name mode size
LICENSE 100644 35.14kB
README.md 100644 1.3kB
analyze_all.sh 100755 650B
analyze_netspeed 100755 6.78kB
analyze_now.sh 100644 6.78kB
lognetspeed.sh 100755 228B
README.md
# automatic_wondershaper_wombat A script in bash (with some assorted tools) to collect info about available bandwidth and to perform traffic shaping requires speedtest (python or perl version) Python: https://github.com/sivel/speedtest-cli Speedtest-cli logs in this manner: Mon 1 00:01 Ping: 16.828 ms Download: 14.05 Mbit/s Upload: 1.02 Mbit/s Perl: https://metacpan.org/pod/App::SpeedTest Perl App::SpeedTest logs in this manner: $ speedtest -1Qv0 DL: 40.721 Mbit/s, UL: 30.307 Mbit/s In my tests, the Perl module reports significantly *slower* speeds across the board than the python version. For comparison, these tests were done sequentially with no other traffic on my network (yeah, I don't have the fastest internet, which is why this is important to me): speedtest (perl) - Up:13.120 Down:0.849 speedtest-cli (python) - Up:16.64 Down:0.91 speetest.net (web) - Up:17.4 Down:1.06 speedguide.net (web) - Up:16.61 Down:0.87 So if you're interested in matching with web-based tools, go with the pythonic version. If you're interested in *ensuring* you don't over- allocate your bandwidth, the perl version will be better. Interestingly, the perl version is pretty close to 80% of the bandwidth reported by other tools, which is what you need for traffic shaping.