What is MPI? MPI = Message Passing InterfaceSpecification of message passing libraries for developers and usersNot a library by itself, but specifies what such a library should beSpecifies application programming interface (API) for such librariesMany libraries implement such APIs on different platforms MPI librariesGoal: provide a standard for writing message passing programsPortable, efficient, flexibleLanguage binding: C, C++, FORTRAN programs

History & Evolution1980s 1990s: incompatible libraries and software tools; need for a standard1994, MPI 1.0; 1995, MPI 1.1, revision and clarification to MPI 1.0Major milestoneC, FORTRANFully implemented in all MPI libraries1997, MPI 1.2Corrections and clarifications to MPI 1.11997, MPI 2Major extension (and clarifications) to MPI 1.1C++, C, FORTRANPartially implemented in most libraries; a few full implementations (e.g. ANL MPICH2)MPI Evolution

Why Use MPI?Standardization: de facto standard for parallel computingNot an IEEE or ISO standard, but industry standardPractically replaced all previous message passing librariesPortability: supported on virtually all HPC platformsNo need to modify source code when migrating to different machinesPerformance: so far the best; high performance and high scalabilityRich functionality:MPI 1.1 125 functionsMPI 2 152 functions.If you know 6 MPI functions, you can do almost everything in parallel.

Programming ModelMessage passing model: data exchange through explicit communications.For distributed memory, as well as shared-memory parallel machinesUser has full control (data partition, distribution): needs to identify parallelism and implement parallel algorithms using MPI function calls.Number of CPUs in computation is staticNew tasks cannot be dynamically spawned during run time (MPI 1.1)MPI 2 specifies dynamic process creation and management, but not available in most implementations.Not necessarily a disadvantageGeneral assumption: one-to-one mapping of MPI processes to processors (although not necessarily always true).

MPI 1.1 OverviewPoint to point communicationsCollective communicationsProcess groups and communicatorsProcess topologiesMPI environment management

MPI 2 OverviewDynamic process creation and managementOne-sided communicationsMPI Input/Output (Parallel I/O)Extended collective communicationsC++ binding

MPI ResourcesMPI Standard:http://www.mpi-forum.org/MPI web sites/tutorials etc, see class web sitePublic domain (free) MPI implementationsMPICH and MPICH2 (from ANL)LAM MPI

General MPI Program Structure

Example#include #include

int main(int argc, char **argv){ int my_rank, num_cpus;

MPI_Init(&argc, &argv); MPI_Comm_rank(MPI_COMM_WORLD, &my_rank); MPI_Comm_size(MPI_COMM_WORLD, &num_cpus); printf(Hello, I am process %d among %d processes\n, my_rank, num_cpus); MPI_Finalize(); return 0;}Hello, I am process 1 among 4 processesHello, I am process 2 among 4 processesHello, I am process 0 among 4 processesHello, I am process 3 among 4 processesOn 4 processors:

Exampleprogram helloimplicit none

include mpif.hinteger :: ierr, my_rank, num_cpus

call MPI_INIT(ierr)call MPI_COMM_RANK(MPI_COMM_WORLD, my_rank)call MPI_COMM_SIZE(MPI_COMM_WORLD, num_cpus)write(*,*) Hello, I am process , my_rank, among & , num_cpus, processescall MPI_FINALIZE(ierr)

end program helloHello, I am process 1 among 4 processesHello, I am process 2 among 4 processesHello, I am process 0 among 4 processesHello, I am process 3 among 4 processesOn 4 processors:

MPI Header FilesIn C/C++:

In FORTRAN:#include include mpif.hor (in FORTRAN90 and later)use MPI

MPI Naming ConventionsAll names have MPI_ prefix.In FORTRAN:All subroutine names upper case, last argument is return code

A few functions without return code

In C: mixed uppercase/lowercase

MPI constants all uppercasecall MPI_XXXX(arg1,arg2,,ierr)call MPI_XXXX_XXXX(arg1,arg2,,ierr)ierr = MPI_Xxxx(arg1,arg2,);ierr = MPI_Xxxx_xxx(arg1,arg2,);MPI_COMM_WORLD, MPI_SUCCESS, MPI_DOUBLE, MPI_SUM, If ierr == MPI_SUCCESS,Everything is ok; otherwise, something is wrong.

