I have a model that is functionally identical to Bayesian Estimation Supersedes the T-Test — PyMC3 3.10.0 documentation
When I feed it ~1M observed rows sampling proceeds swiftly, at a rate of something like 2s/draw. Then, as we approach ~95%, it slows to something like 75s/draw. I’ve got buckets of available ram, CPU capacity, and disk I/O to spare. It appears that this stage proceeds with only a single core, based on utilization. The “fast” stage fully utilizes my processors, as specified. I tried specifying cores=actual-1 to give it some room but it didn’t change the result.
Are there any “gotchas” I should be looking out for that would explain this behaviour, or perhaps is this just a visual/reporting issue?
The only weird thing I’m doing is specifying return_inferencedata=True.
x86_64 ubuntu in a Oracle Virtualbox virtual machine running on Windows 10 w/ PAE/NX and Nested VT-x/AMD-V enabled.