IBM MPI Parallel Environment (PE) Applications

Totalview® for HPC User Guide : PART IV Advanced Tools and Customization : Chapter 19 Setting Up MPI Debugging Sessions : IBM MPI Parallel Environment (PE) Applications

NOTE >> In many cases, you can bypass the procedure described in this section. For more information, see "Debugging MPI Programs".

You can debug IBM MPI Parallel Environment (PE) applications on the IBM RS/6000 and SP platforms.

To take advantage of TotalView’s ability to automatically acquire processes, you must be using release 3,1 or later of the Parallel Environment for AIX.

Topics in this section are:

• "Preparing to Debug a PE Application"

• "Starting TotalView on a PE Program"

• "Setting Breakpoints"

• "Starting Parallel Tasks"

• "Attaching to a PE Job"

Preparing to Debug a PE Application

The following sections describe what you must do before TotalView can debug a PE application.

Using Switch-Based Communications

If you’re using switch-based communications (either IP over the switch or user space) on an SP computer, configure your PE debugging session so that TotalView can use IP over the switch for communicating with the TotalView Server (tvdsvr). Do this by setting the -adapter_use option to shared and the -cpu_use option to multiple, as follows:

• If you’re using a PE host file, add shared multiple after all host names or pool IDs in the host file.

• Always use the following arguments on the poe command line:

-adapter_use shared -cpu_use multiple

If you don’t want to set these arguments on the poe command line, set the following environment variables before starting poe:

setenv MP_ADAPTER_USE shared

setenv MP_CPU_USE multiple

When using IP over the switch, the default is usually shared adapter use and multiple cpu use; we recommend that you set them explicitly using one of these techniques. You must run TotalView on an SP or SP2 node. Since TotalView will be using IP over the switch in this case, you cannot run TotalView on an RS/6000 workstation.

Performing a Remote Login

You must be able to perform a remote login using the ssh command. You also need to enable remote logins by adding the host name of the remote node to the /etc/hosts.equiv file or to your .rhosts file.

When the program is using switch-based communications, TotalView tries to start the TotalView Server by using the ssh command with the switch host name of the node.

Setting Timeouts

If you receive communications timeouts, you can set the value of the MP_TIMEOUT environment variable; for example:

setenv MP_TIMEOUT 1200

If this variable isn’t set, TotalView uses a timeout value of 600 seconds.