InitializationInitialization: MPI_Init() initializes MPI environment; (MPI_Init_thread() if multiple threads)Must be called before any other MPI routine (so put it at the beginning of code) except MPI_Initialized() routine.Can be called only once; subsequent calls are erroneous.MPI_Initialized() to check if MPI_Init() is calledint main(int argc, char ** argv){ MPI_Init(&argc, &argv); int flag; MPI_Initialized(&flag); if(flag != 0) // MPI_Init called MPI_Finalize(); return 0;}int MPI_Init(int *argc, char ***argv)program testinteger ierrcall MPI_INIT(ierr)call MPI_FINALIZE(ierr)end program testMPI_INIT(ierr)

TerminationMPI_Finalize() cleans up MPI environmentMust be called before exits.No other MPI routine can be called after this call, even MPI_INIT()Exception: MPI_Initialized() (and MPI_Get_version(), MPI_Finalized()).Abnormal termination: MPI_Abort()Makes a best attempt to abort all tasks

int MPI_Finalize(void)MPI_FINALIZE(IERR) integer IERR

int MPI_Abort(MPI_Comm comm, int errorcode)MPI_ABORT(COMM,ERRORCODE,IERR) integer COMM, ERRORCODE, IERR

MPI ProcessesMPI is process-oriented: program consists of multiple processes, each corresponding to one processor.MIMD: Each process runs its own code. In practice, runs its own copy of the same code (SPMD).MPI process and threads: MPI process can contain a single thread (common case) or multiple threads.Most MPI implementations do not support multiple threads. Needs special processing with that support.We will assume a single thread per process from now on.MPI processes are identified by their ranks:If total nprocs processes in computation, rank ranges from 0, 1, , nprocs-1. (true in C and FORTRAN).nprocs does not change during computation.

Communicators and Process GroupsCommunicator: is a group of processes that can communicate with one another.Most MPI routines require a communicator argument to specify the collection of processes the communication is based on.All processes in the computation form the communicator MPI_COMM_WORLD.MPI_COMM_WORLD is pre-defined by MPI, available anywhereCan create subgroups/subcommunicators within MPI_COMM_WORLD.A process may belong to different communicators, and have different ranks in different communicators.

How many CPUs, Which one am I How many CPUs: MPI_COMM_SIZE()Who am I: MPI_COMM_RANK()Can compute data decomposition etc.Know total number of grid points, total number of cpus and current cpu id; can calc which portion of data current cpu is to work on.E.g. Poisson equ on a squareRanks also used to specify source and destination of communications.int my_rank, ncpus;MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);MPI_Comm_size(MPI_COMM_WORLD, &ncpus);int MPI_Comm_rank(MPI_Comm comm, int *rank)int MPI_Comm_size(MPI_Comm comm, int *size)MPI_COMM_RANK(comm,rank,ierr)MPI_COMM_SIZE(comm,size,ierr)my_rank value different on different processors !

Compiling, RunningMPI standard does not specify how to start up the programCompiling and running MPI code implementation dependentMPI implementations provide utilities/commands for compiling/running MPI codesCompile: mpicc, mpiCC, mpif77, mpif90, mpCC, mpxlf mpiCC o myprog myfile.C (cluster)mpif90 o myprog myfile.f90 (cluster)CC Ipath_mpi_include o myprog myfile.C lmpi (SGI)mpCC o myprog myfile.C (IBM)Run: mpirun, poe, prun, ibrun mpirun np 2 myprog (cluster)mpiexec np 2 myprog (cluster)poe myprog node 1 tasks_per_node 2 (IBM)

Example#include #include #include

int main(int argc, char **argv){ char message[256]; int my_rank; MPI_Status status;

MPI_Init(&argc, &argv); MPI_Comm_rank(MPI_COMM_WORLD,&my_rank); if(my_rank==0){ strcpy(message,Hello, there!); MPI_Send(message,strlen(message)+1,MPI_CHAR,1,99,MPI_COMM_WORLD); } else if(my_rank==1) { MPI_Recv(message,256,MPI_CHAR,0,99,MPI_COMM_WORLD,&status); printf(Process %d received: %s\n,my_rank,message); } MPI_Finalize(); return 0;}mpirun np 2 test_hello

Process 1 received: Hello, there!6 MPI functions:MPI_Init()MPI_Finalize()MPI_Comm_rank()MPI_Comm_size()MPI_Send()MPI_Recv()

MPI CommunicationsPoint-to-point communicationsInvolves a sender and a receiver, one processor to another processorOnly the two processors participate in communicationCollective communicationsAll processors within a communicator participate in communication (by calling same routine, may pass different arguments); Barrier, reduction operations, gather,

Point-to-Point Communications

Send / ReceiveMessage data: what to send/receive?Where is the message? Where to put it?What kind of data is it? What is the size?Message envelope: where to send/receive?Sender, receiverCommunication contextMessage tag.

