MK's NDT download breaks with M-Lab's new platform

[bassosimone]

  [112%] elapsed: 15.622 s; test_id: 4; num_flows: 1; speed:    3.6 Mbit/s
  [113%] elapsed: 15.875 s; test_id: 4; num_flows: 1; speed:   3.54 Mbit/s
  [115%] elapsed: 16.126 s; test_id: 4; num_flows: 1; speed:   3.48 Mbit/s
  [117%] elapsed: 16.381 s; test_id: 4; num_flows: 1; speed:   3.43 Mbit/s
  [119%] elapsed: 16.634 s; test_id: 4; num_flows: 1; speed:   3.38 Mbit/s
  [121%] elapsed: 16.889 s; test_id: 4; num_flows: 1; speed:   3.33 Mbit/s
  [122%] elapsed: 17.142 s; test_id: 4; num_flows: 1; speed:   3.28 Mbit/s
[!] netx_recv: netx_recv_nonblocking() failed: timed_out
[!] run_download: receiving: timed_out

The new platform does not close the connection at the end of the download, which was what the old platform did. The above log is from libndt. The NDT client in MK behaves much worse than libndt: it takes a ten seconds timeout and then fails the test.

I initially reacted by trying to fix MK code in 605d7ae, then later I considered switching to libndt (#1881) and started working to this end (measurement-kit/libndt#108).

Meanwhile, I reported the issues that I discovered in m-lab/ndt-server#160 and m-lab/ndt-server#161. Changing the ndt5 server to close the connection will fix the main issue. Having more variables will make the results more useful than just providing a bunch of zeroes to the user.

[bassosimone]

We're starting to fix these issues in m-lab/ndt-server#162, where we have a branch that works with a vanilla version of measurement-kit therefore fixing m-lab/ndt-server#160.

[bassosimone]

This has been fixed and deployed in production! (🙏 @stephen-soltesz!)