When a tool or debugger (including PDB) with large-scale jobs is used, it is recommended
that you complete the following tasks:
Ensure that the MP_DBG_TASKINFO environment variable is set to yes. This setting
indicates that the debugger exchanges task information with POE by way of a socket
connection. MP_DBG_TASKINFO is set to yes by default.
Ensure that the MP_DEBUG_ATTACH environment variable is set to no. This setting
indicates that debugger attachment files are not created or updated, which degrades the
performance of a large-scale job.
Create a socket connection that requests the specific task information that is needed by
the tool or debugger, by using the poe.socket.taskinfo API. The socket request specifies
the following types of task information:
– Task ID
– World ID
– Rank in the world
– PID of the task
– Host name
– IP address
– Executable path name
– Node ID (for PDB use only)
– Session ID (for PDB use only)
For more information about creating the socket request, see to the main page for
poe.socket.taskinfo in Parallel Environment Runtime Edition for AIX V1.1: MPI Programming
Guide, SC23-6783.
When MP_DBG_TASKINFO=yes, POE and PMD each create UNIX socket files whose host
names are the same as the names returned by gethostname() system call. By default, these
files are written to /tmp, but the system administrator optionally uses the
MP_DBG_TASKINFO_DIR entry in the /etc/poe.limits file to change the directory into
which these files are stored.
The tool or debugger connects to the sockets created by POE and the PMD and sends the
task information request. POE and PMD respond to the query by sending the requested task
information to the tool by way of the same socket.
Specifying hosts for jobs with large numbers of tasks
For jobs that include large numbers of tasks (up to 1024 K), a simple host list file in which
each host is specified on a separate line is unworkable. Instead, PE provides a shorthand for
specifying the hosts for such large jobs.
Specifying hosts for jobs that include large numbers of tasks requires a different method than
the method used for smaller jobs. For smaller jobs, each host is included in the POE host list
file on a separate line, but the use of this configuration for large jobs is impractical. To make
the host list file usable for large jobs, PE provides a shorthand for specifying the hosts.
On each line of the host list file, you specify a host name, followed by the tasks that run on it.
If you do not specify a value for tasks, a default task mapping is assumed. You provide
multiple hosts and multiple tasks on a single line. If both hosts and tasks are specified, the
number of hosts must match the number of tasks.
The following format of the basic host list file entry is used:
hosts%tasks
Chapter 2. Application integration
133
Need help?
Do you have a question about the Power Systems 775 and is the answer not in the manual?