Message Passing Interface


Message Passing Interface[] is a standardized and portable message-passing standard designed by a group of researchers from academia and industry to function on a wide variety of parallel computing architectures. The standard defines the syntax and semantics of a core of library routines useful to a wide range of users writing portable message-passing programs in C, C++, and Fortran. There are several well-tested and efficient implementations of MPI, many of which are open-source or in the public domain. These fostered the development of a parallel software industry, and encouraged development of portable and scalable large-scale parallel applications.

History

The message passing interface effort began in the summer of 1991 when a small group of researchers started discussions at a mountain retreat in Austria. Out of that discussion came a Workshop on Standards for Message Passing in a Distributed Memory Environment held on April 29–30, 1992 in Williamsburg, Virginia. Attendees at Williamsburg discussed the basic features essential to a standard message-passing interface and established a working group to continue the standardization process. Jack Dongarra, Tony Hey, and David W. Walker put forward a preliminary draft proposal, "MPI1", in November 1992. In November 1992 a meeting of the MPI working group took place in Minneapolis and decided to place the standardization process on a more formal footing. The MPI working group met every 6 weeks throughout the first 9 months of 1993. The draft MPI standard was presented at the Supercomputing '93 conference in November 1993. After a period of public comments, which resulted in some changes in MPI, version 1.0 of MPI was released in June 1994. These meetings and the email discussion together constituted the MPI Forum, membership of which has been open to all members of the high-performance-computing community.
The MPI effort involved about 80 people from 40 organizations, mainly in the United States and Europe. Most of the major vendors of concurrent computers were involved in MPI - along with researchers from universities, government laboratories, and industry.
MPI provides parallel hardware vendors with a clearly defined base set of routines that can be efficiently implemented. As a result, hardware vendors can build upon this collection of standard low-level routines to create higher-level routines for the distributed-memory communication environment supplied with their parallel machines. MPI provides a simple-to-use portable interface for the basic user, yet one powerful enough to allow programmers to use the high-performance message passing operations available on advanced machines.
In an effort to create a universal standard for message passing, researchers incorporated the most useful features of several systems into MPI, rather than choosing one system to adopt as a standard. They adopted features from systems by IBM, Intel, nCUBE, PVM, Express, P4 and PARMACS. The message-passing paradigm is attractive because of wide portability and can be used in communication for distributed-memory and shared-memory multiprocessors, networks of workstations, and a combination of these elements. The paradigm can apply in multiple settings, independent of network speed or of memory architecture.
Support for MPI meetings came in part from DARPA and from the US National Science Foundation under grant ASC-9310330, NSF Science and Technology Center Cooperative agreement number CCR-8809615, and from the Commission of the European Community through Esprit Project P6643. The University of Tennessee also made financial contributions to the MPI Forum.

Overview

MPI is a communication protocol for programming parallel computers. Both point-to-point and collective communication are supported. MPI "is a message-passing application programmer interface, together with protocol and semantic specifications for how its features must behave in any implementation." MPI's goals are high performance, scalability, and portability. MPI remains the dominant model used in high-performance computing today.
MPI is not sanctioned by any major standards body; nevertheless, it has become a de facto standard for communication among processes that model a parallel program running on a distributed memory system. Actual distributed memory supercomputers such as computer clusters often run such programs.
The principal MPI-1 model has no shared memory concept, and MPI-2 has only a limited distributed shared memory concept. Nonetheless, MPI programs are regularly run on shared memory computers, and both MPICH and Open MPI can use shared memory for message transfer if it is available. Designing programs around the MPI model has advantages over NUMA architectures since MPI encourages memory locality. Explicit shared memory programming was introduced in MPI-3.
Although MPI belongs in layers 5 and higher of the OSI Reference Model, implementations may cover most layers, with sockets and Transmission Control Protocol used in the transport layer.
Most MPI implementations consist of a specific set of routines directly callable from C, C++, Fortran and any language able to interface with such libraries, including C#, Java or Python. The advantages of MPI over older message passing libraries are portability and speed.
MPI uses Language Independent Specifications for calls and language bindings. The first MPI standard specified ANSI C and Fortran-77 bindings together with the LIS. The draft was presented at Supercomputing 1994 and finalized soon thereafter. About 128 functions constitute the MPI-1.3 standard which was released as the final end of the MPI-1 series in 2008.
At present, the standard has several versions: version 1.3, which emphasizes message passing and has a static runtime environment, MPI-2.2, which includes new features such as parallel I/O, dynamic process management and remote memory operations, and MPI-3.1, which includes extensions to the collective operations with non-blocking versions and extensions to the one-sided operations.
MPI-2's LIS specifies over 500 functions and provides language bindings for ISO C, ISO C++, and Fortran 90. Object interoperability was also added to allow easier mixed-language message passing programming. A side-effect of standardizing MPI-2, completed in 1996, was clarifying the MPI-1 standard, creating the MPI-1.2.
MPI-2 is mostly a superset of MPI-1, although some functions have been deprecated. MPI-1.3 programs still work under MPI implementations compliant with the MPI-2 standard.
MPI-3 includes new Fortran 2008 bindings, while it removes deprecated C++ bindings as well as many deprecated routines and MPI objects.
MPI is often compared with Parallel Virtual Machine, which is a popular distributed environment and message passing system developed in 1989, and which was one of the systems that motivated the need for standard parallel message passing. Threaded shared memory programming models and message passing programming can be considered as complementary programming approaches, and can occasionally be seen together in applications, e.g. in servers with multiple large shared-memory nodes.

Functionality

The MPI interface is meant to provide essential virtual topology, synchronization, and communication functionality between a set of processes in a language-independent way, with language-specific syntax, plus a few language-specific features. MPI programs always work with processes, but programmers commonly refer to the processes as processors. Typically, for maximum performance, each CPU will be assigned just a single process. This assignment happens at runtime through the agent that starts the MPI program, normally called mpirun or mpiexec.
MPI library functions include, but are not limited to, point-to-point rendezvous-type send/receive operations, choosing between a Cartesian or graph-like logical process topology, exchanging data between process pairs, combining partial results of computations, synchronizing nodes as well as obtaining network-related information such as the number of processes in the computing session, current processor identity that a process is mapped to, neighboring processes accessible in a logical topology, and so on. Point-to-point operations come in synchronous, asynchronous, buffered, and ready forms, to allow both relatively stronger and weaker semantics for the synchronization aspects of a rendezvous-send. Many outstanding operations are possible in asynchronous mode, in most implementations.
MPI-1 and MPI-2 both enable implementations that overlap communication and computation, but practice and theory differ. MPI also specifies thread safe interfaces, which have cohesion and coupling strategies that help avoid hidden state within the interface. It is relatively easy to write multithreaded point-to-point MPI code, and some implementations support such code. Multithreaded collective communication is best accomplished with multiple copies of Communicators, as described below.

Concepts

MPI provides a rich range of abilities. The following concepts help in understanding and providing context for all of those abilities and help the programmer to decide what functionality to use in their application programs. Four of MPI's eight basic concepts are unique to MPI-2.

Communicator

Communicator objects connect groups of processes in the MPI session. Each communicator gives each contained process an independent identifier and arranges its contained processes in an ordered topology. MPI also has explicit groups, but these are mainly good for organizing and reorganizing groups of processes before another communicator is made. MPI understands single group intracommunicator operations, and bilateral intercommunicator communication. In MPI-1, single group operations are most prevalent. Bilateral operations mostly appear in MPI-2 where they include collective communication and dynamic in-process management.
Communicators can be partitioned using several MPI commands. These commands include MPI_COMM_SPLIT, where each process joins one of several colored sub-communicators by declaring itself to have that color.

Point-to-point basics

A number of important MPI functions involve communication between two specific processes. A popular example is MPI_Send, which allows one specified process to send a message to a second specified process. Point-to-point operations, as these are called, are particularly useful in patterned or irregular communication, for example, a data-parallel architecture in which each processor routinely swaps regions of data with specific other processors between calculation steps, or a master-slave architecture in which the master sends new task data to a slave whenever the prior task is completed.
MPI-1 specifies mechanisms for both blocking and non-blocking point-to-point communication mechanisms, as well as the so-called 'ready-send' mechanism whereby a send request can be made only when the matching receive request has already been made.

Collective basics

involve communication among all processes in a process group. A typical function is the MPI_Bcast call. This function takes data from one node and sends it to all processes in the process group. A reverse operation is the MPI_Reduce call, which takes data from all processes in a group, performs an operation, and stores the results on one node. MPI_Reduce is often useful at the start or end of a large distributed calculation, where each processor operates on a part of the data and then combines it into a result.
Other operations perform more sophisticated tasks, such as MPI_Alltoall which rearranges n items of data such that the nth node gets the nth item of data from each.

Derived datatypes

Many MPI functions require that you specify the type of data which is sent between processes. This is because MPI aims to support heterogeneous environments where types might be represented differently on the different nodes, in which case MPI implementations can perform data conversion. Since the C language does not allow a type itself to be passed as a parameter, MPI predefines the constants MPI_INT, MPI_CHAR, MPI_DOUBLE to correspond with int, char, double, etc.
Here is an example in C that passes arrays of ints from all processes to one. The one receiving process is called the "root" process, and it can be any designated process but normally it will be process 0. All the processes ask to send their arrays to the root with MPI_Gather, which is equivalent to having each process call MPI_Send and the root make the corresponding number of ordered MPI_Recv calls to assemble all of these arrays into a larger one:

int send_array;
int root = 0; /* or whatever */
int num_procs, *recv_array;
MPI_Comm_size;
recv_array = malloc;
MPI_Gather / sizeof, MPI_INT,
recv_array, sizeof / sizeof, MPI_INT,
root, comm);

However, you may instead wish to send data as one block as opposed to 100 ints. To do this define a "contiguous block" derived data type:

MPI_Datatype newtype;
MPI_Type_contiguous;
MPI_Type_commit;
MPI_Gather;

For passing a class or a data structure, MPI_Type_create_struct creates an MPI derived data type from MPI_predefined data types, as follows:

int MPI_Type_create_struct

where:
The disp array is needed for data structure alignment, since the compiler may pad the variables in a class or data structure. The safest way to find the distance between different fields is by obtaining their addresses in memory. This is done with MPI_Get_address, which is normally the same as C's & operator but that might not be true when dealing with memory segmentation.
Passing a data structure as one block is significantly faster than passing one item at a time, especially if the operation is to be repeated. This is because fixed-size blocks do not require serialization during transfer.
Given the following data structures:

struct A ;
struct B ;

Here's the C code for building an MPI-derived data type:

static const int blocklen = ;
static const MPI_Aint disp = ;
static MPI_Datatype type = ;
MPI_Datatype newtype;
MPI_Type_create_struct / sizeof;
MPI_Type_commit;

MPI-2 concepts

One-sided communication

MPI-2 defines three one-sided communications operations, MPI_Put, MPI_Get, and MPI_Accumulate, being a write to remote memory, a read from remote memory, and a reduction operation on the same memory across a number of tasks, respectively. Also defined are three different methods to synchronize this communication as the specification does not guarantee that these operations have taken place until a synchronization point.
These types of call can often be useful for algorithms in which synchronization would be inconvenient, or where it is desirable for tasks to be able to balance their load while other processors are operating on data.

Dynamic process management

The key aspect is "the ability of an MPI process to participate in the creation of new MPI processes or to establish communication with MPI processes that have been started separately." The MPI-2 specification describes three main interfaces by which MPI processes can dynamically establish communications, MPI_Comm_spawn, MPI_Comm_accept/MPI_Comm_connect and MPI_Comm_join. The MPI_Comm_spawn interface allows an MPI process to spawn a number of instances of the named MPI process. The newly spawned set of MPI processes form a new MPI_COMM_WORLD intracommunicator but can communicate with the parent and the intercommunicator the function returns. MPI_Comm_spawn_multiple is an alternate interface that allows the different instances spawned to be different binaries with different arguments.

I/O

The parallel I/O feature is sometimes called MPI-IO, and refers to a set of functions designed to abstract I/O management on distributed systems to MPI, and allow files to be easily accessed in a patterned way using the existing derived datatype functionality.
The little research that has been done on this feature indicates that it may not be trivial to get high performance gains by using MPI-IO. For example, an implementation of sparse matrix-vector multiplications using the MPI I/O library shows a general behavior of negligible performance gain, but these results are inconclusive. It was not
until the idea of collective I/O implemented into MPI-IO that MPI-IO started to reach widespread adoption. Collective I/O substantially boosts applications' I/O bandwidth by having processes collectively transform the small and noncontiguous I/O operations into large and contiguous ones, thereby reducing the locking and disk seek overhead. Due to its vast performance benefits, MPI-IO also became the underlying I/O layer for many state-of-the-art I/O libraries, such as HDF5 and Parallel NetCDF. Its popularity also triggered a series of research efforts on collective I/O optimizations, such as layout-aware I/O and cross-file aggregation.

Official implementations

Many other efforts are derivatives of MPICH, LAM, and other works, including, but not limited to, commercial implementations from HP, Intel, Microsoft, and NEC.
While the specifications mandate a C and Fortran interface, the language used to implement MPI is not constrained to match the language or languages it seeks to support at runtime. Most implementations combine C, C++ and assembly language, and target C, C++, and Fortran programmers. Bindings are available for many other languages, including Perl, Python, R, Ruby, Java, and CL.

Hardware implementations

MPI hardware research focuses on implementing MPI directly in hardware, for example via processor-in-memory, building MPI operations into the microcircuitry of the RAM chips in each node. By implication, this approach is independent of the language, OS or CPU, but cannot be readily updated or removed.
Another approach has been to add hardware acceleration to one or more parts of the operation, including hardware processing of MPI queues and using RDMA to directly transfer data between memory and the network interface without CPU or OS kernel intervention.

Compiler wrappers

mpicc is a program that wraps over an existing compiler to set the necessary command-line flags when compiling code that uses MPI. Typically, it adds a few flags that enable the code to be the compiled and linked against the MPI library.

Language bindings

Bindings are libraries that extend MPI support to other languages by wrapping an existing MPI implementation such as MPICH or Open MPI.

Common Language Infrastructure

The two managed Common Language Infrastructure.NET implementations are Pure Mpi.NET and MPI.NET, a research effort at Indiana University licensed under a BSD-style license. It is compatible with Mono, and can make full use of underlying low-latency MPI network fabrics.

Java

Although Java does not have an official MPI binding, several groups attempt to bridge the two, with different degrees of success and compatibility. One of the first attempts was Bryan Carpenter's mpiJava, essentially a set of Java Native Interface wrappers to a local C MPI library, resulting in a hybrid implementation with limited portability, which also has to be compiled against the specific MPI library being used.
However, this original project also defined the mpiJava API which other subsequent Java MPI projects adopted. An alternative, less-used API is MPJ API, designed to be more object-oriented and closer to Sun Microsystems' coding conventions. Beyond the API, Java MPI libraries can be either dependent on a local MPI library, or implement the message passing functions in Java, while some like P2P-MPI also provide peer-to-peer functionality and allow mixed platform operation.
Some of the most challenging parts of Java/MPI arise from Java characteristics such as the lack of explicit pointers and the linear memory address space for its objects, which make transferring multidimensional arrays and complex objects inefficient. Workarounds usually involve transferring one line at a time and/or performing explicit de-serialization and casting at both sending and receiving ends, simulating C or Fortran-like arrays by the use of a one-dimensional array, and pointers to primitive types by the use of single-element arrays, thus resulting in programming styles quite far from Java conventions.
Another Java message passing system is MPJ Express. Recent versions can be executed in cluster and multicore configurations. In the cluster configuration, it can execute parallel Java applications on clusters and clouds. Here Java sockets or specialized I/O interconnects like Myrinet can support messaging between MPJ Express processes. It can also utilize native C implementation of MPI using its native device. In the multicore configuration, a parallel Java application is executed on multicore processors. In this mode, MPJ Express processes are represented by Java threads.

Julia language

The Julia language wrapper for MPI, has been used and found to be faster than for C or Fortran.

MATLAB

There are a few academic implementations of MPI using MATLAB. MATLAB has its own parallel extension library implemented using MPI and PVM.

OCaml

The OCamlMPI Module implements a large subset of MPI functions and is in active use in scientific computing. An eleven-thousand-line OCaml program was "MPI-ified" using the module, with an additional 500 lines of code and slight restructuring and ran with excellent results on up to 170 nodes in a supercomputer.

