Skip to content

Latest commit

 

History

History
138 lines (99 loc) · 4.9 KB

intro-to-mpi.md

File metadata and controls

138 lines (99 loc) · 4.9 KB
Raw

Message passing interface

Message Passing Interface (MPI) is an application programming interface (API) for communication between separate processes.

MPI programs are extremely portable and can have good performance even on the largest of supercomputers. MPI is the most widely used approach for distributed parallel computing with compilers and libraries available on all typical computer architectures. Due to the widely available support, the same MPI program can run on a wide array of computers, from laptops to large supercomputers. The MPI standard defines C and Fortran interfaces, but an unofficial Python interface is provided by MPI for Python (mpi4py).

From a programmer point of view, MPI is both flexible and comprehensive. It contains over 300 procedures, but often only a handful of procedures are actually needed in a single program.

MPI contains routines for:

  • Communication between processes
    • sending and receiving messages between two processes
    • sending and receiving messages between several processes
  • Synchronization between processes
  • Communicator creation and manipulation
  • Advanced features (e.g. user defined datatypes, one-sided communication and parallel I/O)

Execution model

A MPI program is launched as a set of independent, identical processes. Each process executes exactly the same program code and instructions. The processes can reside in different CPU cores, compute nodes or even in different computers. In principle, all that is needed is a network connection between the processes, but in practice, a high-bandwidth, low-latency interconnect is needed for good performance.

A special MPI launcher is needed to start a MPI program. Different systems use different MPI launchers (depending on the MPI implementation and/or queueing system in place), but common ones include: mpirun, mpiexec, and srun. In the virtual machine used in the course we use mpirun

MPI rank

When a MPI program is launched, each process is assigned an unique ID number called rank. It is possible to query the rank within a program and to perform different tasks based on it. In fact this is the basis on which all logic for parallelism is built on in a MPI program.

if (rank == 0):
    # do something
elif (rank == 1):
    # do something else
else:
    # all other processes do something different

MPI communicator

In MPI context, a communicator is a special object representing a group of processes that participate in communication. When a MPI routine is called, the communication will involve some or all of the processes in a communicator.

All processes start in a global communicator called MPI_COMM_WORLD (or MPI.COMM_WORLD in mpi4py), but the user can also define their own custom communicators as needed.

In C and Fortran, all MPI routines expect a communicator as one of the arguments. In Python, most MPI routines are implemented as methods of a communicator object.

Data model

Since all MPI processes are completely independent, this means also a complete separation of data. Each process has its own separate memory space, i.e. all variables and data structures are local to the process. To exchange information, processes need to explicitly send and receive messages.

Getting started

To get started with MPI, it is a good idea to look into two basic methods of a communicator object: Get_size() and Get_rank().

These two routines can be used to find out two pieces of information that virtually all MPI programs need, i.e. the total number of MPI processes executing the program and the unique ID ([rank](#MPI rank)) assigned to this process.

To be more precise, they will give you the number of processes in the communicator and the rank within this communicator. A single process can belong to multiple communicators and have a different rank in each one.

Simple example: Hello world

from mpi4py import MPI

comm = MPI.COMM_WORLD # communicator object containing all processes

size = comm.Get_size()
rank = comm.Get_rank()

print("I am rank %d in group of %d processes" % (rank, size))

Running the example program

As discussed earlier, MPI program is started with a special launcher. When working with Python, the executable that is started is the Python interpreter. In the example below, we start four Python interpreters which all run then the same hello.py program. The communication between the programs is faciliated with the MPI function calls.

$ mpirun -np 4 python3 hello.py

I am rank 2 in group of 4 processes
I am rank 0 in group of 4 processes
I am rank 3 in group of 4 processes
I am rank 1 in group of 4 processes