MPI_GATHERV(3)

📦
version 5.0.11 · 17 versions
, in arch rolling

💖 MPI_GATHERV(3) — Open MPI manual, but make it make sense 💖

okay bestie, buckle up 🎀 this is the Open MPI page for MPI_Gatherv, MPI_Igatherv, and MPI_Gatherv_init, and it's genuinely one of the most useful collectives out there. the vibe? everybody sends a different amount of stuff to the root, and the root catches all of it. no cap.

🌈 NAME

MPI_Gatherv, MPI_Igatherv, MPI_Gatherv_init — the family that gathers varying amounts of data from all processes to the root process. yeah. varying. that's the whole point of the v. 💅

✨ SYNTAX ✨

🧚 C Syntax

#include <mpi.h>

int MPI_Gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
    void *recvbuf, const int recvcounts[], const int displs[], MPI_Datatype recvtype,
    int root, MPI_Comm comm)

int MPI_Igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
    void *recvbuf, const int recvcounts[], const int displs[], MPI_Datatype recvtype,
    int root, MPI_Comm comm, MPI_Request *request)

int MPI_Gatherv_init(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
    void *recvbuf, const int recvcounts[], const int displs[], MPI_Datatype recvtype,
    int root, MPI_Comm comm, MPI_Info info, MPI_Request *request)

three flavours, bestie: the blocking one, the I one (that's the non-blocking icon ✨), and the _init one (persistent request, main character energy).

🦋 Fortran Syntax

USE MPI
! or the older form: INCLUDE 'mpif.h'

MPI_GATHERV(SENDBUF, SENDCOUNT, SENDTYPE, RECVBUF, RECVCOUNTS,
        DISPLS, RECVTYPE, ROOT, COMM, IERROR)
    <type>  SENDBUF(*), RECVBUF(*)
    INTEGER SENDCOUNT, SENDTYPE, RECVCOUNTS(*), DISPLS(*)
    INTEGER RECVTYPE, ROOT, COMM, IERROR

MPI_IGATHERV(SENDBUF, SENDCOUNT, SENDTYPE, RECVBUF, RECVCOUNTS,
        DISPLS, RECVTYPE, ROOT, COMM, REQUEST, IERROR)
    <type>  SENDBUF(*), RECVBUF(*)
    INTEGER SENDCOUNT, SENDTYPE, RECVCOUNTS(*), DISPLS(*)
    INTEGER RECVTYPE, ROOT, COMM, REQUEST, IERROR

MPI_GATHERV_INIT(SENDBUF, SENDCOUNT, SENDTYPE, RECVBUF, RECVCOUNTS,
        DISPLS, RECVTYPE, ROOT, COMM, INFO, REQUEST, IERROR)
    <type>  SENDBUF(*), RECVBUF(*)
    INTEGER SENDCOUNT, SENDTYPE, RECVCOUNTS(*), DISPLS(*)
    INTEGER RECVTYPE, ROOT, COMM, INFO, REQUEST, IERROR

note the USE MPI up top, or the old-school INCLUDE 'mpif.h' if you're feeling vintage 🫶

🌸 Fortran 2008 Syntax

USE mpi_f08

MPI_Gatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs,
        recvtype, root, comm, ierror)
    TYPE(*), DIMENSION(..), INTENT(IN) :: sendbuf
    TYPE(*), DIMENSION(..) :: recvbuf
    INTEGER, INTENT(IN) :: sendcount, recvcounts(*), displs(*), root
    TYPE(MPI_Datatype), INTENT(IN) :: sendtype, recvtype
    TYPE(MPI_Comm), INTENT(IN) :: comm
    INTEGER, OPTIONAL, INTENT(OUT) :: ierror

MPI_Igatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs,
        recvtype, root, comm, request, ierror)
    TYPE(*), DIMENSION(..), INTENT(IN), ASYNCHRONOUS :: sendbuf
    TYPE(*), DIMENSION(..), ASYNCHRONOUS :: recvbuf
    INTEGER, INTENT(IN) :: sendcount, root
    INTEGER, INTENT(IN), ASYNCHRONOUS :: recvcounts(*), displs(*)
    TYPE(MPI_Datatype), INTENT(IN) :: sendtype, recvtype
    TYPE(MPI_Comm), INTENT(IN) :: comm
    TYPE(MPI_Request), INTENT(OUT) :: request
    INTEGER, OPTIONAL, INTENT(OUT) :: ierror

