Set switch interval in a loop with increasing numbers #63
Conversation
After talking to Pablo and Thomas about this, they suggested that we start at a 1ns switch interval and go up to 100ns for every time setting it fails. After three failures, we go back to original switch interval. This makes sure that the tests will run in parallel, even if the switch interval is not changed.
| failed = None | ||
| barrier = threading.Barrier(n_workers) | ||
| original_switch = sys.getswitchinterval() | ||
| new_switch = 1e-6 |
There was a problem hiding this comment.
| new_switch = 1e-6 | |
| new_switch = 1e-9 |
The PR description says nanoseconds
There was a problem hiding this comment.
Oh yeah, sorry. Meant microseconds instead. The plan was to start with the original value and increase by an order of magnitude if it fails.
There was a problem hiding this comment.
Just out of curiosity - when can this code get triggered? Or I guess said another way, when can sys.setswitchinterval fail? I've never had any issues setting the switch interval to a small number.
There was a problem hiding this comment.
My understanding is that on i386 and i586 (and especially the latter), there has always been some weird behavior around floats. This can lead to a rounding error, where the C implementation of sys.getswitchinterval effectively sees a 0 and raises a ValueError.
2 participants
After talking to Pablo and Thomas about this, they suggested that we start at a 1ns switch interval and go up to 100ns for every time setting it fails. After three failures, we go back to original switch interval.
This makes sure that the tests will run in parallel, even if the switch interval is not changed.