Understand parallelisation

[LucaMantani]

I have now performed many different tests, both in my laptop and in the hpc, and using MPI seems to consistently yield slower results.

ultranest: 4.3.1
mpich: 4.2.1

I have also tried downgrading ultranest to 3.6.4, but still the test was unsuccessful.

I have been testing with this custom snippet, to make sure it's not something we do in colibri:

import ultranest
import time
import ultranest.stepsampler as ustepsampler


def log_likelihood(params):
    return -0.5 * sum((params - 0.5) ** 2 / 0.1**2)


def bayesian_prior(cube):
    return cube


Nparams = 50

parameters = ["w%i" % i for i in range(Nparams)]

sampler = ultranest.ReactiveNestedSampler(
    parameters,
    log_likelihood,
    bayesian_prior,
    log_dir="ultranest_log",
)

sampler.stepsampler = ustepsampler.SliceSampler(
                generate_direction=ultranest.stepsampler.generate_mixture_random_direction,
                nsteps=2*Nparams,
            )

t0 = time.time()
ultranest_result = sampler.run()
t1 = time.time()

print("ULTRANEST RUNNING TIME: %f" % (t1 - t0))

# sampler.plot()

[comane]

Any update on this?

[LucaMantani]

Last time I did was when I opened this issue, so it was not working as expected.

[LucaMantani]

This was also experienced by @tgiani, what it seems to be happening to me is that parallelisation is already happening without the need for MPI explicitly, at least when running on a single machine. It might be that this does not happen if one wants to parallelise over multiple CPU/GPU units, this stuff would need careful checks but I do not think it is strictly needed before the release of the code.

[LucaMantani]

This issue is still unclear to me as I get behaviours that change with environments (I have tried colibri and smefit on the script in the first post). With further digging, chatGPT indicates that the two environments use different "threading" because colibri uses "OPENBLAS", while smefit uses "Accelerate" which is something MacOS native.

The thing that would be interesting to understand is whether the behaviour of openblas is convenient or if it would be better to have explicit control over parallelisation, because openblas is deciding on what to do and how many cpu to use in a dynamic way apparently.