MPI_Gatherv_init(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs,
        recvtype, root, comm, info, request, ierror)
    TYPE(*), DIMENSION(..), INTENT(IN), ASYNCHRONOUS :: sendbuf
    TYPE(*), DIMENSION(..), ASYNCHRONOUS :: recvbuf
    INTEGER, INTENT(IN) :: sendcount, root
    INTEGER, INTENT(IN), ASYNCHRONOUS :: recvcounts(*), displs(*)
    TYPE(MPI_Datatype), INTENT(IN) :: sendtype, recvtype
    TYPE(MPI_Comm), INTENT(IN) :: comm
    TYPE(MPI_Info), INTENT(IN) :: info
    TYPE(MPI_Request), INTENT(OUT) :: request
    INTEGER, OPTIONAL, INTENT(OUT) :: ierror

see how the async ones slap ASYNCHRONOUS on sendbuf, recvbuf, recvcounts, and displs? that's not decoration, that's the whole contract 😤 keep it exactly as written.

👑 INPUT PARAMETERS

  • sendbuf — starting address of send buffer (choice). where your stuff lives before it gets yeeted to the root 🚀
  • sendcount — number of elements in send buffer (integer). one number, yours, nobody else's business.
  • sendtype — datatype of send buffer elements (handle).
  • recvcounts — integer array (of length group size) containing the number of elements that are received from each process (significant only at root). significant only at root, so if you're not root, don't stress about it 💅
  • displs — integer array (of length group size). entry i specifies the displacement relative to recvbuf at which to place the incoming data from process i (significant only at root). this is literally the map that says "process i's stuff goes here" 🗺️
  • recvtype — datatype of recv buffer elements (significant only at root) (handle).
  • root — rank of receiving process (integer). the one collecting everything. the main character of this operation fr fr.
  • comm — communicator (handle).
  • info — info (handle, persistent only). yeah, only the _init one takes this 🎀

🍓 OUTPUT PARAMETERS

  • recvbuf — address of receive buffer (choice, significant only at root). this is where the pile lands 📦
  • request — request (handle, non-blocking only). your little ticket to come back and check on it later.
  • ierror — Fortran only: error status (integer).

✨ DESCRIPTION ✨

okay so MPI_Gatherv is literally MPI_Gather's glow-up 💅 — it does everything Gather does but it lets every process contribute a different amount of data, because recvcounts is now a whole array instead of one sad little number. And it's more flexible about where the data lands on the root, thanks to the brand-new argument displs. she gave us options, she gave us range 👑

The outcome is as if each process, including the root process, sends a message to the root:

   MPI_Send(sendbuf, sendcount, sendtype, root, ...);

and the root executes n receives:

   MPI_Recv(recvbuf + disp[i] * extent(recvtype), recvcounts[i],
            recvtype, i, ...);

Messages get placed in the receive buffer of the root process in rank order — so the data sent from process j goes into the jth portion of the receive buffer recvbuf on process root. And that jth portion of recvbuf begins at offset displs[j] elements (in terms of recvtype) into recvbuf. yes bestie, the displacements are in elements of recvtype, not bytes — do not come for me, that's just how it is 🌸

The receive buffer is ignored for all nonroot processes. like, they literally don't look at it, it's not their business 🙈

The type signature implied by sendcount, sendtype on process i must be equal to the type signature implied by recvcounts[i], recvtype at the root. Which means the amount of data sent must equal the amount of data received, pairwise between each process and the root. Distinct type maps between sender and receiver are still allowed though — Example 2 below is literally the icon that proves it 💖

All arguments to the function are significant on process root, while on other processes, only sendbuf, sendcount, sendtype, root, comm are significant. And root and comm must have identical values on all processes — no cap, everyone has to agree or it's chaos 😭

⚠️ The specification of counts, types, and displacements should not cause any location on the root to be written more than once. Such a call is erroneous. Double-writing is not a vibe. Don't do it. periodt.

