中科院计算所谭光明老师课件C02-高级MPI编程技术.ppt

MPI_Alltoallv函数 int MPI_Alltoallv( void* sendbuf /* in */, int sendcounts[] /* in */, int sdispls[] /* in */, MPI_Datatype sendtype /* in */, void* recvbuf /*out*/, int recvcounts[] /* in */, int rdispls[] /* in */, MPI_Datatype recvtype /* in */, MPI_Comm comm /* in */) MPI_Alltoallv函数详解 正如MPI_Allgatherv 和MPI_Allgather 的关系一样MPI_Alltoallv在MPI_Alltoall的基础上进一步增加了灵活性。它可以由sdispls指定待发送数据的位置，在接收方则由rdispls指定接收的数据存放在缓冲区的偏移量 所有参数对每个进程都是有意义的，并且所有进程中的comm值必须一致 MPI_Alltoall和MPI_Alltoallv可以实现n次独立的点对点通信但也有限制：1)所有数据必须是同一类型；2)所有的消息必须按顺序进行散发和收集 MPI_Alltoall vs MPI_Alltoallv 全局归约MPI_Reduce int MPI_Reduce ( void* openand /* in */, void* result /* in */, int count /* in */, MPI_Datatype datatype /* in */, MPI_Op operator /*out*/, int root /* in */, MPI_Comm comm /* in */) openand 操作数(发送缓冲区)起始地址 result 接收缓冲区(结果)的地址 count 发送缓冲区数据个数 operator 归约操作符 组归约函数MPI_Allreduce int MPI_Reduce ( void* openand /* in */, void* result /* in */, int count /* in */, MPI_Datatype datatype /* in */, MPI_Op operator /*out*/, MPI_Comm comm /* in */) MPI_Allreduce比MPI_Reduce少一个root 参数, 其它参数及含义与后者一样 MPI_Allreduce函数详解 MPI_Allreduce相当于组中每一个进程都作为root分别进行了一次归约操作。 MPI_Allreduce 相当于在MPI_Reduce 后马上再将结果进行一次广播MPI_Bcast 归约的结果不只是某一个进程所有而是所有的进程都所有。 它在某种程度上和组收集与收集的关系很相似 MPI_Allreduce函数图示 Sendbuf Recvbuf MPI_Allreduce: np = 3; count = 3; Op = MPI_SUM P0 P1 P2 A0 A1 A2 B0 B1 B2 C0 C1 C2 A0+B0+C0 MPI_Allreduce A1+B1+C1 A2+B2+C2 P0 P1 P2 A0+B0+C0 A1+B1+C1 A2+B2+C2 A0+B0+C0 A1+B1+C1 A2+B2+C2 归约发散函数MPI_Reduce_scatter int MPI_Reduce_scatter ( void* sendbuf /* in */, void* recvbuf /* in */, int* recvcounts /* in */, MPI_Datatype datatype /* in */, MPI_Op operator /*out*/, MPI_Comm comm /* in */) MPI_Reduce_scatter函数图示 MPI_Reduce_scatter: np = 3; count = 3; Op = MPI_SUM Sendbuf Recvbuf P0 P1 P2 A0 A1 A2 B0 B1 B2 C0 C1 C2 A0+B0+C0 P0 P1 P2 A1+B1+C1 A2+B2+C2 MPI_Reduce_scat