Python

MPI Python implementations include: pyMPI, mpi4py, pypar, MYMPI, and the MPI submodule in ScientificPython. pyMPI is notable because it is a variant python interpreter, while pypar, MYMPI, and ScientificPython's module are import modules. They make it the coder's job to decide where the call to MPI_Init belongs.
Recently the Boost C++ Libraries acquired Boost:MPI which included the MPI Python Bindings. This is of particular help for mixing C++ and Python. Boost:MPI's Python bindings still have unfixed packaging bugs in CentOS.

R

bindings of MPI include Rmpi and pbdMPI, where Rmpi focuses on manager-workers parallelism while pbdMPI focuses on SPMD parallelism. Both implementations fully support Open MPI or MPICH2.

Example program

Here is a "Hello World" program in MPI written in C. In this example, we send a "hello" message to each processor, manipulate it trivially, return the results to the main process, and print the messages.

/*
"Hello World" MPI Test Program
  • /
  1. include
  2. include
  3. include
  4. include
int main

When run with 4 processes, it should produce the following output:

$ mpicc example.c && mpiexec -n 4./a.out
We have 4 processes.
Process 1 reporting for duty.
Process 2 reporting for duty.
Process 3 reporting for duty.

Here, mpiexec is a command used to execute the example program with 4 processes, each of which is an independent instance of the program at run time and assigned ranks 0, 1, 2, and 3. The name mpiexec is recommended by the MPI standard, although some implementations provide a similar command under the name mpirun. The MPI_COMM_WORLD is the communicator that consists of all the processes.
A single program, multiple data programming model is thereby facilitated, but not required; many MPI implementations allow multiple, different, executables to be started in the same MPI job. Each process has its own rank, the total number of processes in the world, and the ability to communicate between them either with point-to-point communication, or by collective communication among the group. It is enough for MPI to provide an SPMD-style program with MPI_COMM_WORLD, its own rank, and the size of the world to allow algorithms to decide what to do. In more realistic situations, I/O is more carefully managed than in this example. MPI does not stipulate how standard I/O should work on a given system. It generally works as expected on the rank-0 process, and some implementations also capture and funnel the output from other processes.
MPI uses the notion of process rather than processor. Program copies are mapped to processors by the MPI runtime. In that sense, the parallel machine can map to 1 physical processor, or N where N is the total number of processors available, or something in between. For maximum parallel speedup, more physical processors are used. This example adjusts its behavior to the size of the world N, so it also seeks to scale to the runtime configuration without compilation for each size variation, although runtime decisions might vary depending on that absolute amount of concurrency available.

MPI-2 adoption

Adoption of MPI-1.2 has been universal, particularly in cluster computing, but acceptance of MPI-2.1 has been more limited. Issues include:
  1. MPI-2 implementations include I/O and dynamic process management, and the size of the middleware is substantially larger. Most sites that use batch scheduling systems cannot support dynamic process management. MPI-2's parallel I/O is well accepted.
  2. Many MPI-1.2 programs were developed before MPI-2. Portability concerns initially slowed, although wider support has lessened this.
  3. Many MPI-1.2 applications use only a subset of that standard with no real need for MPI-2 functionality.

    Future

Some aspects of the MPI's future appear solid; others less so. The MPI Forum reconvened in 2007, to clarify some MPI-2 issues and explore developments for a possible MPI-3, which resulted in versions MPI-3.0 and MPI-3.1.
Architectures are changing, with greater internal concurrency, better fine-grain concurrency control, and more levels of memory hierarchy. Multithreaded programs can take advantage of these developments more easily than single-threaded applications. This has already yielded separate, complementary standards for symmetric multiprocessing, namely OpenMP. MPI-2 defines how standard-conforming implementations should deal with multithreaded issues, but does not require that implementations be multithreaded, or even thread-safe. MPI-3 adds the ability to use shared-memory parallelism within a node. Implementations of MPI such as Adaptive MPI, Hybrid MPI, Fine-Grained MPI, MPC and others offer extensions to the MPI standard that address different challenges in MPI.
Astrophysicist Jonathan Dursi wrote an opinion piece that MPI is obsolescent, pointing to newer technologies like Chapel, Unified Parallel C, Hadoop, Spark and Flink.