Question

我试图找出为什么我的OpenMPI 1.6版本不起作用。我在CentOS 6.6上使用gcc-4.7.2。给定玩具程序（即hello.c）

#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>

int main(int argc, char * argv[])
{
    int taskID = -1; 
    int NTasks = -1; 

    /* MPI Initializations */
    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &taskID);
    MPI_Comm_size(MPI_COMM_WORLD, &NTasks);

    printf("Hello World from Task %i\n", taskID);

    MPI_Finalize();
    return 0;
}

并使用mpicc hello.c进行编译并运行mpirun -np 8 ./a.out，我收到错误：

--------------------------------------------------------------------------
WARNING: No preset parameters were found for the device that Open MPI
detected:

  Local host:            qmaster02.cluster
  Device name:           mlx4_0
  Device vendor ID:      0x02c9
  Device vendor part ID: 4103

Default device parameters will be used, which may result in lower
performance.  You can edit any of the files specified by the
btl_openib_device_param_files MCA parameter to set values for your
device.

NOTE: You can turn off this warning by setting the MCA parameter
      btl_openib_warn_no_device_params_found to 0.
--------------------------------------------------------------------------
Hello World from Task 4
Hello World from Task 7
Hello World from Task 5
Hello World from Task 0
Hello World from Task 2
Hello World from Task 3
Hello World from Task 6
Hello World from Task 1
[headnode.cluster:22557] 7 more processes have sent help message help-mpi-btl-openib.txt / no device params found
[headnode.cluster:22557] Set MCA parameter "orte_base_help_aggregate" to 0 to see all help / error messages

如果我使用mvapich2-2.1和gcc-4.7.2运行它，我只是得到Hello World from Task N而没有任何这些错误/警告。

查看链接到a.out的库，我得到：

$ ldd a.out 
    linux-vdso.so.1 =>  (0x00007fff05ad2000)
    libmpi.so.1 => /act/openmpi-1.6/gcc-4.7.2/lib/libmpi.so.1 (0x00002b0f8e196000)
    libdl.so.2 => /lib64/libdl.so.2 (0x0000003954800000)
    libm.so.6 => /lib64/libm.so.6 (0x0000003955400000)
    librt.so.1 => /lib64/librt.so.1 (0x0000003955c00000)
    libnsl.so.1 => /lib64/libnsl.so.1 (0x0000003965000000)
    libutil.so.1 => /lib64/libutil.so.1 (0x0000003964c00000)
    libpthread.so.0 => /lib64/libpthread.so.0 (0x0000003955000000)
    libc.so.6 => /lib64/libc.so.6 (0x0000003954c00000)
    /lib64/ld-linux-x86-64.so.2 (0x0000003954400000)

如果我用mvapich2重新编译它，

$ ldd a.out
linux-vdso.so.1 =>  (0x00007fffcdbcb000)
libmpi.so.12 => /act/mvapich2-2.1/gcc-4.7.2/lib/libmpi.so.12 (0x00002af3be445000)
libc.so.6 => /lib64/libc.so.6 (0x0000003954c00000)
libxml2.so.2 => /usr/lib64/libxml2.so.2 (0x000000395e800000)
libibmad.so.5 => /usr/lib64/libibmad.so.5 (0x0000003955400000)
librdmacm.so.1 => /usr/lib64/librdmacm.so.1 (0x0000003146400000)
libibumad.so.3 => /usr/lib64/libibumad.so.3 (0x0000003955800000)
libibverbs.so.1 => /usr/lib64/libibverbs.so.1 (0x0000003956000000)
libdl.so.2 => /lib64/libdl.so.2 (0x0000003954800000)
librt.so.1 => /lib64/librt.so.1 (0x0000003955c00000)
libgfortran.so.3 => /act/gcc-4.7.2/lib64/libgfortran.so.3 (0x00002af3beaf6000)
libm.so.6 => /lib64/libm.so.6 (0x00002af3bee0a000)
libpthread.so.0 => /lib64/libpthread.so.0 (0x0000003955000000)
libgcc_s.so.1 => /act/gcc-4.7.2/lib64/libgcc_s.so.1 (0x00002af3bf08e000)
libquadmath.so.0 => /act/gcc-4.7.2/lib64/libquadmath.so.0 (0x00002af3bf2a4000)
/lib64/ld-linux-x86-64.so.2 (0x0000003954400000)
libz.so.1 => /lib64/libz.so.1 (0x00002af3bf4d9000)
libnl.so.1 => /lib64/libnl.so.1 (0x0000003958800000)

这里有什么问题？这是因为infiniband库没有在openmpi案例中链接吗？

Answer 1

默认情况下，Open MPI 1.6不附带部件ID为4103的Mellanox ConnectX HCA的设备参数，可以轻松修复。找到[Mellanox Hermon]中的$PREFIX/share/openmpi/mca-btl-openib-device-params.ini部分，并将4103追加到部件ID列表的末尾：

[Mellanox Hermon]
vendor_id = 0x2c9,0x5ad,0x66a,0x8f1,0x1708,0x03ba,0x15b3,0x119f
vendor_part_id = 25408,25418,25428,...<skipped>...,26488,4099,4103
use_eager_rdma = 1                                           ^^^^^
mtu = 2048
max_inline_data = 128

将$PREFIX替换为Open MPI安装的路径。在您的情况下，/act/openmpi-1.6/gcc-4.7.2。

玩具程序使用OpenMPI 1.6失败但与Mvapich2一起使用

1 个答案: