Closed adrn closed 2 months ago
I'd be surprised if this is true in general. If so, we should definitely figure it out! The emcee pool can do some load balancing, but it doesn't look like that's what's happening here. I know that @rodluger managed to get a significant speedup by specializing the pool to pass only numpy objects, but that wasn't ever added to emcee. It would be interesting to know where the bottleneck is.
Is there a huge speed difference?
See: https://github.com/gnarayan/WDmodel/commit/a65c676e0b2d055939d54ebc5d1744c0e1f25780