A major implementation problem with implicitly parallel languages, is that small grain size can lead to execution overheads and reduced performance. We present a new granularity-analysis scheme that produces estimators, at compile time, of the relative execution weight of each procedure invocation. These estimators can be cheaply evaluated at runtime to approximate the relative task granularities, enabling intelligent scheduling decisions. Our method seeks to balance tradeoffs between analysis complexity, estimator accuracy, and runtime overhead of evaluating the estimator. To this end, rather than analyze data, size or dependencies, we introduce iteration parameters to handle recursive procedures. This simplification facilitates solving the recurrence equations that describe the granularity estimators, and reduces the runtime overhead of evaluating these estimators. The algorithm is described in the context of concurrent logic programming languages, although the concepts are applicable to functional languages in general. We show, for a bench­mark suite, that the method accurately estimates cost. Multiprocessor simulation results quantify the advantage of dynamically scheduling tasks with the granularity information.