CyclicDist¶
Usage
use CyclicDist;
-
class
Cyclic
¶ This Cyclic distribution maps indices to locales in a round-robin pattern starting at a given index.
Formally, consider a Cyclic distribution with:
rank d
start index (s_1, ...., s_d)
over locales targetLocales: [0..N_1-1, ...., 0..N_d-1] locale
It maps an index
(i_1, ...., i_d)
to the localetargetLocales[j_1, ...., j_d]
where, for eachk
in1..d
, we have:j_k = (i_k - s_k) (mod N_k)
Example
The following code declares a domain
D
distributed using a Cyclic distribution with a start index of(1,1)
, and declares an arrayA
over that domain. The forall loop sets each array element to the ID of the locale to which it is mapped.use CyclicDist; const Space = {1..8, 1..8}; const D: domain(2) dmapped Cyclic(startIdx=Space.low) = Space; var A: [D] int; forall a in A do a = a.locale.id; writeln(A);
When run on 6 locales, the output is:
0 1 0 1 0 1 0 1 2 3 2 3 2 3 2 3 4 5 4 5 4 5 4 5 0 1 0 1 0 1 0 1 2 3 2 3 2 3 2 3 4 5 4 5 4 5 4 5 0 1 0 1 0 1 0 1 2 3 2 3 2 3 2 3
Initializer Arguments
The
Cyclic
class initializer is defined as follows:proc Cyclic.init( startIdx, targetLocales: [] locale = Locales, dataParTasksPerLocale = // value of dataParTasksPerLocale config const, dataParIgnoreRunningTasks = // value of dataParIgnoreRunningTasks config const, dataParMinGranularity = // value of dataParMinGranularity config const, param rank: int = // inferred from startIdx argument, type idxType = // inferred from startIdx argument )
The argument
startIdx
is a tuple of integers defining an index that will be distributed to the first locale intargetLocales
. In the 1-dimensional case,startIdx
can be an integer or a tuple with a single element.The argument
targetLocales
is an array containing the target locales to which this distribution maps indices and data. The rank oftargetLocales
must match the rank of the distribution, or be1
. If the rank oftargetLocales
is1
, a greedy heuristic is used to reshape the array of target locales so that it matches the rank of the distribution and each dimension contains an approximately equal number of indices.The arguments
dataParTasksPerLocale
,dataParIgnoreRunningTasks
, anddataParMinGranularity
set the knobs that are used to control intra-locale data parallelism for Cyclic-distributed domains and arrays in the same way that the like-named config constants control data parallelism for ranges and default-distributed domains and arrays.The
rank
andidxType
arguments are inferred from thestartIdx
argument unless explicitly set. They must match the rank and index type of the domains “dmapped” using that Cyclic instance.Convenience Initializer Functions
It is common for a
Cyclic
distribution to distribute its indices across all locales. In this case, a convenience function can be used to declare variables of cyclic-distributed domain or array type. These functions take a domain or list of ranges as arguments and return a cyclic-distributed domain or array.use CyclicDist; var CyclicDom1 = newCyclicDom({1..5, 1..5}); var CyclicArr1 = newCyclicArr({1..5, 1..5}, real); var CyclicDom2 = newCyclicDom(1..5, 1..5); var CyclicArr2 = newCyclicArr(1..5, 1..5, real);
Data-Parallel Iteration
A forall loop over a Cyclic-distributed domain or array executes each iteration on the locale where that iteration’s index is mapped to.
Parallelism within each locale is guided by the values of
dataParTasksPerLocale
,dataParIgnoreRunningTasks
, anddataParMinGranularity
of the respective Cyclic instance. Updates to these values, if any, take effect only on the locale where the updates are made.Limitations
This distribution has not been tuned for performance.