I find a way to trigger a remoting problem using tcp fault injection with netem. I'm able to trigger this wait call at hudson.remoting.Request.call(Request.java:146):

while(response==null && !channel.isInClosed())
  // I don't know exactly when this can happen, as pendingCalls are cleaned up by Channel,
  // but in production I've observed that in rare occasion it can block forever, even after a channel
  // is gone. So be defensive against that.
  wait(30*1000);

When this wait is triggered, the running build is stuck and consumes a executor. It loops over and over on the wait.

To reproduce, setup a SSH slave using the attached Dockerfile, and setup netem on the docker0 bridge like this:

tc qdisc add dev docker0 root netem
tc qdisc change dev docker0 root netem corrupt 1

Testing requires to run the job one time before configuring netem, as netem settings are applied to all network streams, it could fail while downloading Maven dependencies. I just launched a Maven build of a example project to trigger the problem. It might be a Maven specific problem...

To remove netem settings, just run tc qdisc del dev docker0 root.

I've attached the Dockerfile, the command I used to launch it and a threaddump of a Jenkins stuck master.