Dask simulator

[erinaldi]

Hello,

I would like to understand what the DaskSimulator is actually doing when store.simulate() is called and there is a LocalCluster object setup.
Right now I observe (in the Dark dashboard) that there is usually activity on all the workers are the beginning, but then most of the work is done by a single worker for 90% of the total time.
Is this the expected behavior?

Thank you.

[bkmi]

@maxwelltsai @fnattino @meiertgrootes @rogerkuou do you have any thoughts on this?

[fnattino]

Hi @erinaldi - apologies for the very late reply, I completely missed the message (and @bkmi's ping!).

In case this is still relevant to you, the issue here could be the setting of batch_size, which regulates the size of the batches employed to run the simulations (see also #97 and #108). The default when calling store.simulate() is that all simulations are run in a single batch. This leads to all simulations being run sequentially on a single worker, which is probably what you have observed. Setting batch_size to ~ (total number of simulations) / (number of workers) or smaller should lead to the desired behaviour in which simulations are split and run across the various workers.

[erinaldi]

Hi @fnattino

Thanks for the reply. Using batch_size might be indeed what is needed.
I will try it out.