throw off some of the stats, and also look like a DoS.

> This issue is just burning up a lot of my tech supports time trying 
> to educate the end users.  I just feel that a cooperative effort that 
> yields more accurate and consistent results may be a better way to 
> approach this.

        We've seen the same issue.  People just don't get it and
we've spent a lot of time educating customers.

        - jared

-- 
Jared Mauch  | pgp key available via finger from jared () puck nether net
clue++;      | http://puck.nether.net/~jared/  My statements are only mine.