Testing Bandwidth performance

Wed Jun 26 14:32:21 UTC 2002

Hi Alan,

>What are some tools to test bandwidth perfomance?  I've used iperf, but =
>are there other tools or ways to generate traffic for testing purposes =
>to see a links maximum capacity?  Especially greater than a 100mb.

Iperf can be used to generate OC3+ class TCP flows above if the host 
configuration is correctly tuned.

See, for example, the excellent paper by Stanislav Shalunov at

  http://www.internet2.edu/~shalunov/gigatcp/

...and systems have only gotten better since that was written. For example, 
relative to what folks had available a year or so ago (which still performed
quite well), you can now buy a system like the Supermicro SuperServer 6022L-6 
2U ( http://www.supermicro.com/PRODUCT/SUPERServer/SuperServer6022L-6.htm ) 
off the shelf with:

-- 2x Xeon 2.4GHz/512K L2 cache (Stas's quoted box was working with only
   2x1GHz PIII CPUs)

-- PCI buses which have gone from 64 bit/66Mhz to PCI-X 64 bit/133MHz

-- DDR with 2 way interleaving is now an inexpensive production memory option
   (the quoted box had used SDR)

-- The previous ServerWorks ServerSet III/HE has now been upgraded and
   improved in the form of the ServerWorks GC-LE chipset (see:
   http://www.serverworks.com/products/GCLE.html )

I would also suggest one of the Syskonnect gig cards, rather than the 3Com 
or Netgear gig cards cards. In some configurations we've seen a 300Mbps
increase in throughput over the local area (relative to what an Intel gig 
card delivered) simply by swapping in a Syskonnect SK-9843.

Of course, the real killer is still that the path MTU across most of the
world is still abysmally small, typically 1500 octets. If folks really want
to routinely go fast, they need to be working toward getting 9K frame sizes
supported end to end (but that will be tremendously hard/impossible to make
happen across the generic Internet). 

Regards,

Joe St Sauver (joe at oregon.uoregon.edu)
University of Oregon Computing Center