Title: COMP309/509 - Lecture 22 class: middle, center, inverse

COMP309/509 - Parallel and Distributed Computing

Lecture 22 - Advanced MPI

By Mitchell Welch

University of New England

Reading

• A users guide to MPI (This respurce is a bit dated, however the examples are still current)
• man mpirun
• Assignment 4
• Project (COMP509)

Summary

• Advanced MPI Commumication
• Matrix Multiplication in MPI

Advanced MPI Commumication

• We have seen two forms of collective communication in MPI:
  • MPI_Bcast
  • MPI_Reduce
• MPI_Bcast acts:
  • like a send for the process whose rank is root.
  • like a receive for all other processes.

Advanced MPI Commumication

• MPI_Reduce acts:
  • like a receive for the process whose rank is root.
  • like a send for all other processes.
  • the thing received is the accumulation of all the sends.
  • it is accumulated using the binary operation MPI_Op.
• Built-in MPI_Ops:
  • MPI_MAX MPI_MIN MPI_SUM MPI_PROD MPI_LAND MPI_LOR

Advanced MPI Commumication

• Thus these calls are a form of:
  • communication.
  • synchronization!.
• Similarly MPI_Barrier can be thought of in the same light.
• Another example is MPI_Gather

Advanced MPI Commumication

• MPI_Gather acts:
  • like a receive for the process whose rank is root.
  • like a send for all processes, including root.
  • the thing received is the accumulation of all the sends.
  • it is accumulated via concatenation in the receive buffer, ordered by rank.
• Thus MPI_Gather would be perfect for collecting the rows of the C matrix in the algorithm used in pvmmm!

Advanced MPI Commumication

#include "mpi.h"
int MPI_Gather (void *sndbuf,  int sndcnt, MPI_Datatype sndtyp, 
                void *rcvbuf,  int rcvcnt, MPI_Datatype rcvtyp, 
                int root, MPI_Comm comm )

• Gathers together values from a group of processes.
• sndbuf is the address of send buffer.
• sndcnt is the number of elements in send buffer.

Advanced MPI Commumication

• sndtyp is the type of send buffer items.
• rcvbuf is the address of receive buffer.
• significant only at root.
• rcvcnt is the number of elements for any single receive
• significant only at root.
• root is the rank of receiving process.
• comm is the communicator of the processes participating in the gathering.

Advanced MPI Commumication

• Each process in the comm communicator: sends the contents of sndbuf to the process with rank root.
• This includes the process with rank root.
• The process root concatenates the received data in process rank order in rcvbuff.
• That means:
  • The data from process 0 comes first,
  • Followed by the data from process 1,
  • and so on
  • all the way to the last process in the communicator comm.

Advanced MPI Commumication

• The rcv arguments are only significant at the root process.
• rcvcnt is the number of elements received from each particular process, not the total.
• Thus rcvbuf should be capable of storing rcvcnt times the number of processes in the communicator.

Advanced MPI Commumication

#include "mpi.h"
int MPI_Scatter(void *sndbuf, int sndcnt, MPI_Datatype sndtyp, 
                void *rcvbuf, int rcvcnt, MPI_Datatype rcvtyp, 
                int root, MPI_Comm comm )

• Scatters the values to a group of processes.

Advanced MPI Commumication

• sndbuf is the address of send buffer.
  • significant only at root.
• sndcnt is the number of elements in send buffer.
  • significant only at root.
• sndtype is the type of send buffer items.
• rcvbuf is the address of receive buffer.
• rcvcnt is the number of elements for any single receive
• root is the rank of scattering process.
• comm is the communicator of the processes participating in the scattering.

Advanced MPI Commumication

• Another procedure that would great for porting!
• MPI_Scatter is the inverse operation to a gathering.
• The process with rank root distributes the contents of sndbuf among the processes in the communicator comm.
• The contents of the sndbuf are split into nproc segments, each consisting of sndcnt items.

Advanced MPI Commumication

• Scattering means:
  • process 0 gets the first segment.
  • process 1 gets the second segment.
  • and so on
  • all the way to the last process in the communicator comm which gets the nprocth segment.
• The snd arguments are only significant at the root process.
• The rcv arguments are significant at all processes.

Advanced MPI Commumication

• Lets use them to do matrix multiplication ring style.
• There are two versions of the program in the Examples subdirectory.
  • mpimm.c contains oodles of error checking.
  • abbreviated-mpimm.c contains no error checking.
• You should know by now that for good marks your solutions should look like the former not the latter!!

Matrix Multiplication in MPI

• Review the Examples
  • mpimm.c
  • abbreviated-mpimm.c

class: middle, center, inverse

Questions?

Reading

• A users guide to MPI (This respurce is a bit dated, however the examples are still current)
• man mpirun
• Assignment 4
• Project (COMP509)