MPI_Send(message,strlen(message)+1,MPI_CHAR,1,99,MPI_COMM_WORLD);MPI_Recv(message,256,MPI_CHAR,0,99,MPI_COMM_WORLD,&status);

Sendbuf memory address of start of messagecount number of data itemsdatatype what type each data item is (integer, character, double, float )dest rank of receiving processtag additional identification of messagecomm communicator, usually MPI_COMM_WORLDint MPI_Send(void *buf,int count,MPI_Datatype datatype, int dest, int tag, MPI_Comm comm)MPI_SEND(BUF,COUNT,DATATYPE,DEST,TAG,COMM,IERROR) BUF(*) integer COUNT,DATATYPE,DEST,TAG,COMM,IERRORchar message[256];MPI_Send(message,strlen(message)+1,MPI_CHAR,1,99,MPI_COMM_WORLD);

Receivebuf initial address of receive buffercount number of elements in receive buffer (size of receive buffer)may not equal to the count of items actually receivedActual number of data items received can be obtained by calling MPI_Get_count().datatype data type in receive buffersource rank of sending processtag additional identification for messagecomm communicator, usually MPI_COMM_WORLDstatus object containing additional info of received messageierror return codeint MPI_Recv(void *buf,int count,MPI_Datatype datatype,int source,int tag, MPI_Comm comm,MPI_Status *status)MPI_RECV(BUF,COUNT,DATATYPE,SOURCE,TAG,COMM,STATUS,IERROR) BUF(*) integer COUNT,DATATYPE,SOURCE,TAG,COMM,STATUS(MPI_STATUS_SIZE),IERRORActual number of data items received can be queried from status object; it may be smaller than count, but cannot be larger (if larger overflow error).char message[256];MPI_Recv(message,256,MPI_CHAR,0,99,MPI_COMM_WORLD,&status);

MPI_Recv StatusIn C: MPI_Status structure, 3 members; MPI_Status statusstatus.MPI_TAG tag of received messagestatus.MPI_SOURCE source rank of messagestatus.MPI_ERROR error codeIn FORTRAN: integer array; integer status(MPI_STATUS_SIZE)Status(MPI_TAG) tag of received messagestatus(MPI_SOURCE) source rank of messagestatus(MPI_ERROR) error codeLength of received message: MPI_Get_count()Int MPI_Get_count(MPI_Status *status, MPI_Datatype datatype, int *count)MPI_GET_COUNT(STATUS,DATATYPE,COUNT,IERROR) integer STATUS(MPI_STATUS_SIZE),DATATYPE,COUNT,IERRORMPI_Status status;int count;MPI_Recv(message,256,MPI_CHAR,0,99,MPI_COMM_WORLD,&status);MPI_Get_count(&status, MPI_CHAR, &count); // count contains actual length

Message DataConsists of count successive entries of the type indicated by datatype, starting with the entry at the address buf.MPI data types:Basic data types: one for each data type in hosting languages of C/C++, FORTRANDerived data type: will learn later.

Basic MPI Data Types

Exampleint num_students;double grade[10];char note[1024];int tag1=1001, tag2=1002;int rank,ncpus;MPI_Status status;MPI_Comm_rank(MPI_COMM_WORLD,&rank); // set num_students,grade,note in rank=0 cpuif(rank==0){ MPI_Send(&num_students,1,MPI_INT,1,tag1,MPI_COMM_WORLD); MPI_Send(grade, 10, MPI_DOUBLE,2,tag1,MPI_COMM_WORLD); MPI_Send(note,strlen(note)+1,MPI_CHAR,1,tag2,MPI_COMM_WORLD);}if(rank==1){ MPI_Recv(&num_students,1,MPI_INT,0,tag1,MPI_COMM_WORLD,&status); MPI_Recv(note,1024,MPI_CHAR,0,tag2,MPI_COMM_WORLD,&status);}if(rank==2){ MPI_Recv(grade,10,MPI_DOUBLE,0,tag1,MPI_COMM_WORLD,&status);}num_students: 0 1note: 0 1grade: 0 2

Message EnvelopeEnvelope:SourceDestinationTagCommunicatorMPI_Send: source is the calling process, destination process specified;Destination can be MPI_PROC_NULL will return asap, no effect.MPI_Recv: destination is the calling process, source is specifiedSource can be a wildcard, MPI_ANY_SOURCE.Source can also be MPI_PROC_NULL - Return asap, no effect, receive buffer not modified.Tag: non-negative number, 0, 1, , UB; UB can be determined by querying MPI environment (UB>=32767).For MPI_Recv, can be a wildcard, MPI_ANY_TAG.Communicator: specified, usually MPI_COMM_WORLD.

