MPI (Message Passing Interface)
This is a short introduction to the Message Passing Interface (MPI) system that was designed to enable parallel programming by communication on distributed-memory machines. MPI has become a standard for multiple-processor programming of code that runs on a variety of machines, from small Beowulf installations to shared-memory high-performance machines and hybrid clusters such as the ones at HPCVL. Due to the complexity of the system, it is of course impossible to give a detailed guide or even tutorial to MPI-programming in the framework of a "Frequently Asked Questions" file. This is not an introduction to MPI programming. References and links for further details are given.
What is the Message Passing Interface (MPI) ?
The Message Passing Interface is a communication system that was designed by a group of researchers to supply programmers with a standard for distributed-memory parallel programming that is portable and usable on a variety of platforms. The standardization process was initiated at a Workshop in 1992, and a draft of the standard was available a year later. Version 1.0 was released in the Summer of 1994. The current version of MPI is 3.0.
At present, the system is implemented in the form of "bindings" for Fortran, C, and C++. MPI-1 (i.e. the original version 1.0 of the standard) consisted of about 100 functions/subroutines which form the MPI core and were available for Fortran and C. MPI-2 approximately doubles the number of routines, adding new features (which we will not discuss in this FAQ), and introducing C++ bindings.
The number of essential MPI routines is quite limited. In our experience, well-functioning MPI programs can be written with a dozen or so. The routines are supplied in the form of libraries, header files and usually scripts for easy compilation. Almost any hardware/OS platform supplies a native MPI version these days, and public-domain versions are available as well.
Here is an overview of the main components of MPI-1:
MPI-2 adds to this:
What kind of system uses MPI ?
MPI was designed to handle distributed-memory systems, i.e. clusters. This does not mean that its usage is restricted to such systems. In fact, there are good arguments for MPI running even better on shared-memory machines. It will of course also be usable on hybrids such as the HPCVL Sunfire cluster.
Since MPI provides a means to enable communication between different CPUs, it does not depend on shared-memory architectures, as is the case for multi-threading systems such as OpenMP. On the other hand, it can make use of shared memory for fast amd improved communication.
Note that the status of MPI as a distributed-memory system implies that multiple processes are started from the beginning and run, usually on different CPUs to completion. These processes do not have anything in common, and each has its own memory space. Any information exchange requires communication of data, for which MPI was designed.
MPI is a set of subroutines which are used explicitly to communicate between processes. As such, MPI programs are truly "multi-processing". Parallelisation can not be done automatically or semi-automatically as in "multi-threading" programs. Instead, function and subroutine calls have to be inserted into the code and form an integral part of the program. Often it is beneficial to alter the algorithm of the code with respect to the serial version.
The need to include the parallelism explicitly in the program is both a curse and a blessing: while it means more work and requires more planning than multi-threading, it also often leads to more reliable and scalable code since the behaviour of the latter is in the hands of the programmer. Well-written MPI codes can be made to scale for thousands of CPUs.
To create an MPI program, you need to:
The working principle of MPI is perhaps best illustrated on the grounds of a programming example. The following program, written in Fortran 90 computes the sum of all square-roots of integers from 0 up to a specific limit m:
module mpi include 'mpif.h' end module mpi module cpuids integer::myid,totps, ierr end module cpuids program example02 use mpi use cpuids call mpiinit call demo02 call mpi_finalize(ierr) stop end subroutine mpiinit use mpi use cpuids call mpi_init( ierr ) call mpi_comm_rank(mpi_comm_world,myid,ierr) call mpi_comm_size(mpi_comm_world,totps,ierr) return end subroutine demo02 use mpi use cpuids integer:: m, i real*8 :: s, mys if(myid.eq.0) then write(*,*)'how many terms?' read(*,*) m end if call mpi_bcast(m,1,mpi_integer,0,mpi_comm_world,ierr) mys=0.0d0 do i=myid,m,totps mys=mys+dsqrt(dfloat(i)) end do write(*,*)'rank:', myid,'mys=',mys, ' m:',m s=0.0d0 call mpi_reduce(mys,s,1,mpi_real8,mpi_sum,0,mpi_comm_world,ierr) if(myid.eq.0) then write(*,*)'total sum: ', s end if return end
Some of the common tasks that need to be performed in every MPI program are done in the subroutine mpiinit in this program. Namely, we need to call the routine mpi_init to prepare the usage of MPI. This has to be done before any other MPI routine is called. The two routine calls to mpi_comm_size and call mpi_comm_rank determine how many processes are running and what is the unique ID number of the present, i.e. the calling process. Both pieces of information are essential. The results are stored in the variables totps and myid, respectively. Note that these variables appear in a module cpuids so that they may be accessed from all routines that "use" that module.
The main work in the example is done in the subroutine demo02. Note that this routine does use the module cpuids. The first operation is to determine the maximum integer m in the sum by requesting input from the user. The if-clause if(myid.eq.0) then serves to restrict this I/O operation to only one process, the so-called "root process", usually chosen to be the one with rank (i.e. unique ID number) zero.
