IBM MP Utility Descriptions

System and Other Tuning Information

MPI is provided as part of the IBM AIX offering package "Parallel Operating Environment" (POE).

• poe: MPI driver for AIX.

  Usage:  submit = poe $command ...
  
    Command-line parameters:
  
      -procs $ranks            The number of ranks to run. Can also be specified with the MP_PROCS environment variable.
  	-hostfile $hostfile      Name of file containing the list of names of hosts that will run each rank. Can also
  	                         be specified with the MP_HOSTFILE environment variable.
  
    Parameters specified by environment variables:
  
       MP_PGMMODEL=spmd              Set program model to Single-Program/Multiple-Data.
  	 MP_MSG_API=mpi                Specify that MPI is the message-passing protocol to use.
  	 MP_DEVTYPE=ib                 Use InfiniBand devices for communication.
  	 MP_CLOCK_SOURCE=AIX           Retreive time values from the OS.
       MP_ADAPTER_USE=shared         Determines how the node's adapter should be used (e.g. "shared" for shared between ranks).
       MP_EUILIB=us                  Determines the communication subsystem implementation to use (e.g. "us" for user mode).
       MP_EUIDEVICE=sn_single        Determines the adapter set to use for message passing (e.g. "sn_single" for one window per task).
       MP_EUIDEVICE=sn_all            Using "sn_all" causes user-space windows to be striped per task.
       MP_SHARED_MEMORY=yes          Do not use switch to transmit messages within the same node.
       MP_SINGLE_THREAD=yes          Skip mutex locking in single-threaded ranks.
       MP_WAIT_MODE=poll             Poll across blocked threads.
       MP_WAIT_MODE=yield            Suspend blocked threads.
       MP_EAGER_LIMIT=65536          Use rendezvous protocol above this message size.
       MP_BUFFER_MEM=67108864        The size of the Early Arrival (EA) buffer that is used by the communication subsystem.
       MP_POLLING_INTERVAL=80000000  Length of polling interval, in microseconds.
  	 MP_PULSE=0                    The interval at which POE checks the remote nodes to ensure that they are communicating with the home node. Use 0 to disable.
       MP_USE_BULK_XFER=yes          Use the bulk data transfer mechanism.
       MP_BULK_MIN_MSG_SIZE=65536    Use bulk data transfer for messages of at least this many bytes.
       MP_STDINMODE=none             Tasks will not read input from STDIN.
  	 MP_STDOUTMODE=unordered       Determines the formatting of task output.
       MP_LABELIO=no                 Do not label lines of output with the task ID number.
  	 MP_RESD=no                    Determines whether or not the Partition Manager should connect to the LoadLeveler utility.
  	 MP_CSS_INTERRUPT=no           Determines whether or not arriving message packets cause interrupts. 
  	 MP_INFOLEVEL=0                Determines the level of message reporting.
  	 MP_PMDLOG=no                  Determines whether or not diagnostic messages should be logged to a file in /tmp on each of the remote nodes.
  	 MP_PARTITION=64               An identifying number that is the same for all tasks in a job.
  	 NRT_WINDOW_COUNT=1            Set to the number of communication windows used by each task. 
  

• petaskbind.sh: script to assign ranks to cores within a node, with the core index incremented by the stride parameter.

  Usage:  submit = poe petaskbind.sh $command ...
  

---