|
XMPI is an X/Motif based graphical user interface for running,
debugging and visualizing MPI programs. Extensive MPI information is
extracted from a running application on demand, or from a cummulative
log of communication. Both sources are tightly integrated with an
application overview window and any number of single process focus
windows.
XMPI is an excellent console for teaching MPI because students can
vividly see the results of message-passing functions.
There are currently two versions of XMPI
available:
- XMPI 2.2 works with LAM/MPI version 6.3.2.
- For those needed XMPI for LAM/MPI version 6.5.9,
there is a beta release on the beta web page.
Both versions are available from the LAM Tools download site.
Neither version of XMPI will work with versions of LAM prior to
the one listed. Please upgrade your LAM/MPI to work with XMPI
properly.
Key Features
- runtime snapshot of MPI process synchronization
- runtime snapshot of unreceived message synchronization
- single process focus detailing communicator, tag, message length, and
datatype
- runtime and post-mortem execution tracing with timeline and
cummulative visualizations
- highly integrated snapshot from communication trace timeline
- process group and datatype type map displays
- matrix display of unreceived message sources
- assembles MPI applications from local or remote programs
- easy startup and takedown of applications
Application Overview
After interacting with XMPI to run an MPI application, a honeycomb
representation of all processes is displayed in the main (overview)
window. Icons inside each cell indicate the execution state of the
process and alert the user to any unreceived messages. Processes are
identified by their rank in MPI_COMM_WORLD. A button click updates
the information across the entire application - while the application
is running (or perhaps deadlocked or otherwise hung). This simple
capability to inspect, at runtime, the synchronization state of
processes and messages is a very effective debugging tool. Source
code steppers and "printf" are often used to get at the same
information, but in a more cumbersome manner.
Application Overview: processes and messages
at a glance
Process Focus
A process window with full MPI details on a single process and its
unreceived messages is popped up by a mouse click on an interesting
cell in the application overview. The key MPI communcation parameters
are displayed: communicator, source rank, destination rank, and
element count. For the datatype, a button pops up a description of
the type map. For the communicator, a button highlights the group
membership in the overview window. Several process windows may be on
the screen simultaneously so that the user can focus on problem
interactions between a few processes. As with the overview window,
all of the process windows are updated when a new snapshot is
taken.
Process Focus: details on process status
and its unreceived messages
 The datatype button displays
the type map.
 The next button shows
the next group of identical messages.
 The group button highlights
the members of the communicator's process group.
Communication Trace Visualization
XMPI can cause the communication activity of an application to be
traced. The resulting trace data can be extracted at runtime or after
the application completes. The trace data is visualized in two tried
and true ways: the communication timeline and the kiviat radial chart.
Within XMPI however, these common views are highly intergrated into
the overall debugging picture. Using a dial in the timeline window, a
snapshot of the application state can be taken as if the same
snapshot was taken at the selected time during the application's run.
The results are displayed in the same manner, using the overview
window and the processes windows.
Communication Timeline: Placing the dial
gives a full snapshot of MPI details.
Kiviat: Process states are cummulative
from the start to the dial time.
Matrix: Unreceived messages counted by
source rank.
How to Get XMPI
XMPI was originally developed at the Ohio Supercomputer Center and is
currently being developed by the Open Systems Laboratory at Indiana
University. It is freely available from the
LAM download site.
|