Example 1: Now have each process send 100 ints to root, but place each set (of 100) stride ints apart at the receiving end. Use MPI_Gatherv and the displs argument to achieve this effect. Assume stride >= 100.

   MPI_Comm comm;
   int gsize, sendarray[100];
   int root, *rbuf, stride;
   int *displs, i, rcounts;
   ...

   MPI_Comm_size(comm, &gsize);
   rbuf = (int)malloc(gsize * stride * sizeof(int));
   displs = (int)malloc(gsize * sizeof(int));
   rcounts = (int )malloc(gsize * sizeof(int));

   for (i=0; i<gsize; ++i) {
     displs[i] = i * stride;
     rcounts[i] = 100;
   }
   MPI_Gatherv(sendarray, 100, MPI_INT, rbuf, rcounts, displs, MPI_INT,
               root, comm);

Note that the program is erroneous if stride < 100. 🫠

Example 2: Same as Example 1 on the receiving side, but send the 100 ints from the 0th column of a 100 x 150 int array, in C.

   MPI_Comm comm;
   int gsize, sendarray[100][150];
   int root, *rbuf, stride;
   MPI_Datatype stype;
   int displs,i, rcounts;
   ...

   MPI_Comm_size(comm, &gsize);
   rbuf = (int )malloc(gsize * stride * sizeof(int));
   displs = (int)malloc(gsize * sizeof(int));
   rcounts = (int )malloc(gsize * sizeof(int));

   for (i=0; i<gsize; ++i) {
     displs[i] = i * stride;
     rcounts[i] = 100;
   }

   // Create datatype for 1 column of array
   MPI_Type_vector(100, 1, 150, MPI_INT, &stype);
   MPI_Type_commit( &stype );
   MPI_Gatherv(sendarray, 1, stype, rbuf, rcounts, displs, MPI_INT,
               root, comm);

see how the sending side sends 1 of stype and the root still receives 100 MPI_INTs? type maps differ, type signatures match. that's the whole slay of this example 💅

Example 3: Process i sends (100-i) ints from the ith column of a 100 x 150 int array, in C. It is received into a buffer with stride, as in the previous two examples.

   MPI_Comm comm;
   int gsize, sendarray[100][150], *sptr;
   int root, *rbuf, stride, myrank;
   MPI_Datatype stype;
   int displs, i, rcounts;
   ...

   MPI_Comm_size(comm, &gsize);
   MPI_Comm_rank( comm, &myrank );
   rbuf = (int)malloc(gsize * stride * sizeof(int));
   displs = (int)malloc(gsize * sizeof(int));
   rcounts = (int )malloc(gsize * sizeof(int));

   for (i=0; i<gsize; ++i) {
     displs[i] = i * stride;
     rcounts[i] = 100-i; // note change from previous example
   }

   // Create datatype for the column we are sending
   MPI_Type_vector(100-myrank, 1, 150, MPI_INT, &stype);
   MPI_Type_commit( &stype );
   // sptr is the address of start of "myrank" column
   sptr = &sendarray[0][myrank];
   MPI_Gatherv(sptr, 1, stype, rbuf, rcounts, displs, MPI_INT,
               root, comm);

Note that a different amount of data is received from each process. that's the entire point bestie, varying counts, main character energy ✨

Example 4: Same as Example 3, but done in a different way at the sending end. We create a datatype that causes the correct striding at the sending end so that we read a column of a C array.

   MPI_Comm comm;
   int gsize, sendarray[100][150], *sptr;
   int root, *rbuf, stride, myrank, disp[2], blocklen[2];
   MPI_Datatype stype, type[2];
   int displs, i, rcounts;
   ...

