Parallel Computation - Lecture Slides | COSC 4397, Study notes of Computer Science

Download Parallel Computation - Lecture Slides | COSC 4397 and more Study notes Computer Science in PDF only on Docsity!

COSC 4397 – Parallel ComputationEdgar Gabriel

COSC 4397

Parallel ComputationGroup management

Edgar GabrielSpring 2006

2

COSC 4397 – Parallel ComputationEdgar Gabriel

Terminology (I)

•^

an

MPI_Group

is the object describing the list of processes

forming a logical entity– a group has a size

MPI_Group_size

• every process in the group has a unique rank between 0

and (size of group -1)

MPI_Group_rank

• a group is a local object, and cannot be used for any

communication

4

COSC 4397 – Parallel ComputationEdgar Gabriel

Predefined communicators

• MPI_COMM_WORLD

• contains all processes started with

mpirun/mpiexec

• exist upon exiting

MPI_Init

• can not be modified, freed etc.
  • MPI_COMM_SELF
    • contains just the local process itself, size is always 1– exist upon exiting

MPI_Init

• can not be modified, freed etc.

5

COSC 4397 – Parallel ComputationEdgar Gabriel

Creating new communicators

-^

All communicators in MPI-1 are derived from MPI_COMM_WORLD

and

MPI_COMM_SELF

-^

Creating a freeing a communicator is a

collective

operation

Ä

all processes of the participating original communicator have to call the function with the same arguments

-^

Methods to create new communicators– splitting the original communicator into n-parts– creating subgroups of the original communicator– re-ordering of processes based on topology information– spawn new processes– connect two applications and merge their

communicators

7

COSC 4397 – Parallel ComputationEdgar Gabriel

Example for MPI_Comm_split (I)

-^

odd/even splitting of processes

-^

a process– can just be part of

one

of the generated communicators

MPI_Comm newcomm;int color, rank;MPI_Comm_rank (MPI_COMM_WORLD, &rank);color = rank%2;MPI_Comm_split (MPI_COMM_WORLD,– can not “see” the other communicators– can not “see” how many communicators have been created

color, rank, &newcomm);

MPI_Comm_size (newcomm, &size);MPI_Comm_rank (newcomm, &rank);

8

COSC 4397 – Parallel ComputationEdgar Gabriel

Example for MPI_Comm_split (II)

•^

rank and size of the new communicator

MPI_COMM_WORLDnewcomm,

color=0,

size

=

4

0

1

2

3

4

5

6

0

1

2

3

0

1

2

newcomm,

color=1,

size

=

3

10

COSC 4397 – Parallel ComputationEdgar Gabriel

Modifying the group of processes

comm

MPI_Comm_group() MPI_Group_incl/excl()

MPI_Comm_create()

newcomm

original communicatorExtract the group of processesfrom the original communicatorModify the groupCreate new communicator basedon the modified groupnew communicator

11

COSC 4397 – Parallel ComputationEdgar Gabriel

Extracting the group of processes

with

• comm

: original communicator

• group

: the group object describing the list of

participating processes in

comm

MPI_Comm_group (MPI_Comm comm, MPI_Group *group);

13

COSC 4397 – Parallel ComputationEdgar Gabriel

Modifying groups (II)

•^

for more group-constructors, see also^ – MPI_Group_range_incl– MPI_Group_range_excl– MPI_Group_difference– MPI_Group_intersection– MPI_Group_union

14

COSC 4397 – Parallel ComputationEdgar Gabriel

Creating a new communicator based on a

group

with

• comm

: original communicator

• group

: the group object describing the list of processes

for the new communicator

• newcomm

: resulting communicator

Note:

• newcomm

is always a subset of

comm

• you can generate

one

communicator at a time (in contrary

to

MPI_Comm_split

)

• list of arguments has to be identical on all processes of

comm

• newcomm

will be

MPI_COMM_NULL

for processes which

have been excluded/not included in

newgroup

MPI_Comm_create ( MPI_Comm comm, MPI_Group newgroup,

MPI_Comm *newcomm);

16

COSC 4397 – Parallel ComputationEdgar Gabriel

Possibility 1: using MPI_Group_incl

MPI_Comm newcomm;MPI_Group group, newgroup;int color, size, ranks[5], cnt;MPI_Comm_size (MPI_COMM_WORLD, &size);cnt = 5;ranks[0] = 0; ranks[1] = 1; ranks[2] = 2; ranks[3] = 3;ranks[4] = size-1MPI_Comm_group (MPI_COMM_WORLD, &group);MPI_Group_incl (group,

cnt, ranks, &newgroup)

MPI_Comm_create (comm,

newgroup, &newcomm);

if ( newcomm != MPI_COMM_NULL ) {

MPI_Comm_rank (newcomm, &nrank);MPI_Comm_free (&newcomm);MPI_Group_free (&newgroup);

} MPI_Group_free (&group);

17

COSC 4397 – Parallel ComputationEdgar Gabriel

Possibility 2: using MPI_Group_excl MPI_Comm newcomm;MPI_Group group, newgroup;int color, size, ranks[…], cnt;/*

NOTE:

Assuming

that

size

5,

ranks

is

large

enough

etc.

*/

MPI_Comm_size (MPI_COMM_WORLD, &size);cnt = 0;for ( i=4; i<(size-1); i++) {

ranks[cnt++] = I; } MPI_Comm_group (MPI_COMM_WORLD, &group);MPI_Group_excl (group,

cnt-1, ranks, &newgroup)

MPI_Comm_create (comm, newgroup, &newcomm);if ( newcomm != MPI_COMM_NULL ) {

MPI_Comm_rank (newcomm, &nrank);MPI_Comm_free (&newcomm);MPI_Group_free (&newgroup);

} MPI_Group_free (&group);

19

COSC 4397 – Parallel ComputationEdgar Gabriel

What else is there?

•^

Creating a communicator using topology information– e.g. generate a communicator, where the processes

form a 2-D/3-D logical cartesian grid with MPI_Cart_create

• functions to determine the ranks of neighbor

processes in all dimensions available

• e.g. extract the processes in a certain dimension of

the cartesian grid using

MPI_Cart_sub

• e.g. generate a communicator, where the processes

are ordered logically as described by a directed graphusing

MPI_Graph_create

20

COSC 4397 – Parallel ComputationEdgar Gabriel

intra vs. inter-communicators (I)

•^

An intra-communicator contains of a single group ofprocesses– all processes have a unique rank in the group

-^

An inter-communicator contains of two groups– a local group and a remote group– a process is always part of a local group, e.g. has a

unique rank in the local group

• local and remote group have however separate

ranking scheme

Ä

you have two processes having the rank 0, one inthe local group and one in the remote group