Question

我在C中使用MPI（openMPI 3.0.1）来完成我的工作，我想知道如何设置mpi_comm_spawn，就好像我用mpirun运行我的可执行文件一样？

我使用这个主机文件：

# This is a single processor node: 

# This is a height-processor node. Oversubscribing
# to it is prevented by setting max-slots=4:

# host  | rank | nb procs | max nb procs
  node7   0,2    slots=2    max-slots=4

# This is a height-processor node. Oversubscribing
# to it is prevented by setting max-slots=8:

# host  | rank        | nb procs | max nb procs
  node8   1,3,4,5,6,7   slots=6    max-slots=8

所以当我用

启动hostname时

mpirun -hostfile /path/to/mpi_hostfile hostname

我给了

node7
node7
node8
node8
node8
node8
node8
node8

但当我生成新进程时（警告这是伪代码）

maxprocs = to_unsigned(arg)
info_key = "add-hostfile"
info_value = "/path/to/mpi_hostfile"

mpi_comm_spawn(....)

这是我使用的代码

mpi->comm_spawn->set_attr_command(argv[0]);
mpi->comm_spawn->set_attr_argv(argv + 1);
mpi->comm_spawn->set_attr_maxprocs(universe_size - 1);

if (mc_options.hostfile == nullptr)
  mpi->comm_spawn->set_attr_info(spot::mpi::info_null);
else
  {
     // OpenMPI specific
     const char* hostfile = "add-hostfile";

     mpi->info_create->set_attr_info(&info);
     mpi->info_create->do_info_create();

     mpi->info_set->set_attr_info(info);
     mpi->info_set->set_attr_key(hostfile);
     mpi->info_set->set_attr_value(mc_options.hostfile);
     mpi->info_set->do_info_set();

     mpi->comm_spawn->set_attr_info(info);
   }

 mpi->comm_spawn->set_attr_root(master);
 mpi->comm_spawn->set_attr_comm(spot::mpi::comm_self);
 mpi->comm_spawn->set_attr_intercomm(&children);
 mpi->comm_spawn->
 set_attr_array_of_errcodes(spot::mpi::errcodes_ignore);
 mpi->comm_spawn->do_comm_spawn();

我在node7上拥有所有进程，并且我无法启动2个以上的进程（比如我在node7的mpi_hostfile中指定）

Process #0: on HOST node7
Process #1: on HOST node7

如果我尝试推出更多

--------------------------------------------------------------------------
There are not enough slots available in the system to satisfy the 3 slots
that were requested by the application:
/* my executable*/

Either request fewer slots for your application, or make more slots available
for use.
--------------------------------------------------------------------------
[node7:2306] *** An error occurred in MPI_Comm_spawn
[node7:2306] *** reported by process [4042260481,0]
[node7:2306] *** on communicator MPI_COMM_SELF
[node7:2306] *** MPI_ERR_SPAWN: could not spawn processes
[node7:2306] *** MPI_ERRORS_ARE_FATAL (processes in this communicator will now abort,
[node7:2306] ***    and potentially your MPI job)

我希望

node7
node7
node8
node8
node8
node8
node8
node8

和mpirun一样我不明白为什么会有所不同？

感谢。

如何将MPI_COMM_SPAWN设置为像mpirun一样？

0 个答案: