When you distribute an array to a number of workers, MATLAB^{®} software
partitions the array into segments and assigns one segment of the
array to each worker. You can partition a two-dimensional array horizontally,
assigning columns of the original array to the different workers,
or vertically, by assigning rows. An array with N dimensions can be
partitioned along any of its N dimensions. You choose which dimension
of the array is to be partitioned by specifying it in the array constructor
command.

For example, to distribute an 80-by-1000 array to four workers, you can partition it either by columns, giving each worker an 80-by-250 segment, or by rows, with each worker getting a 20-by-1000 segment. If the array dimension does not divide evenly over the number of workers, MATLAB partitions it as evenly as possible.

The following example creates an 80-by-1000 replicated array
and assigns it to variable `A`

. In doing so, each
worker creates an identical array in its own workspace and assigns
it to variable `A`

, where `A`

is
local to that worker. The second command distributes `A`

,
creating a single 80-by-1000 array `D`

that spans
all four workers. Worker 1 stores columns 1 through 250, worker 2
stores columns 251 through 500, and so on. The default distribution
is by the last nonsingleton dimension, thus, columns in this case
of a 2-dimensional array.

spmd A = zeros(80, 1000); D = codistributed(A) end Lab 1: This lab stores D(:,1:250). Lab 2: This lab stores D(:,251:500). Lab 3: This lab stores D(:,501:750). Lab 4: This lab stores D(:,751:1000).

Each worker has access to all segments of the array. Access to the local segment is faster than to a remote segment, because the latter requires sending and receiving data between workers and thus takes more time.

For each worker, the MATLAB Parallel Command Window displays information about the codistributed array, the local portion, and the codistributor. For example, an 8-by-8 identity matrix codistributed among four workers, with two columns on each worker, displays like this:

>> spmd II = eye(8,'codistributed') end Lab 1: This lab stores II(:,1:2). LocalPart: [8x2 double] Codistributor: [1x1 codistributor1d] Lab 2: This lab stores II(:,3:4). LocalPart: [8x2 double] Codistributor: [1x1 codistributor1d] Lab 3: This lab stores II(:,5:6). LocalPart: [8x2 double] Codistributor: [1x1 codistributor1d] Lab 4: This lab stores II(:,7:8). LocalPart: [8x2 double] Codistributor: [1x1 codistributor1d]

To see the actual data in the local segment of the array, use
the `getLocalPart`

function.

In distributing an array of `N`

rows, if `N`

is
evenly divisible by the number of workers, MATLAB stores the
same number of rows (`N/numlabs`

) on each worker.
When this number is not evenly divisible by the number of workers, MATLAB partitions
the array as evenly as possible.

MATLAB provides codistributor object properties called `Dimension`

and `Partition`

that
you can use to determine the exact distribution of an array. See Indexing into a Codistributed Array for
more information on indexing with codistributed arrays.

You can distribute arrays of any MATLAB built-in data type, and also numeric arrays that are complex or sparse, but not arrays of function handles or object types.

You can create a codistributed array in any of the following ways:

Partitioning a Larger Array — Start with a large array that is replicated on all workers, and partition it so that the pieces are distributed across the workers. This is most useful when you have sufficient memory to store the initial replicated array.

Building from Smaller Arrays — Start with smaller variant or replicated arrays stored on each worker, and combine them so that each array becomes a segment of a larger codistributed array. This method reduces memory requirements as it lets you build a codistributed array from smaller pieces.

Using MATLAB Constructor Functions — Use any of the MATLAB constructor functions like

`rand`

or`zeros`

with a codistributor object argument. These functions offer a quick means of constructing a codistributed array of any size in just one step.

If you have a large array already in memory that you want MATLAB to
process more quickly, you can partition it into smaller segments and
distribute these segments to all of the workers using the `codistributed`

function.
Each worker then has an array that is a fraction the size of the original,
thus reducing the time required to access the data that is local to
each worker.

As a simple example, the following line of code creates a 4-by-8
replicated matrix on each worker assigned to the variable `A`

:

spmd, A = [11:18; 21:28; 31:38; 41:48], end A = 11 12 13 14 15 16 17 18 21 22 23 24 25 26 27 28 31 32 33 34 35 36 37 38 41 42 43 44 45 46 47 48

The next line uses the `codistributed`

function
to construct a single 4-by-8 matrix `D`

that is distributed
along the second dimension of the array:

spmd D = codistributed(A); getLocalPart(D) end 1: Local Part | 2: Local Part | 3: Local Part | 4: Local Part 11 12 | 13 14 | 15 16 | 17 18 21 22 | 23 24 | 25 26 | 27 28 31 32 | 33 34 | 35 36 | 37 38 41 42 | 43 44 | 45 46 | 47 48

Arrays `A`

and `D`

are the
same size (4-by-8). Array `A`

exists in its full
size on each worker, while only a segment of array `D`

exists
on each worker.

spmd, size(A), size(D), end

Examining the variables in the client workspace, an array that
is codistributed among the workers inside an `spmd`

statement,
is a distributed array from the perspective of the client outside
the `spmd`

statement. Variables that are not codistributed
inside the spmd, are Composites in the client outside the spmd.

whos Name Size Bytes Class A 1x4 613 Composite D 4x8 649 distributed

See the `codistributed`

function reference page
for syntax and usage information.

The `codistributed`

function is less useful
for reducing the amount of memory required to store data when you
first construct the full array in one workspace and then partition
it into distributed segments. To save on memory, you can construct
the smaller pieces (local part) on each worker first, and then use `codistributed.build`

to combine them
into a single array that is distributed across the workers.

This example creates a 4-by-250 variant array A on each of four
workers and then uses `codistributor`

to distribute
these segments across four workers, creating a 4-by-1000 codistributed
array. Here is the variant array, `A`

:

spmd A = [1:250; 251:500; 501:750; 751:1000] + 250 * (labindex - 1); end WORKER 1 WORKER 2 WORKER 3 1 2 ... 250 | 251 252 ... 500 | 501 502 ... 750 | etc. 251 252 ... 500 | 501 502 ... 750 | 751 752 ...1000 | etc. 501 502 ... 750 | 751 752 ...1000 | 1001 1002 ...1250 | etc. 751 752 ...1000 | 1001 1002 ...1250 | 1251 1252 ...1500 | etc. | | |

Now combine these segments into an array that is distributed by the first dimension (rows). The array is now 16-by-250, with a 4-by-250 segment residing on each worker:

spmd D = codistributed.build(A, codistributor1d(1,[4 4 4 4],[16 250])) end Lab 1: This lab stores D(1:4,:). LocalPart: [4x250 double] Codistributor: [1x1 codistributor1d] whos Name Size Bytes Class A 1x4 613 Composite D 16x250 649 distributed

You could also use replicated arrays in the same fashion, if
you wanted to create a codistributed array whose segments were all
identical to start with. See the `codistributed`

function
reference page for syntax and usage information.

MATLAB provides several array constructor functions that
you can use to build codistributed arrays of specific values, sizes,
and classes. These functions operate in the same way as their nondistributed
counterparts in the MATLAB language, except that they distribute
the resultant array across the workers using the specified codistributor
object, `codist`

.

**Constructor Functions. **The codistributed constructor functions are listed here. Use
the `codist`

argument (created by the `codistributor`

function: `codist=codistributor()`

)
to specify over which dimension to distribute the array. See the individual
reference pages for these functions for further syntax and usage information.

`eye`

(___,codist)`false`

(___,codist)`Inf`

(___,codist)`NaN`

(___,codist)`ones`

(___,codist)`rand`

(___,codist)`randi`

(___,codist)`randn`

(___,codist)`true`

(___,codist)`zeros`

(___,codist)`codistributed.cell`

(m,n,...,codist)`codistributed.colon`

(a,d,b) codistributed.`linspace`

(m,n,...,codist) codistributed.`logspace`

(m,n,...,codist)`sparse`

(m,n,codist)`codistributed.speye`

(m,...,codist)`codistributed.sprand`

(m,n,density,codist)`codistributed.sprandn`

(m,n,density,codist)

That part of a codistributed array that resides on each worker
is a piece of a larger array. Each worker can work on its own segment
of the common array, or it can make a copy of that segment in a variant
or private array of its own. This local copy of a codistributed array
segment is called a *local array*.

The `getLocalPart`

function copies the segments
of a codistributed array to a separate variant array. This example
makes a local copy `L`

of each segment of codistributed
array `D`

. The size of `L`

shows
that it contains only the local part of `D`

for each
worker. Suppose you distribute an array across four workers:

spmd(4) A = [1:80; 81:160; 161:240]; D = codistributed(A); size(D) L = getLocalPart(D); size(L) end

returns on each worker:

3 80 3 20

Each worker recognizes that the codistributed array `D`

is
3-by-80. However, notice that the size of the local part, `L`

,
is 3-by-20 on each worker, because the 80 columns of `D`

are
distributed over four workers.

Use the `codistributed.build`

function
to perform the reverse operation. This function, described in Building from Smaller Arrays, combines
the local variant arrays into a single array distributed along the
specified dimension.

Continuing the previous example, take the local variant arrays `L`

and
put them together as segments to build a new codistributed array `X`

.

spmd codist = codistributor1d(2,[20 20 20 20],[3 80]); X = codistributed.build(L,codist); size(X) end

returns on each worker:

3 80

MATLAB offers several functions that provide information on any particular array. In addition to these standard functions, there are also two functions that are useful solely with codistributed arrays.

The `iscodistributed`

function
returns a logical `1`

(`true`

) if
the input array is codistributed, and logical `0`

(`false`

)
otherwise. The syntax is

spmd, TF = iscodistributed(D), end

where `D`

is any MATLAB array.

The codistributor object determines how an array is partitioned
and its dimension of distribution. To access the codistributor of
an array, use the `getCodistributor`

function.
This returns two properties, `Dimension`

and `Partition`

:

spmd, getCodistributor(X), end Dimension: 2 Partition: [20 20 20 20]

The `Dimension`

value of `2`

means
the array `X`

is distributed by columns (dimension
2); and the `Partition`

value of ```
[20 20
20 20]
```

means that twenty columns reside on each of the four
workers.

To get these properties programmatically, return the output
of `getCodistributor`

to a variable, then use dot
notation to access each property:

spmd C = getCodistributor(X); part = C.Partition dim = C.Dimension end

Other functions that provide information about standard arrays also work on codistributed arrays and use the same syntax.

When constructing an array, you distribute the parts of the
array along one of the array's dimensions. You can change the direction
of this distribution on an existing array using the `redistribute`

function
with a different codistributor object.

Construct an 8-by-16 codistributed array `D`

of
random values distributed by columns on four workers:

spmd D = rand(8,16,codistributor()); size(getLocalPart(D)) end

returns on each worker:

8 4

Create a new codistributed array distributed by rows from an existing one already distributed by columns:

spmd X = redistribute(D, codistributor1d(1)); size(getLocalPart(X)) end

returns on each worker:

2 16

You can restore a codistributed array to its undistributed form
using the `gather`

function. `gather`

takes
the segments of an array that reside on different workers and combines
them into a replicated array on all workers, or into a single array
on one worker.

Distribute a 4-by-10 array to four workers along the second dimension:

spmd, A = [11:20; 21:30; 31:40; 41:50], end A = 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 spmd, D = codistributed(A), end WORKER 1 WORKER 2 WORKER 3 WORKER 4 11 12 13 | 14 15 16 | 17 18 | 19 20 21 22 23 | 24 25 26 | 27 28 | 29 30 31 32 33 | 34 35 36 | 37 38 | 39 40 41 42 43 | 44 45 46 | 47 48 | 49 50 | | | spmd, size(getLocalPart(D)), end Lab 1: 4 3 Lab 2: 4 3 Lab 3: 4 2 Lab 4: 4 2

Restore the undistributed segments to the full array form by gathering the segments:

spmd, X = gather(D), end X = 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 spmd, size(X), end 4 10

While indexing into a nondistributed array is fairly straightforward,
codistributed arrays require additional considerations. Each dimension
of a nondistributed array is indexed within a range of 1 to the final
subscript, which is represented in MATLAB by the `end`

keyword.
The length of any dimension can be easily determined using either
the `size`

or `length`

function.

With codistributed arrays, these values are not so easily obtained.
For example, the second segment of an array (that which resides in
the workspace of worker 2) has a starting index that depends on the
array distribution. For a 200-by-1000 array with a default distribution
by columns over four workers, the starting index on worker 2 is 251.
For a 1000-by-200 array also distributed by columns, that same index
would be 51. As for the ending index, this is not given by using the `end`

keyword,
as `end`

in this case refers to the end of the entire
array; that is, the last subscript of the final segment. The length
of each segment is also not given by using the `length`

or `size`

functions,
as they only return the length of the entire array.

The MATLAB `colon`

operator and `end`

keyword
are two of the basic tools for indexing into nondistributed arrays.
For codistributed arrays, MATLAB provides a version of the `colon`

operator,
called `codistributed.colon`

. This actually is a
function, not a symbolic operator like `colon`

.

When using arrays to index into codistributed arrays, you can
use only replicated or codistributed arrays for indexing. The toolbox
does not check to ensure that the index is replicated, as that would
require global communications. Therefore, the use of unsupported variants
(such as `labindex`

) to index
into codistributed arrays might create unexpected results.

Suppose you have a row vector of 1 million elements, distributed
among several workers, and you want to locate its element number 225,000.
That is, you want to know what worker contains this element, and in
what position in the local part of the vector on that worker. The `globalIndices`

function provides a correlation
between the local and global indexing of the codistributed array.

D = rand(1,1e6,'distributed'); %Distributed by columns spmd globalInd = globalIndices(D,2); pos = find(globalInd == 225e3); if ~isempty(pos) fprintf(... 'Element is in position %d on worker %d.\n', pos, labindex); end end

If you run this code on a pool of four workers you get this result:

Lab 1: Element is in position 225000 on worker 1.

If you run this code on a pool of five workers you get this result:

Lab 2: Element is in position 25000 on worker 2.

Notice if you use a pool of a different size, the element ends up in a different location on a different worker, but the same code can be used to locate the element.

As
an alternative to distributing by a single dimension of rows or columns,
you can distribute a matrix by blocks using `'2dbc'`

or
two-dimensional block-cyclic distribution. Instead of segments that
comprise a number of complete rows or columns of the matrix, the segments
of the codistributed array are 2-dimensional square blocks.

For
example, consider a simple 8-by-8 matrix with ascending element values.
You can create this array in an `spmd`

statement,
communicating job, or pmode. This example uses pmode for a visual
display.

P>> A = reshape(1:64, 8, 8)

The result is the replicated array:

1 9 17 25 33 41 49 57 2 10 18 26 34 42 50 58 3 11 19 27 35 43 51 59 4 12 20 28 36 44 52 60 5 13 21 29 37 45 53 61 6 14 22 30 38 46 54 62 7 15 23 31 39 47 55 63 8 16 24 32 40 48 56 64

Suppose you want to distribute this array among four workers, with a 4-by-4 block as the local part on each worker. In this case, the lab grid is a 2-by-2 arrangement of the workers, and the block size is a square of four elements on a side (i.e., each block is a 4-by-4 square). With this information, you can define the codistributor object:

P>> DIST = codistributor2dbc([2 2], 4)

Now you can use this codistributor object to distribute the original matrix:

P>> AA = codistributed(A, DIST)

This distributes the array among the workers according to this scheme:

If the lab grid does not perfectly overlay the dimensions of
the codistributed array, you can still use `'2dbc'`

distribution,
which is block cyclic. In this case, you can imagine the lab grid
being repeatedly overlaid in both dimensions until all the original
matrix elements are included.

Using the same original 8-by-8 matrix and 2-by-2 lab grid, consider a block size of 3 instead of 4, so that 3-by-3 square blocks are distributed among the workers. The code looks like this:

P>> DIST = codistributor2dbc([2 2], 3) P>> AA = codistributed(A, DIST)

The first “row” of the lab grid is distributed to worker 1 and worker 2, but that contains only six of the eight columns of the original matrix. Therefore, the next two columns are distributed to worker 1. This process continues until all columns in the first rows are distributed. Then a similar process applies to the rows as you proceed down the matrix, as shown in the following distribution scheme:

The diagram above shows a scheme that requires four overlays of the lab grid to accommodate the entire original matrix. The following pmode session shows the code and resulting distribution of data to each of the workers:

The following points are worth noting:

`'2dbc'`

distribution might not offer any performance enhancement unless the block size is at least a few dozen. The default block size is 64.The lab grid should be as close to a square as possible.

Not all functions that are enhanced to work on

`'1d'`

codistributed arrays work on`'2dbc'`

codistributed arrays.