   MPI_Comm_size(comm, &gsize);
   MPI_Comm_rank(comm, &myrank );
   rbuf = (int )alloc(gsize * stride * sizeof(int));
   displs = (int )malloc(gsize * sizeof(int));
   rcounts = (int)malloc(gsize * sizeof(int));

   for (i=0; i<gsize; ++i) {
     displs[i] = i* stride;
     rcounts[i] = 100-i;
   }
   // Create datatype for one int, with extent of entire row
   disp[0] = 0;
   disp[1] = 150 * sizeof(int);
   type[0] = MPI_INT;
   type[1] = MPI_UB;
   blocklen[0] = 1;
   blocklen[1] = 1;

   MPI_Type_struct( 2, blocklen, disp, type, &stype );
   MPI_Type_commit(&stype );
   sptr = &sendarray[0][myrank];
   MPI_Gatherv(sptr, 100-myrank, stype, rbuf, rcounts, displs, MPI_INT,
               root, comm);

Example 5: Same as Example 3 at sending side, but at receiving side we make the stride between received blocks vary from block to block.

   MPI_Comm comm;
   int gsize, sendarray[100][150], *sptr;
   int root, *rbuf, *stride, myrank, bufsize;
   MPI_Datatype stype;
   int *displs, i, *rcounts, offset;
   ...

   MPI_Comm_size( comm, &gsize);
   MPI_Comm_rank( comm, &myrank );
   de = (int )malloc(gsize * sizeof(int));
   ...
   // stride[i] for i = 0 to gsize-1 is set somehow

   // set up displs and rcounts vectors first
   displs = (int)malloc(gsize * sizeof(int));
   rcounts = (int )malloc(gsize * sizeof(int));
   offset = 0;

   for (i=0; i<gsize; ++i) {
     displs[i] = offset;
     offset += stride[i];
     rcounts[i] = 100-i;
   }

   // the required buffer size for rbuf is now easily obtained
   bufsize = displs[gsize-1]+rcounts[gsize-1];
   rbuf = (int )malloc(bufsize * sizeof(int));
   // Create datatype for the column we are sending
   MPI_Type_vector(100-myrank, 1, 150, MPI_INT, &stype);
   MPI_Type_commit( &stype );
   sptr = &sendarray[0][myrank];
   MPI_Gatherv(sptr, 1, stype, rbuf, rcounts, displs, MPI_INT,
               root, comm);

look at her computing bufsize from displs[gsize-1]+rcounts[gsize-1] — iconic, efficient, no wasted malloc 🔥

Example 6: Process i sends num ints from the ith column of a 100 x 150 int array, in C. The complicating factor is that the various values of num are not known to root, so a separate gather must first be run to find these out. The data is placed contiguously at the receiving end.

   MPI_Comm comm;
   int gsize, sendarray[100][150], *sptr;
   int root, *rbuf, stride, myrank, disp[2], blocklen[2];
   MPI_Datatype stype,types[2];
   int *displs, i, *rcounts, num;
   ...

   MPI_Comm_size( comm, &gsize);
   MPI_Comm_rank( comm, &myrank );

   // First, gather nums to root
   rcounts = (int )malloc(gsize * sizeof(int));
   MPI_Gather( &num, 1, MPI_INT, rcounts, 1, MPI_INT, root, comm);
   // root now has correct rcounts, using these we set
   // displs[] so that data is placed contiguously (or concatenated) at receive end

   displs = (int)malloc(gsize * sizeof(int));
   displs[0] = 0;
   for (i=1; i<gsize; ++i) {
     displs[i] = displs[i-1]+rcounts[i-1];
   }

   // And, create receive buffer
   rbuf = (int *)malloc(gsize * (displs[gsize-1]+rcounts[gsize-1]) * sizeof(int));
   // Create datatype for one int, with extent of entire row
   disp[0] = 0;
   disp[1] = 150 * sizeof(int);
   type[0] = MPI_INT;
   type[1] = MPI_UB;
   blocklen[0] = 1;
   blocklen[1] = 1;
   MPI_Type_struct(2, blocklen, disp, type, &stype );
   MPI_Type_commit( &stype );
   sptr = &sendarray[0][myrank];
   MPI_Gatherv(sptr, num, stype, rbuf, rcounts, displs, MPI_INT, root, comm);

two-step queen behavior: gather the counts first, then gather the data. lowkey the most real-world example of the bunch 🫶

