NAME
Parallel::MPI::SimpleSYNOPSIS
mpirun -np 2 perl script.pl
#!perl
use Parallel::MPI::Simple;
MPI_Init();
my $rank = MPI_Comm_rank(MPI_COMM_WORLD);
if ($rank == 1) {
my $msg = "Hello, I'm $rank";
MPI_Send($msg, 0, 123, MPI_COMM_WORLD);
}
else {
my $msg = MPI_Recv(1, 123, MPI_COMM_WORLD);
print "$rank received: '$msg'\n";
}
MPI_Finalise();INTRODUCTION
Perl is not a strongly typed language, Perl does not enforce data structures of a fixed size or dimensionality, Perl makes things easy. Parallel processing solves problems faster and is commonly programmed using a message passing paradigm. Traditional message passing systems are designed for strongly typed languages like C or Fortran, there exist implementations of these for Perl but they concentrate on perfectly mimicing the standards forcing the poor Perl programmer to use strongly typed data despite all his best instincts.
This module provides a non-compliant wrapper around the widely implemented MPI libraries, allowing messages to consist of arbitarily nested Perl data structures whose size is limited by available memory. This hybrid approach should allow you to quickly write programs which run anywhere which supports MPI (both Beowulf and traditional MPP machines).
Message Passing and Multiprocessing
The message passing paradigm is simple and easy to use. Multiple versions of the same program are run on multiple processors (or nodes). Each running copy should call MPI_Init to announce that it is running. It can then find out who it is by calling MPI_Comm_rank and who else it can talk to by calling MPI_Comm_size. Using this information to decide what part it is to play in the ensuing computation, it the exchanges messages, or parcels of data, with other nodes allowing all to cooperate.
Once the computation is finished, the node calls MPI_Finalize and exits cleanly, ready to run another day.
These processes are all copies of the same perl script and are invoked using: mpirun -np [number of nodes] perl script.pl .
Starting and Stopping a process
A process must formally enter and leave the MPI pool by calling these functions.
MPI_Init
MPI_Init()Initialises the message passing layer. This should be the first MPI_* call made by the program and ideally one of the first things the program does. After completing this call, all processes will be synchronised and will become members of the MPI_COMM_WORLD communicator. It is an error for programs invoked with mpirun to fail to call MPI_Init (not to mention being a little silly).
MPI_Finalize
MPI_Finalize()Shuts down the message passing layer. This should be called by every participating process before exiting. No more MPI_* calls may be made after this function has been called. It is an error for a program to exit without calling this function.
Communicators
All processes are members of one or more communicators. These are like channels over which messages are broadcast. Any operation involving more than one process will take place in a communicator, operations involving one communicator will not interfere with those in another.
On calling MPI_Init all nodes automatically join the MPI_COMM_WORLD communicator. A communicator can be split into smaller subgroups using the MPI_Comm_split function.
MPI_COMM_WORLD
$global_comm = MPI_COMM_WORLD;Returns the global communicator shared by all processes launched at the same time. Can be used as a "constant" where a communicator is required. Most MPI applications can get by using only this communicator.
MPI_Comm_rank
$rank = MPI_Comm_rank($comm);Returns the rank of the process within the communicator given by $comm. Processes have ranks from 0..(size-1).
MPI_Comm_size
$size = MPI_Comm_size($comm);Returns the number of processes in communicator $comm.
MPI_Comm_compare
$result = MPI_Comm_compare($comm1, $comm2);Compares the two communicators $comm1 and $comm2. $result will be equal to:
MPI_IDENT : communicators are identical
MPI_CONGRUENT: membership is same, ranks are equal
MPI_SIMILAR : membership is same, ranks not equal
MPI_UNEQUAL : at least one member of one is not in the otherMPI_Comm_dup
$newcomm = MPI_Comm_dup($comm);Duplicates $comm but creates a new context for messages.
MPI_Comm_split
$newcomm = MPI_Comm_split($comm, $colour, $key);Every process in $comm calls MPI_Comm_split at the same time. A new set of communicators is produced, one for each distinct value of $colour. All those processes which specified the same value of $colour end up in the same comminicator and are ranked on the values of $key, with their original ranks in $comm being used to settle ties.
If $colour is negative (or MPI_UNDEFINED), the process will not be allocated to any of the new communicators and undef will be returned.
MPI_Comm_free
MPI_Comm_free($comm, [$comm2, ...] );Frees the underlying object in communicator $comm, do not attempt to do this to MPI_COMM_WORLD, wise to do this for any other comminicators that you have created. If given a list of comminicators, will free all of them, make sure there are no duplicates...
