300 likes | 410 Views
MPI: Message Passing Interface. Prabhaker Mateti Wright State University. Overview. MPI Hello World! Introduction to programming with MPI MPI library calls. MPI Overview. Similar to PVM Network of Heterogeneous Machines Multiple implementations Open source: MPICH LAM Vendor specific.
E N D
MPI: Message Passing Interface Prabhaker Mateti Wright State University
Overview • MPI Hello World! • Introduction to programming with MPI • MPI library calls Mateti, MPI
MPI Overview • Similar to PVM • Network of Heterogeneous Machines • Multiple implementations • Open source: • MPICH • LAM • Vendor specific Mateti, MPI
MPI Features • Rigorously specified standard • Portable source code • Enables third party libraries • Derived data types to minimize overhead • Process topologies for efficiency on MPP • Van fully overlap communication • Extensive group communication Mateti, MPI
MPI 2 • Dynamic Process Management • One-Sided Communication • Extended Collective Operations • External Interfaces • Parallel I/O • Language Bindings (C++ and Fortran-90) • http://www.mpi-forum.org/ Mateti, MPI
MPI Overview • 125+ functions • typical applications need only about 6 Mateti, MPI
#include <mpi.h> main(int argc, char *argv[]) { int myrank; MPI_Init(&argc, &argv); MPI_Comm_rank(MPI_COMM_WORLD, &myrank); if (myrank == 0) manager(); else worker(); MPI_Finalize(); } MPI_Initinitializes the MPI system MPI_Finalizecalled last by all processes MPI_Comm_rank identifies a process by its rank MPI_COMM_WORLD is the group that this process belongs to MPI: manager+workers Mateti, MPI
manager() { MPI_Status status; MPI_Comm_size( MPI_COMM_WORLD, &ntasks); for (i = 1;i < ntasks;++i){ work= nextWork(); MPI_Send(&work, 1, MPI_INT,i,WORKTAG, MPI_COMM_WORLD); } … MPI_Reduce(&sub, &pi, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD); } MPI_Comm_size MPI_Send MPI: manager() Mateti, MPI
worker() { MPI_Status status; for (;;) { MPI_Recv(&work, 1, MPI_INT, 0, MPI_ANY_TAG, MPI_COMM_WORLD, &status); result = doWork(); MPI_Send(&result, 1, MPI_DOUBLE, 0, 0, MPI_COMM_WORLD); } } MPI_Recv MPI: worker() Mateti, MPI
#include "mpi.h" int main(int argc, char *argv[]) { MPI_Init(&argc,&argv); MPI_Comm_size(MPI_COMM_WORLD,&np); MPI_Comm_rank(MPI_COMM_WORLD,&myid); n = ...; /* intervals */ MPI_Bcast(&n, 1, MPI_INT, 0, MPI_COMM_WORLD); sub = series_sum(n, np); MPI_Reduce(&sub, &pi, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD); if (myid == 0) printf("pi is %.16f\n", pi); MPI_Finalize(); return 0; } MPI computes Mateti, MPI
Process groups • Group membership is static. • There are no race conditions caused by processes independently entering and leaving a group. • New group formation is collective and group membership information is distributed, not centralized. Mateti, MPI
MPI_Send blocking send MPI_Send( &sendbuffer, /* message buffer */ n, /* n items of */ MPI_type, /* datatype in message */ destination, /* process rank */ WORKTAG, /* user chosen tag */ MPI_COMM /* group */ ); Mateti, MPI
MPI_Recv blocking receive MPI_Recv( &recvbuffer, /* message buffer */ n, /* n data items */ MPI_type, /* of type */ MPI_ANY_SOURCE, /* from any sender */ MPI_ANY_TAG, /* any type of message */ MPI_COMM, /* group */ &status ); Mateti, MPI
Send-receive succeeds … • Sender’s destination is a valid process rank • Receiver specified a valid source process • Communicator is the same for both • Tags match • Message data types match • Receiver’s buffer is large enough Mateti, MPI
Message Order • P sends messages m1 first then m2 to Q • Q will receive m1 before m2 • P sends m1 to Q, then m2 to R • In terms of a global wall clock, conclude nothing re R receiving m2 before/after Q receiving m1. Mateti, MPI