💅 USE OF IN-PLACE OPTION 💅

The in-place option operates in the same way as it does for MPI_Gather. When the communicator is an intracommunicator, you can perform a gather operation in-place (the output buffer is used as the input buffer). Use the variable MPI_IN_PLACE as the value of the root process sendbuf. In this case, sendcount and sendtype are ignored, and the contribution of the root process to the gathered vector is assumed to already be in the correct place in the receive buffer. she already put herself there, no notes 💅

Note that MPI_IN_PLACE is a special kind of value; it has the same restrictions on its use as MPI_BOTTOM. so like, treat it with the same respect, okay? 🙏

Because the in-place option converts the receive buffer into a send-and-receive buffer, a Fortran binding that includes INTENT must mark these as INOUT, not OUT. this one matters, do not skip it, fr fr 🧚

🌈 WHEN COMMUNICATOR IS AN INTER-COMMUNICATOR 🌈

When the communicator is an inter-communicator, the root process in the first group gathers data from all the processes in the second group. The first group defines the root process. That process uses MPI_ROOT as the value of its root argument. The remaining processes use MPI_PROC_NULL as the value of their root argument. All processes in the second group use the rank of that root process in the first group as the value of their root argument. The send buffer argument of the processes in the first group must be consistent with the receive buffer argument of the root process in the second group. 😭 yes it's a whole cross-group situation, but everybody has a role and everybody plays it 👑

💥 ERRORS 💥

okay bestie, error-handling arc, let's get into it 🫶

Almost all MPI routines return an error value; C routines as the return result of the function and Fortran routines in the last argument.

Before the error value is returned, the current MPI error handler associated with the communication object (e.g., communicator, window, file) is called. If no communication object is associated with the MPI call, then the call is considered attached to MPI_COMM_SELF and will call the associated MPI error handler. When MPI_COMM_SELF is not initialized (i.e., before MPI_Init/MPI_Init_thread, after MPI_Finalize, or when using the Sessions Model exclusively) the error raises the initial error handler. The initial error handler can be changed by calling MPI_Comm_set_errhandler on MPI_COMM_SELF when using the World model, or the mpi_initial_errhandler CLI argument to mpiexec or info key to MPI_Comm_spawn/MPI_Comm_spawn_multiple. If no other appropriate error handler has been set, then the MPI_ERRORS_RETURN error handler is called for MPI I/O functions and the MPI_ERRORS_ABORT error handler is called for all other MPI functions.

Open MPI includes three predefined error handlers that can be used:

  • MPI_ERRORS_ARE_FATAL — causes the program to abort all connected MPI processes. 💀
  • MPI_ERRORS_ABORT — an error handler that can be invoked on a communicator, window, file, or session. When called on a communicator, it acts as if MPI_Abort was called on that communicator. If called on a window or file, acts as if MPI_Abort was called on a communicator containing the group of processes in the corresponding window or file. If called on a session, aborts only the local process.
  • MPI_ERRORS_RETURN — returns an error code to the application. 🎀

MPI applications can also implement their own error handlers by calling:

  • MPI_Comm_create_errhandler then MPI_Comm_set_errhandler
  • MPI_File_create_errhandler then MPI_File_set_errhandler
  • MPI_Session_create_errhandler then MPI_Session_set_errhandler or at MPI_Session_init
  • MPI_Win_create_errhandler then MPI_Win_set_errhandler
Note that MPI does not guarantee that an MPI program can continue past an error. like, no cap, that's important 😭

See the MPI man page for a full list of MPI error codes.

See the Error Handling section of the MPI-3.1 standard for more information.

🔎 SEE ALSO 🔎

  • MPI_Gather — yes bestie, go peek at that one too 👀✨

2003-2026, The Open MPI Community

okay that's the whole page, mwah 💋 this was MPI_GATHERV(3) from the Sep 11, 2026 footer drop, and we are soooo done — bye bestie 🫶🌈✨

📦 where this page comes from