Communications operations
MPI_Barrier
MPI_Barrier($comm);Waits for every process in $comm to call MPI_Barrier, once done, all continue to execute. This causes synchronisation of processes. Be sure that every process does call this, else your computation will hang.
MPI_Send
MPI_Send($scalar, $dest, $msg_tag, $comm);This takes a scalar (which can be an anonymous reference to a more complicated data structure) and sends it to process with rank $dest in communicator $comm. The message also carries $msg_tag as an identfier, allowing nodes to receive and send out of order. Completion of this call does not imply anything about the progress of the receiving node.
MPI_Recv
$scalar = MPI_Recv($source, $msg_tag, $comm);Receives a scalar from another process. $source and $msg_tag must both match a message sent via MPI_Send (or one which will be sent in future) to the same communicator given by $comm.
if ($rank == 0) {
MPI_Send([qw(This is a message)], 1, 0, MPI_COMM_WORLD);
}
elsif ($rank == 1) {
my $msg = MPI_Recv(1,0,MPI_COMM_WORLD);
print join(' ', @{ $msg } );
}Will output "This is a message". Messages with the same source, destination, tag and comminicator will be delivered in the order in which they were sent. No other guarantees of timeliness or ordering can be given. If needed, use MPI_Barrier.
MPI_Bcast
$data = MPI_Bcast($scalar, $root, $comm);This sends $scalar in process $root from the root process to every other process in $comm, returning this scalar in every process. All non-root processes should provide a dummy message (such as undef), this is a bit ugly, but maintains a consistant interface between the other communication operations. The scalar can be a complicated data structure.
if ($rank == 0) { # send from 0
my $msg = [1,2,3, {3=>1, 5=>6} ];
MPI_Bcast( $msg, 0, MPI_COMM_WORLD);
}
else { # everything else receives, note dummy message
my $msg = MPI_Bcast(undef, 0, MPI_COMM_WORLD);
}MPI_Gather
# if root:
@list = MPI_Gather($scalar, $root, $comm);
#otherwise
(nothing) = MPI_Gather($scalar, $root, $comm);Sends $scalar from every process in $comm (each $scalar can be different, root's data is also sent) to the root process which collects them as a list of scalars, sorted by process rank order in $comm.
MPI_Scatter
$data = MPI_Scatter([N items of data], $root, $comm);Sends list of scalars (anon array as 1st arg) from $root to all processes in $comm, with process of rank N-1 receiving the Nth item in the array. Very bad things might happen if number of elements in array != N. This does not call the C function at any time, so do not expect any implicit synchronisation.
MPI_Allgather
@list = MPI_Allgather($scalar, $comm);Every process receives an ordered list containing an element from every other process. Again, this is implemented without a call to the C function.
MPI_Alltoall
@list = MPI_Alltoall([ list of scalars ], $comm);Simillar to Allgather, each process (with rank rank) ends up with a list such that element i contains the data which started in element rank of process is data.
MPI_Reduce
$value = MPI_Reduce($input, \&operation, $comm);Every process receives in $value the result of performing &operation between every processes $input. If there are three processes in $comm, then $value = $input_0 op $input_1 op $input_2.
Operation should be a sub which takes two scalar values (the $input above) and returns a single value. The operation it performs should be commutative and associative, otherwise the result will be undefined.
For instance, to return the sum of some number held by each process, perform:
$sum = MPI_Reduce($number, sub {$_[0] + $_[1]}, $comm);To find which process holds the greatest value of some number:
($max, $mrank) = @{ MPI_Reduce([$number, $rank],
sub { $_[0]->[0] > $_[1]->[0] ? $_[0] : $_[1]}
, $comm) };PHILOSOPHY
I have decided to loosely follow the MPI calling and naming conventions but do not want to stick strictly to them in all cases. In the interests of being simple, I have decided that all errors should result in the death of the MPI process rather than huge amounts of error checking being foisted onto the module's user.
Many of the MPI functions have not been implemented, some of this is because I feel they would complicate the module (I chose to have a single version of the Send command, for instance) but some of this is due to my not having finished yet. I certainly do not expect to provide process topologies or inter-communicators, I also do not expect to provide anything in MPI-2 for some time.
ISSUES
This module has been tested on a variety of platforms. I have not been able to get it running with the mpich MPI implementation in a clustered environment.
In general, I expect that most programs using this module will make use of little other than MPI_Init, MPI_Send, MPI_Recv, MPI_COMM_WORLD, MPI_Barrier, MPI_Comm_size, MPI_Comm_rank and MPI_Finalize.
AUTHOR
Alex Gough (alex@rcon.org)COPYRIGHT
This module is copyright (c) Alex Gough, 2001.
You may use and redistribute this software under the Artistic License as
supplied with Perl.