DistributedIters

Usage

use DistributedIters;

This module contains iterators that can be used to distribute a forall loop for a range or domain by dynamically splitting iterations between locales.

config param debugDistributedIters: bool = false

Toggle debugging output.

config param timeDistributedIters: bool = false

Toggle per-locale performance timing and output.

config const infoDistributedIters: bool = false

Toggle invocation information output.

iter distributedDynamic(c, chunkSize: int = 1, numTasks: int = 0, parDim: int = 1, localeChunkSize: int = 0, coordinated: bool = false, workerLocales = Locales)
Arguments:
  • c : range(?) or domain -- The range (or domain) to iterate over. The range (domain) size must be positive.
  • chunkSize : int -- The chunk size to yield to each task. Must be positive. Defaults to 1.
  • numTasks : int -- The number of tasks to use. Must be nonnegative. If this argument has value 0, the iterator will use the value indicated by dataParTasksPerLocale.
  • parDim : int -- If c is a domain, then this specifies the dimension index to parallelize across. Must be positive, and must be at most the rank of the domain c. Defaults to 1.
  • localeChunkSize : int -- Chunk size to yield to each locale. Must be nonnegative. If this argument has value 0, the iterator will use an undefined heuristic in an attempt to choose a value that will perform well.
  • coordinated : bool -- If true (and multi-locale), then have the locale invoking the iterator coordinate task distribution only; that is, disallow it from receiving work.
  • workerLocales : [] locale -- An array of locales over which to distribute the work. Defaults to Locales (all available locales).
Yields:

Indices in the range c.

This iterator is equivalent to a distributed version of the dynamic policy of OpenMP.

Given an input range (or domain) c, each locale (except the calling locale, if coordinated is true) receives chunks of size localeChunkSize from c (or the remaining iterations if there are fewer than localeChunkSize). Each locale then distributes sub-chunks of size chunkSize as tasks, using the dynamic iterator from the DynamicIters module.

Available for serial and zippered contexts.

iter distributedGuided(c, numTasks: int = 0, parDim: int = 1, minChunkSize: int = 1, coordinated: bool = false, workerLocales = Locales)
Arguments:
  • c : range(?) or domain -- The range (or domain) to iterate over. The range (domain) size must be positive.
  • numTasks : int -- The number of tasks to use. Must be nonnegative. If this argument has value 0, the iterator will use the value indicated by dataParTasksPerLocale.
  • parDim : int -- If c is a domain, then this specifies the dimension index to parallelize across. Must be positive, and must be at most the rank of the domain c. Defaults to 1.
  • minChunkSize : int -- The smallest allowable chunk size. Must be positive. Defaults to 1.
  • coordinated : bool -- If true (and multi-locale), then have the locale invoking the iterator coordinate task distribution only; that is, disallow it from receiving work.
  • workerLocales : [] locale -- An array of locales over which to distribute the work. Defaults to Locales (all available locales).
Yields:

Indices in the range c.

This iterator is equivalent to a distributed version of the guided policy of OpenMP.

Given an input range (or domain) c, each locale (except the calling locale, if coordinated is true) receives chunks of approximately exponentially decreasing size, until the remaining iterations reaches a minimum value, minChunkSize, or there are no remaining iterations in c. The chunk size is the number of unassigned iterations divided by the number of locales. Each locale then distributes sub-chunks as tasks, where each sub-chunk size is the number of unassigned local iterations divided by the number of tasks, numTasks, and decreases approximately exponentially to 1. The splitting strategy is therefore adaptive.

Available for serial and zippered contexts.