I don't know how they calculate bandwidth, but I was surprised that their system
gave such wrong results under what were effectively lab conditions.
The only reliable way to really test performance is to saturate the
pipe (Iperf) and have a sufficiently well provisioned target. NDT does
a good job using short non-saturation tests, but it is susceptible to
slow start and other challenges. In general, NDT results will be more
conservative than best case, whereas a lot of other tests are very
optimistic best cases.
FWIW, the actively maintained code has moved to:
https://code.google.com/p/ndt/ and 3.6.4 is a bit more stable and
flexible on some platforms than 3.6.5.
You can either standup your own test server or point at the public
sites run by a few universities and MeasurementLab
(http://www.measurementlab.net/mlab_sites), which are not as widely
distributed as the Ookla / speedtest.net targets, but they tend to be
better provisioned and the result data for the Mlab targets is made
available to the public.
Once you've compiled the client, you can run again the closest host via:
$ web100clt -n ndt.iupui.donar.measurement-lab.org
(default install will put the test client in /usr/local/bin)
If anybody want to host Mlab collection servers, they're always
looking for more hosts (http://measurementlab.net/getinvolved).
It'd be nice to know if NDT was not accurate as well. Anyone tested it?
We've been using it for a few years. On my laptop that runs linux I get fairly consistent results (around 935Mb/s up and down right now) over a 1Gig routed link (a couple routers and a firewall in between.) On the Windows boxes I usually see a 100 to 200 Mb/s drop on the upload side. The last time I checked, you can compile a commandline version of the client. I seem to remember the commandline client not taking quite as bad a hit on the tests compared to running it on linux, but it's been a while since I tried it.
For us it's been way more accurate than the various speedtest servers our customers insist on trying. A while back I switched from compiling my own kernel and NDT to using perfSONAR-PS (http://psps.perfsonar.net/). I like that they've got live-cd and net-install versions. If nothing else it's useful for pointing out the difference between a local network issue and Internet Suckage.
[Plug alert]
For longer term monitoring, Project BISmark provides an easy-to-use
system. It's an open source, customizable OpenWRT-based home router that
runs periodic network measurements (latency, throughput, packetloss,
jitter, etc) to nearby MLab servers.
It uses netperf (single and multiple TCP threads), and shaperprobe,
(UDP) for throughput measurements.
Although its original target audience is home users, it can also be used
as a monitoring tool in bigger networks.
More information here:
http://projectbismark.net/
Slides from the talk at NANOG 53:
https://www.nanog.org/meetings/nanog53/presentations/Monday/Sundaresan.pdf
- Srikanth