In my cluster, I am running some experiments that uses the network bandwidth aggressively. (It's on Cent OS 7 Linux)
Right after starting up the experiment, the machines use their maximum network bandwidth. After a short period of time, their network speed drops significantly (it becomes 20~25 times slower).
After long period of time, when most of machines terminate, the remaining few machines are still struggling with the slow network speed.
I suspect the TCP's congestion control, so am cosidering trying turning it off.
How can I completely turn off TCP congestion control?