Over at bufferbloat.net we have developed several pretty accurate bandwidth and latency measurement tests, that work at speeds up to 40GigE.
We wrap the popular with the linux-netdev's "netperf" tool with something that can aggregate and plot the results, called "netperf-wrapper". The most popular test in the suite is called "rrul" which stands for "Realtime Response Under Load", but there are many others in the suite.
It has been used to extensively tune several fair queuing and aqm algorithms, notably "fq_codel" which is in cerowrt, openwrt, and many other 3rd party firmwares. Its been used to debug network hardware, wifi, cablemodems, and most recently during the 40GigE batch-bql patchset now entering the linux kernel.
Some examples of use to tune a smarter queue management system against modern day cable modems:
There are also netperf-wrapper results for 40GigE, DSL, and wifi spread around the Internet. The intermediate format netperf-wrapper uses to store its results are in json and parsable by anything, and I keep hoping someone will get around to writing a web interface for the datafiles...
Nothing else I've ever seen even comes close to netperf-wrapper for finding good, accurate, long term numbers and multiple forms of anomoly. Pretty much all the web based tests get increasingly inaccurate above 20Mbits. Single threaded TCP tests are bad also as they generally result in someone defeating TCP congestion avoidance in pursuit of the best benchmark numbers. 
Far more important to the debloaters is not the bandwidth attained but the latency induced while getting it. 
We maintain several public servers for netperf-wrapper, all connected via a gigE connection to the internet. Thus far we haven't overloaded them (nor advertised them widely), but if you want to give netperf-wrapper a try, and can't set up your own netperf server on the other side, feel free to ping us on the bloat mailing list for some addresses on various continents.
 A brief rant: Bandwidth != speed. Bandwidth is capacity/interval. Real perceived speed is obtained via low
 I really hate that all the web network measurement tests don't simultaneously measure ping while running their upload and downloads. IF ONLY those tests would do that, people would start to realize that there is a huge tradeoff between good latency and high bandwidth, and that they are doing their networks in, by optimizing for bandwidth only.
Networks engineered for speedtest's current test, *suck* for voip and gaming. I'd like to petition them to at least report ping times under load to the 98th percentile.