Conversation
Default runners have 2 CPUs. We can achieve a nice speed-up (but not quite linear) by leveraging runners with more CPUs.
|
With the 4 CPU runners (sample size of 1):
Perfect scaling of machine time would be 4486m. Since cost scales linearly with CPU count, this change will increase costs by ~18%. #1022 will help with CPU scaling by making PGO+BOLT run tests in parallel. So probably looking at a 10-15% cost premium after that merges. IMO this change is worth it since people time costs more than machine time and faster CI enables people to be more productive. I haven't tested 8 CPU runners thoroughly. But they do yield additional speed-ups. |
|
I defer to @zanieb who knows more about our Depot setup and costs, but I would say yes, the turnaround time for getting a Python release out is very important and currently that requires two CI runs (on PR and on HEAD) so paying to get that cut down seems valuable. (This is probably overengineering but I would likely be in favor of using even larger runners conditionally if the commit bumps a CPython patch version, for the same reason.) |
Default runners have 2 CPUs. We can achieve a nice speed-up (but not quite linear) by leveraging runners with more CPUs.