After this initial operation, communication has become necessary, since only one process has the right value of m. This is done by a call to the MPI collective operation routine mpi_bcast. This call has the effect of "broadcasting" the integer m. This call needs to be made by all processes, and after they have done so, all of them know m.
The sum over the square root is then executed on each process in a slightly different manner. Each term is added to a local variable mys. A stride of totps (the number of processes) in the do-loop ensures that each process adds different terms to its local sum, by skipping all others. For instance, if there are 10 processes, process 0 will add the square-roots of 0,10,20,30,..., while process 7 will add the square-roots of 7,17,27,37,...
After the sums have been completed, further communication is necessary, since each process only has computed a partial, local sum. We need to collect these local sums into one total, and we do so by calling mpi_reduce. The effect of this call is to "reduce" a value local to each process to a variable that is local to only one process, usually the root process. We can do this in various ways, but in our case we choose to sum the values up by specifying mpi_sum in the function call. Afterwards, the total sum resides in the variable s, which is printed out by the root process.
The last operation done in our example is finalizing MPI usage by a call to mpi_finalize, which is necessary for proper program completion.
In this simple example, we have distributed the tasks of computing many square roots among processes, each of which only did a part of the work. We used communication to exchange information about the tasks that needed to be performed, and to collect results. This mode of programming is called "task parallel". Often it is necessary to distribute large amounts of data among processes as well, leading to "data parallel" programs. Of course, the distinction is not always clear.
While MPI itself is a portable, platform independent standard, much like a programming language, the actual implementation is necessarily platform dependent since it has to take into account the architecture of the machine or cluster in question.
The most commonly used implementation of MPI for the Linux platform is called OpenMPI. The following considerations will be focussed on this implementation.
Our machines are small to mid-sized shared-memory machines that form a cluster. Since the interconnect between the individual nodes is a bottleneck in efficient program execution, most of the MPI programs running on our machines are executed within a node. This alloows processes to commuincate rapidly through a so-called "shared-memory layer". Our cluster is configured in to preferably schedule processes within a single node.
Currently, two versions of the OpenMPI parallel environment are in common use:
We do not recommend to have both versions set up simultaneously.
Compiling MPI code
The compilation of MPI programs requires a few compiler options to direct the compiler to the location of header files and libraries. Since these switches are always the same, they have been collected in a macro to avoid unnecessary typing. The macro is has an mpi prefix before the normal compiler name. The commands are mpiifort for the Intel Fortran compiler, mpiicc for the gnu C compilers, respectively. For instance, if a serial C program is compiled by
gcc -O3 -c test.c
the corresponding parallel (MPI) program is compiled (using gnu compiler) by
mpicc -xO3 -c test_mpi.c
In the linking stage, the usage of mpi* macros also includes the proper specification of the MPI libraries. For example, the above MPI program should be linked with something like this:
mpicc -o test_mpi.exe test_mpi.o
Compiling and linking may also be combined by omitting the -c option and including the naming option (-o) in the compilation line.
Here are the corresponding MPI macros for the 6 commonly used compilers on our systems:
Running MPI programs
To run MPI programs, a special Runtime Environment is required. This includes commands for the control of multi-process jobs.
mpirun is used to start a multi-process run of a program. This required to run MPI programs. The most commonly used command line option is -np to specify the number of processes to be started. For instance, the following line will start the program test_mpi.exe with 9 processes:
mpirun -np 9 test_mpi.exe
The mpirun command offers additional options that are sometimes useful or required. Most tend to interfere with the scheduling of jobs in a multi-user environment such as ours and should be used with caution. Please consult the man pages for details.
Note that the usage of a scheduler is mandatory for production jobs on our system. This option is therefore used frequently. For a details about Gridengine and jobs submission on HPCVL machines and clusters, go here.
As already pointed out, this FAQ is not an introduction to MPI programming. The standard reference text on MPI is:
Marc Snir, Steve Otto, Steven Huss-Lederman, David Walker, and Jack Dongarra:
This text specifies all MPI routines and concepts, and includes a large number of examples. Most people will find it sufficient for all their needs.
A quite good online tutorial for MPI programming can be found at the Maui HPCC site.
There is also an official MPI webpage which contains the standards documents for MPI and gives access to the MPI Forum.
Standard debugging and profiling tools such as Sun Studio are designed for serial or multi-threaded programs. They do not handle multi-process runs very well.
Quite often, the best way to check the performance of an MPI program is timing it by insertion of suitable routines. MPI supplies a "wall-clock" routine called MPI_WTIME(), that lets you determine how much actual time was spent in a specific segment of your code. An other method is calling the subroutines ETIME and DTIME, which can give you information about the actual CPU time used. However, it is advisable to carefully read the documentation before using them with MPI programs. In this case, refer to the Sun Studio 12: Fortran Library Reference.
We also provide a package called the HPCVL Working Template (HWT), which was created by Gang Liu. The HWT provides 3 main functionalities:
The HWT is based on libraries and script files. It is easy to use and portable (written largely in Fortran). Fortran, C, C++, and any mixture thereof are supported, as well as MPI and OpenMP for parallelism. Documentation of the HWT is available. The package is installed on our clusters in /opt/hwt.