Is increasing granularity a task-specific process, or are there ways that work to increase the granularity of any given task?

How does one determine what is the best granularity for a program like this other than just trying different numbers out?

@victor Most of the time the types of programs that we parallelize are the ones the perform a large number of arithmetic operations on a large set of data. Therefore, you can imagine that we can split this large data into tasks of different sizes in any way we want.