In Order for a Message To be Received Message envelopes must matchMessage must be directed to the process calling MPI_RecvMessage source must match that specified by MPI_Recv, unless MPI_ANY_SOURCE is specified.Message tag must match that specified by MPI_Recv, unless MPI_ANY_TAG is specifiedMessage communicator must match that specified by MPI_Recv.Data type must matchDatatype specified by MPI_Send and MPI_Recv must match.(MPI_PACKED can match any other data type.)Can be more complicated when derived data types are involved.

Exampledouble A[10], B[15];int rank, tag = 1001, tag1=1002;MPI_Status status;MPI_Comm_rank(MPI_COMM_WORLD,&rank);if(rank==0) MPI_Send(A, 10, MPI_DOUBLE, 1, tag, MPI_COMM_WORLD);else if(rank==1){ MPI_Recv(B, 15, MPI_DOUBLE, 0, tag, MPI_COMM_WORLD, &status); // ok // MPI_Recv(B, 15, MPI_FLOAT, 0, tag, MPI_COMM_WORLD, &status); wrong // MPI_Recv(B,15,MPI_DOUBLE,0,tag1,MPI_COMM_WORLD,&status); un-match // MPI_Recv(B,15,MPI_DOUBLE,1,tag,MPI_COMM_WORLD,&status); un-match // MPI_Recv(B,15,MPI_DOUBLE,MPI_ANY_SOURCE,tag,MPI_COMM_WORLD,&status); ok // MPI_Recv(B,15,MPI_DOUBLE,0,MPI_ANY_TAG,MPI_COMM_WORLD,&status); ok}A: 01

Blocking Send/RecvMPI_Send is blocking: will not return until the message data and envelope is safely stored away.The message data might be delivered to the matching receive buffer, or copied to some temporary system buffer.After MPI_Send returns, user can safely access or overwrite the send buffer.MPI_Recv is blocking: returns only after the receive buffer has the received messageAfter it returns, the data is here and ready for use.Non-blocking send/recv: will be discussed later.Non-blocking calls will return immediately; however, not safe to access the send/receive buffers. Need to call other functions to complete send/recv, then safe to access/modify send/receive buffers.

BufferingSend and matching receive operations may not be (and are not) synchronized in reality. MPI implementation must decide what happens when send/recv are out of sync.Consider:Send occurs 5 seconds before receive is ready; where is the message when receive is pending?Multiple sends arrive at the same receiving task which can receive one send at a time what happens to the messages that are backing up?MPI implementation (not the MPI standard) decides what happens in these cases. Typically a system buffer is used to hold data in transit.

BufferingSystem buffer:Invisible to users and managed by MPI libraryFinite resource that can be easily exhaustedMay exist on sending or receiving side, or bothMay improve performance.User can attach own buffer for MPI message buffering.

Communication Modes for SendStandard mode: MPI_SendSystem decides whether the outgoing message will be buffered or notUsually, small messages buffering mode; large messages, no buffering, synchronous mode.Buffered mode: MPI_BsendMessage will be copied to buffer; Send call then returnsUser can attach own buffer for useSynchronous mode: MPI_SsendNo buffering.Will block until a matching receive starts receiving dataReady mode: MPI_RsendCan be used only if a matching receive is already posted; avoid handshake etc. otherwise erroneous.

Communication ModesMPI_Send(void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm)MPI_Bsend(void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm)MPI_Ssend(void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm)MPI_Rsend(void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm)There is only one MPI_Recv; will match any send mode

PropertiesOrder: MPI messages are non-overtakingIf a sender sends two messages in succession to same destination, and both match the same receive, then this receive will receive the first message no matter which message physically arrives the receiving end first.If a receiver posts two receives in succession and both match the same message, then the first receive will be satisfied.Note: if a receive matches two messages from two different senders, the receive may receive either one (implementation dependent).MPI_Comm_rank(MPI_COMM_WORLD, &rank);if(rank==0) { MPI_Bsend(buf1,count,MPI_DOUBLE,1,tag,comm); MPI_Bsend(buf2,count,MPI_DOUBLE,1,tag,comm);}else if(rank==1) { MPI_Recv(buf1,count,MPI_DOUBLE,0,MPI_ANY_TAG,comm,&status); MPI_Recv(buf2,count,MPI_DOUBLE,0,tag,comm,&status);}

PropertiesProgress: if a pair of matching send/recv are initiated on two processes, at least one of them will completeSend will complete, unless the receive is satisfied by some other messageReceive will complete, unless send is consumed by some other matching receive.

MPI_Comm_rank(MPI_COMM_WORLD, &rank);If(rank==0) { MPI_Bsend(buf1,count,MPI_DOUBLE,1,tag1,comm); MPI_Ssend(buf2,count,MPI_DOUBLE,1,tag2,comm);} else if(rank==1) { MPI_Recv(buf1,count,MPI_DOUBLE,0,tag2,comm); MPI_Recv(buf2,count,MPI_DOUBLE,0,tag1,comm);}

PropertiesFairness: no guarantee of fairnessIf a message is sent to a destination, and the destination process repeatedly posts a receive that matches this send, however the message may never be received since it is each time overtaken by another message sent from another source.It is users responsibility to prevent the starvation in such situations.

PropertiesResource limitation:Pending communications consume system resources (e.g. buffer space)Lack of resource may cause error or prevent execution of MPI call.e.g. MPI_Bsend that cannot complete due to lack of buffer space is erroneous.MPI_Send that cannot complete due to lack of buffer space will only block, waiting for buffer space to be available or for a matching receive.

DeadlockDeadlock is a state when the program cannot proceed.Cyclic dependencies cause deadlock

MPI_Comm_rank(MPI_COMM_WORLD,&rank);If(rank==0){ MPI_Recv(buf1,count,MPI_DOUBLE,1,tag,comm); MPI_Send(buf2,count,MPI_DOUBLE,1,tag,comm);} else if (rank==1) { MPI_Recv(buf1,count,MPI_DOUBLE,0,tag,comm); MPI_Send(buf2,count,MPI_DOUBLE,0,tag,comm);}MPI_Comm_rank(MPI_COMM_WORLD,&rank);If(rank==0){ MPI_Ssend(buf1,count,MPI_DOUBLE,1,tag,comm); MPI_Recv(buf2,count,MPI_DOUBLE,1,tag,comm);} else if (rank==1) { MPI_Ssend(buf1,count,MPI_DOUBLE,0,tag,comm); MPI_Recv(buf2,count,MPI_DOUBLE,0,tag,comm);}

DeadlockLack of buffer space may also cause deadlockMPI_Comm_rank(MPI_COMM_WORLD,&rank);If(rank==0){ MPI_Send(buf1,count,MPI_DOUBLE,1,tag,comm); MPI_Recv(buf2,count,MPI_DOUBLE,1,tag,comm);} else if (rank==1) { MPI_Send(buf1,count,MPI_DOUBLE,0,tag,comm); MPI_Recv(buf2,count,MPI_DOUBLE,0,tag,comm);}Deadlock if not enough buffer space!

Send-receiveTwo remedies: non-blocking communication, send-recvMPI_SENDRECV: combine send and recv in one callUseful in shift operations; Avoid possible deadlock with circular shift and like operationsEquivalent to: execute a nonblocking send and a nonblocking recv, and then wait for them to complete

int MPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, int dest, int sendtag, void *recvbuf, int recvcount, MPI_Datatype recvtype, int source, int recvtag, MPI_Comm comm, MPI_Status *status)

MPI_SENDRECV(SENDBUF, SENDCOUNT, SENDTYPE, DEST, SENDTAG, RECVBUF, RECVCOUNT, RECVTYPE, SOURCE, RECVTAG, COMM, STATUS, IERROR) SENDBUF(*), RECVBUF(*) INTEGER SENDCOUNT, SENDTYPE, SENDTAG, DEST, RECVCOUNT, RECVTYPE, SOURCE, RECVTAG, COMM, IERROR, STATUS(MPI_STATUS_SIZE)*sendbuf and *recvbuf may not be the same memory address

Example#include #include

int main(int argc, char **argv){ int my_rank, ncpus; int left_neighbor, right_neighbor; int data_received=-1; int send_tag = 101, recv_tag=101; MPI_Status status;

MPI_Init(&argc, &argv); MPI_Comm_rank(MPI_COMM_WORLD, &my_rank); MPI_Comm_size(MPI_COMM_WORLD, &ncpus);

left_neighbor = (my_rank-1 + ncpus)%ncpus; right_neighbor = (my_rank+1)%ncpus;

MPI_Sendrecv(&my_rank, 1, MPI_INT, left_neighbor, send_tag, &data_received, 1, MPI_INT, right_neighbor, recv_tag, MPI_COMM_WORLD, &status); printf("Among %d processes, process %d received from right neighbor: %d\n", ncpus, my_rank, data_received);

// clean up MPI_Finalize(); return 0;} mpirun np 4 test_shift

Among 4 processes, process 3 received from right neighbor: 0Among 4 processes, process 2 received from right neighbor: 3Among 4 processes, process 0 received from right neighbor: 1Among 4 processes, process 1 received from right neighbor: 2