[AutoTVM][RPCRunner] timeout is not passed correctly

[bernhardklein]

If we hand-over the timeout to the constructor of the LocalExecutor, like done in the LocalBuilder class it seems to solve the issue and the timeout is now correctly passed.

[bernhardklein]

relates to this 6924 issue.

[comaniac]

LGTM

cc @merrymercy

[merrymercy]

We cannot do this because it changes the meaning of the current timeout.
Currently, the timeout argument of RPCRunner only controls the time spent on remote devices. But after your change, the timeout will take all time into consideration, which includes the waiting time to get the device.

For example, if the timeout is 10, we have one remote device and we launch two measurement tasks.
With the existing code, each of the two requests can get 10 seconds to do the measurement on the remote device. So the later request actually has a 20 seconds timeout in LocalExecutor (10s for waiting for the first request to finish).
But after your change, the later request can only get 10 seconds in LocalExecutor.

The correct way is

self.executor = LocalExecutor(timeout=(timeout * self.n_parallel + 1))

We use self.n_parallel+1 instead of self.n_parallel to give more time for other overhead.

[bernhardklein]

@merrymercy as you suggested I set the constructor argument to (timeout * (self.n_parallel + 1)), also changed the self.timeout in the same way. Hope this is the correct way to do it. For n_parallel=None the base class Runner sets to cpu count, which seems to be working correctly. Hope now it's ok.

[merrymercy]

You should only change the timeout setting in LocalExecutor.
The self.timeout is used in other places and it is correct.
I helped you fix the problems.

[bernhardklein]

ok, thank you!