How to scale out to more workers for a fan-out / fan-in function?

[SaxonDruce]

Hi,

I have a single orchestration function which fans out 100 activity functions in parallel, then collates the results.

Each activity does some heavy numeric processing, running for about a minute at 100% CPU single threaded.

Ideally, all 100 tasks would run in parallel and so the final output of the orchestration could be available in one minute (plus some overhead). I expect it to take a while for the workers to be scaled up and so it won't be quite as perfectly parallel as 100 tasks in 1 minute. However I am only seeing about 20 workers being allocated for the tasks, and so overall it takes about 5 minutes to get through the tasks.

Is there something I can do to allow the workers to scale up more quickly to meet the number of tasks?

I am using the Consumption plan, and so since each worker only has a single core, there isn't any benefit to allocating multiple tasks to a single worker at a time - each is using 100% CPU and so they will just cause CPU contention with each other. So I have set maxConcurrentActivityFunctions and maxConcurrentOrchestratorFunctions to 1, which did seem to help a bit.

I've run tests with varying numbers of tasks, and always end up with less workers than tasks, eg:

10 tasks: about 5 workers, about 2-3 minutes total run time
100 tasks: about 20 workers, about 5 minutes total run time
200 tasks: about 30 workers, about 7 minutes total run time

The following shows the number of workers I saw for 5 runs of the orchestration at each of these task sizes:

Is there anything else I can adjust to improve the throughput of the tasks?

Thanks,
Saxon