The bytes Argument

The meaning of the bytes argument varies slightly depending on the situation. Here is some general information for different types of sends and receives, followed by some examples of how the byte argument works with them.

• point-to-point blocking sends - The start probes for routines that initiate point-to-point blocking sends report the number of bytes to be sent. These routines are:

  • MPI_Bsend

  • MPI_Rsend

  • MPI_Send

  • MPI_Ssend

  • MPI_Sendrecv

  • MPI_Sendrecv_replace

• point-to-point nonblocking sends - The end probes for routines that initiate point-to-point nonblocking sends report the number of bytes to be sent. These routines are:

  • MPI_Bsend_init

  • MPI_Ibsend

  • MPI_Irsend

  • MPI_Isend

  • MPI_Issend

  • MPI_Rsend_init

  • MPI_Send_init

  • MPI_Ssend_init

• point-to-point receives - The end probes for routines that terminate or could terminate point-to-point nonblocking sends report the number of bytes actually received. These routines are:

  • MPI_Iprobe

  • MPI_Probe

  • MPI_Recv

  • MPI_Test

  • MPI_Testall

  • MPI_Testany

  • MPI_Testsome

  • MPI_Wait

  • MPI_Waitall

  • MPI_Waitany

  • MPI_Waitsome

  • MPI_Sendrecv

  • MPI_Sendrecv_replace

• collectives - The start probes for collective routines report the number of bytes to be sent from an MPI process and the number to be received at the process. Such byte counts are independent of the algorithm used. For example, the number of bytes sent from the root in a broadcast is given as the number of bytes in the broadcast message, regardless of whether the root sends this message multiple times as part of a binary-tree fan-out. These collective routines are:

  • MPI_Allgather, MPI_Allgatherv

sendbytes - Number of bytes to be sent from this process.

recvbytes - Total number of bytes to be received at any process from all processes.

• MPI_Allreduce, MPI_Reduce, MPI_Reduce_scatter

bytes - Number of bytes on any process to be reduced.

• MPI_Alltoall, MPI_Alltoallv

sendbytes - Total number of bytes to be sent from this process to all processes.

recvbytes - Total number of bytes to be received at this process from all processes.

• MPI_Bcast

bytes - Number of bytes to be broadcast.

• MPI_Gather, MPI_Gatherv

sendbytes - Root reports total number of bytes to be sent; other processes report 0.

recvbytes - Root reports total number of bytes to be received; other processes report 0.

• MPI_Scan

bytes - Number of bytes contributed by any process.

• MPI_Scatter, MPI_Scatterv

sendbytes - Root reports total number of bytes to be sent; other processes report 0.

recvbytes - Number of bytes to be received at this process from the root.

• pack and unpack - The start probes for these routines report the number of bytes packed or unpaced. These routines are:

  • MPI_Pack

  • MPI_Unpack

Examples:

• MPI_Send

call MPI_Send(x,m,MPI_REAL8,...)

Probe mpi_send_start reports that 8*m bytes are to be sent.

• MPI_Recv

call MPI_Recv(x,n,MPI_REAL8,...)

Probe mpi_recv_end reports the number of bytes that were actually received, which must be at most 8*n.

• MPI_Sendrecv

call MPI_Sendrecv(x,m,MPI_REAL8,...,y,n,MPI_REAL8,...)

Probe mpi_sendrecv_start reports that 8*m bytes are to be sent, and probe mpi_sendrecv_end reports the number of bytes that were actually received, which must be at most 8*n.

• MPI_Irecv, MPI_Wait

integer req
call MPI_Irecv(x,n,MPI_REAL8,...,req,...)
call MPI_Wait(req,...) 

Probe mpi_wait_end reports the number of bytes that were actually received, which must be at most 8*n.

• MPI_Isend, MPI_Irecv, MPI_Wait

integer reqs(2)
call MPI_Isend(x,m,MPI_REAL8,...,reqs(1),...)
call MPI_Irecv(Y,N,MPI_REAL8,...,reqs(2),...)
call MPI_Waitany(2,reqs,...)
call MPI_Waitany(2,reqs,...) 

Probe mpi_isend_start reports that 8*m bytes are to be sent. The MPI_Waitany call that completes the receive will show the number of bytes that were actually received, which must be at most 8*n, in its mpi_waitany_end probe. The other MPI_Waitany call, which completes the send, will report 0 bytes received.

• MPI_Waitall

integer reqs(8)
call MPI_Isend(x1,m,MPI_REAL8,...,reqs(1),...)
call MPI_Isend(x2,m,MPI_REAL8,...,reqs(2),...)
call MPI_Isend(x3,m,MPI_REAL8,...,reqs(3),...)
call MPI_Isend(x4,m,MPI_REAL8,...,reqs(4),...)
call MPI_Irecv(x5,n,MPI_REAL8,...,reqs(5),...)
call MPI_Irecv(x6,n,MPI_REAL8,...,reqs(6),...)
call MPI_Irecv(x7,n,MPI_REAL8,...,reqs(7),...)
call MPI_Irecv(x8,n,MPI_REAL8,...,reqs(8),...)
call MPI_Waitall(8,reqs,..) 

Probe mpi_isend_start reports that 8*m bytes are to be sent in each of the four MPI_Isend cases. Probe mpi_waitall_end reports the number of bytes that were actually received, which must be at most 4*8*n.