Blocking and Non-blocking • Send, receive can be blocking or not • A blocking send can be coupled with a non-blocking receive, and vice-versa • Non-blocking send can use • Standard mode MPI_Isend • Synchronous mode MPI_Issend • Buffered mode MPI_Ibsend • Ready mode MPI_Irsend Mateti, MPI
MPI_Isend non-blocking MPI_Isend( &buffer, /* message buffer */ n, /* n items of */ MPI_type, /* datatype in message */ destination, /* process rank */ WORKTAG, /* user chosen tag */ MPI_COMM, /* group */ &handle ); Mateti, MPI
MPI_Irecv MPI_Irecv( &result, /* message buffer */ n, /* n data items */ MPI_type, /* of type */ MPI_ANY_SOURCE, /* from any sender */ MPI_ANY_TAG, /* any type of message */ MPI_COMM_WORLD, /* group */ &handle ); Mateti, MPI
MPI_Wait MPI_Wait( handle, &status ); Mateti, MPI
MPI_Wait( handle, &status ); MPI_Test( handle, &status ); MPI_Wait, MPI_Test Mateti, MPI
Collective Communication Mateti, MPI
MPI_Bcast( buffer, count, MPI_Datatype, root, MPI_Comm ); All processes use the same count, data type, root, and communicator. Before the operation, the root’s buffer contains a message. After the operation, all buffers contain the message from the root MPI_Bcast Mateti, MPI
MPI_Scatter( sendbuffer, sendcount, MPI_Datatype, recvbuffer, recvcount, MPI_Datatype, root, MPI_Comm); All processes use the same send and receive counts, data types, root and communicator. Before the operation, the root’s send buffer contains a message of length sendcount * N', where N is the number of processes. After the operation, the message is divided equally and dispersed to all processes (including the root) following rank order. MPI_Scatter Mateti, MPI
MPI_Gather( sendbuffer, sendcount, MPI_Datatype, recvbuffer, recvcount, MPI_Datatype, root, MPI_Comm); This is the “reverse” of MPI_Scatter(). After the operation the root process has in its receive buffer the concatenation of the send buffers of all processes (including its own), with a total message length of recvcount * N, where N is the number of processes. The message is gathered following rank order. MPI_Gather Mateti, MPI
MPI_Reduce( sndbuf, rcvbuf, count, MPI_Datatype datatype, MPI_Op, root, MPI_Comm); After the operation, the root process has in its receive buffer the result of the pair-wise reduction of the send buffers of all processes, including its own. MPI_Reduce Mateti, MPI
MPI_MAX MPI_MIN MPI_SUM MPI_PROD MPI_LAND MPI_BAND MPI_LOR MPI_BOR MPI_LXOR MPI_BXOR MPI_MAXLOC MPI_MINLOC L logical B bit-wise Predefined Reduction Ops Mateti, MPI
User Defined Reduction Ops void myOperator ( void * invector, void * inoutvector, int * length, MPI_Datatype * datatype) { … } Mateti, MPI
Ten Reasons to Prefer MPI over PVM • MPI has more than one free, and quality implementations. • MPI can efficiently program MPP and clusters. • MPI is rigorously specified. • MPI efficiently manages message buffers. • MPI has full asynchronous communication. • MPI groups are solid, efficient, and deterministic. • MPI defines a 3rd party profiling mechanism. • MPI synchronization protects 3rd party software. • MPI is portable. • MPI is a standard. Mateti, MPI
Summary • Introduction to MPI • Reinforced Manager-Workers paradigm • Send, receive: blocked, non-blocked • Process groups Mateti, MPI
MPI resources • Open source implementations • MPICH • LAM • Books • Using MPIWilliam Gropp, Ewing Lusk, Anthony Skjellum • Using MPI-2William Gropp, Ewing Lusk, Rajeev Thakur • On-line tutorials • www.tc.cornell.edu/Edu/Tutor/MPI/ Mateti, MPI