dadi-cli runs take very long

[Maycon Oliveira]

Hey, I was wondering, what is the expected runtime for dadi-cli in general? For example, I'm running a 2D split with secondary contact model with 32 cores for more than 24h now and it still only output 39 runs to the results file. Is this within the expected runtime for the pipeline? Is there a way to improve runtimes without sacrificing too much in terms of inference potential?

[Ryan Gutenkunst]

Hello Maycon,

It depends strongly on the sample size and number of populations. For typical sample sizes that does seem slow. Another important factor is to set the parameter bounds reasonably. Small population sizes and long divergence times are very expensive to compute. How do you have those set?

Best,

Ryan

On Mar 4, 2026, at 10:17 AM, Maycon Oliveira <oliveira...@gmail.com> wrote:

Hey, I was wondering, what is the expected runtime for dadi-cli in general? For example, I'm running a 2D split with secondary contact model with 32 cores for more than 24h now and it still only output 39 runs to the results file. Is this within the expected runtime for the pipeline? Is there a way to improve runtimes without sacrificing too much in terms of inference potential?

--
You received this message because you are subscribed to the Google Groups "dadi-user" group.
To unsubscribe from this group and stop receiving emails from it, send an email to dadi-user+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/dadi-user/0722868e-dedd-4054-b6f2-addd08b3b39fn%40googlegroups.com.

[Maycon Oliveira]

Hi Ryan, thanks for the reply! I'm setting my lower bounds at 1e-3 1e-3 0 0 0 and the upper bounds at 100 100 10 10 10. I'm working with sea turtles so I expect divergence times to be long.

[Ryan Gutenkunst]

Hello Maycon,

Note that there’s a degeneracy in population genetics… If you have a long split time with a small population size, then the population will settle back into equilibrium and become very insensitive to the parameter changes. Have you fit 1D models, what do those suggest for size changes and timing?

Best,

Ryan

To view this discussion visit https://groups.google.com/d/msgid/dadi-user/5d794d61-1b22-4a0d-9d6b-76d558136d66n%40googlegroups.com.

[Maycon Oliveira]

I haven't tried anything with 1D models yet, good call. I'll take a look at those, while also testing the 2D models with some narrower bounds.

[Maycon Oliveira]

Ryan, I have another question regarding models: is it possible for a more complex model (secondary contact with symmetric migration) have a equal or even slightly lower log likelihood than a simpler model (continuous migration)? In both cases, model runs converged. I tried 100 runs for each model, with global optimization on.

[Ryan Gutenkunst]

Hello Maycon,

If the models are nested, meaning that more complex model can reproduce all scenarios produced by the simpler model, then mathematically the true maximum likelihood for the more complex model must be greater than or equal to that of the simpler model. (Since the more complex model can at the very least reproduce the result from the simpler model.)

If the differences are small, then you may just be seeing imperfections in the optimization and convergence heuristics.

Best,

Ryan

To view this discussion visit https://groups.google.com/d/msgid/dadi-user/c17220b7-9961-4f84-8f87-28668014dce0n%40googlegroups.com.

[Maycon Oliveira]

If that's the case, should I run the models again with broader parameter bounds or a higher number of replicates? The biggest difference I saw was 60 likelihood units between the more complex (lower likelihood) and the simpler (higher likelihood) model.

[Ryan Gutenkunst]

Hello Maycon,

If the bounds of the more complex model were a superset of the less complex model, then that suggests that it’s a convergence issue, and more replicates would be helpful.

Best,

Ryan

To view this discussion visit https://groups.google.com/d/msgid/dadi-user/aa18f438-d4d8-4996-8a41-de2b8f3c8b0en%40googlegroups.com.