Collective Operations

Objectives

1. Discuss Collective Operations

Notes

• Communicators
  • We have seen one communicators MPI_COMM_WORLD
  • This is probably sufficient for most simple computations.
  • But if we want to have different groups of processes working on different tasks we need communicators.
  • We can build a communicator group with the MPI_Comm_split
    • MPI_Comm_split( MPI_Comm comm, int color, int key, MPI_Comm* newcomm)
    • The existing communicator,
    • An integer we wish to split on, this is assigned per processes.
      • All processes with the same color value are in a new communicator group.
    • The key determines the rank order in the group.
    • The new communicator group.
    • So if we had 16 processes we wanted to configure by rows
    • I have never done this so I took some code from https://mpitutorial.com/tutorials/introduction-to-groups-and-communicators/
    • The code is in makeComm.cpp
  • There are other ways to build a communicator, but this will do for now.
• Collectives
  • For us to do our monte carlo pi simulation in MPI we need collectives at a minimum.
  • Four of them
    • Broadcast - send a single value to every process
    • Scatter - send an entry of an array to each process
    • Reduction - combine a single value from every process to the host with a given operator.
    • Gather - bring a single entry of an array from each process.
  • To start either a gather or reduction will do.
  • int MPI_Gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
    • reference
    • sendbuf - the data to send
    • sendcount - the number of this data to send
    • sendtype - The type of data
      • see this reference
      • Normally MPI_typename (MPI_INT)
    • recvbuf - where we will put the data
    • recvcount - how many we will receive (each, not total)
    • recvtype - the type we will receive.
    • root - where the data will be collected
    • communicator: the group involved.
  • int MPI_Allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm)
    • reference.
    • As before but no root.
    • Everyone gets the data.
    • See gather.cpp
  • Scatter
    • reference.
    • int MPI_Scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
    • See scatter.cpp
  • int MPI_Reduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)
    • reference
    • See Collectie Operations.
  • int MPI_Allreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
    • reference.
    • See reduce.cpp