tourstepbystep.txt

MPICH是MPI的重要研究,提供了一系列的接口函数,为并行计算的实现提供了编程环境.

A Tour of Jumpshot-3

Anthony Chan, William Gropp, Ewing Lusk

4/27/2000

Abstract

Jumpshot-3 is a display program for trace data in scalable log format,
[slog1], which is designed to store a large number of drawable objects[slog2]
generated from the backend of a parallel program, for example AIX's
UTE slogmerge or MPICH's MPE profiling library. Basically, Jumpshot-3
is a GUI to display rectangles that represent the various MPI and
user-defined states and arrows indicating messages exchanged between
those states. This article illustrates the basic functionalities of
Jumpshot-3 through a step-by-step guide of how one would usually use
Jumpshot-3 to view a big logfile.

 Background

Assume you are trying to view a slog file called sppm_ic2a.slog([footnote] The
logfile is provided by Dave Wootton at IBM.) , which is generated from
a run of the MPI program sppm on IBM's SP with shared-memory nodes.
The logfile is located in the directory ~/jumpshot-3/logfiles/. Also
assume that Jumpshot-3 has been successfully installed([footnote] For
details on how to install and start running Jumpshot-3, see README.slog
and UserGuide.txt in ~/jumpshot-3.) . Typing the full pathname to the
Jumpshot script which is located at ~/jumsphot-3/bin/ will bring up
Jumpshot-3.

 Main Control Window([fig] <fig: main_win>The Main Control Window
of Jumpshot-3)  

The first window Jumpshot-3 pops up is the Main Control window, which
is shown in Figure [fig: main_win]. The File menu tab brings up the
file selection menu. Here, doubly click to the right directory where
sppm.slog is located, and highlight sppm.slog as the logfile to be
processed. Then the path of the logfile relative to the current working
directory will be printed in the logfile box, in the middle of the
main control window. As soon as the pathname of the logfile appears,
the Read button will be enabled. Click on the Read button to display
the "Preview" of the logfile.

The System menu tab allows you to choose the specific "Look & Feel"
for the whole GUI system of Jumpshot-3([footnote] For Unix system,
only Metal and Motif Look & Feel are allowed because of the license
issue( from Sun's documentation ).) . The Help menu tab provides electronic
documentations to Jumpshot-3, including three submenus: Manual, Tour,
and About. Manual provides a brief explanation of all the buttons
in Jumpshot-3. Tour is the HTML version of this document. About provides
the version and contact information for the copy of Jumpshot-3 being
used.

 Statistical Preview([fig] <fig: preview>The Preview: View & Frame
Selector Window.) 

The goal of the Statistical Preview is to help users select a frame
for viewing. It consists of two windows. The major window shown in
Figure [fig: preview], is titled "View & Frame Selector", but is usually
called "Preview". It contains a graphical representation of all the
activities going on during the run of the program sppm. Activities
in "Preview" are computed in the following way. The whole duration
of the job sppm is divided into 512 time bins. For each time bin,
there are counters for each state and arrow that occur in the time
bin. The counter of each state/arrow will be incremented by a statistical
weight that equals the ratio of the duration of state/arrow to the
width of time bin that the state/arrow is in. Because of this definition
of the statistical weight for each state/arrow, very short duration
states or arrows become statistically insignificant and will not contribute
much to the accumulative activities of all the states and arrows.
The visual consequence is that states/arrows with very short duration
are not noticeable in "Preview". The graphical representation of the
activities of the logfiles is, by default, in accumulative histogram
mode, which can be turned on/off through the button selection sequence
Graph->Type->Bar. The other mode is the noncumulative curve mode which
can be enabled by the button selection sequence Graph->Type->Line.
It is shown in Figure [fig: preview mode line]. Since it is noncumulative,
as in histogram mode, it distinguishes the relative importance of
various states.([fig] <fig: preview mode line>Preview in non-accumulative
curve mode.)  In the lower left-hand corner of "Preview", there is a
"Frame Information" panel, which lists the number of frames in the
logfile, N. In this case, there are only 3 frames in the file. The
frame size is determined by the program that generates the slog files
([footnote] In
AIX's UTE environment, a program called slogmerge allows user to define
the size of the frame. In MPICH's MPE environment, direct logging
mechanism implicitly assume a frame size. However, end user can change
the frame size through the use of program clog2slog.) . In general,
the smaller the frame, the better the performance of Jumpshot-3 will
be in refreshing the screen. But if an overly small frame size is
used, it will be increasingly difficult to navigate to the frame that
is being selected. And each frame will contain too little data for
the user to form a coherent overall picture of the code. Next to the
label Current Frame, there is a dialog box that shows the current
frame index, which proceeds from 0 to the last frame of the file,
N-1, and is initialized to 0 when the GUI starts. The user can change
the frame index by typing in the dialog box or through buttons in
the "Frame Operations", where Previous and Next buttons decrement
and increment the frame index, respectively. But the user is also
allowed to click on the graphical display to select a frame of interest.
A red line will then show up in the graph to highlight the frame selected.
Currently the red line is located in the middle of the frame, so the
red line is moved nonuniformly along the x-axis. The redundancy in
frame selection operations allows the end user to fine tune which
frame to view when the logfile contains a lot of frames. For an MPI
program running in a multithreaded environment as in the AIX SMP box,
the logfile may contain information about threads. In that case, the
"Connectivity Options" and "View Options" panels become very useful.
The user can choose different combinations of these options to see
how threads are dispatched and used in an MPI Program. More details
will be in the next section.([fig] <fig: preview's legends>The Legends
of Preview) Next to the "Preview" there is a smaller window titled "Legend",
which contains a set of radio buttons with different colors. They
are tagged with the names of the MPI and user-defined calls made in
the code sppm. An illustration of "Legend" is shown in Figure [fig: preview's legends].
Below the list of radio buttons there are Select/Deselect, All, None,
and Change Color buttons. 

"Preview" and its corresponding "Legend" windows meant to be used side
by side to help you select frames that are of interest. If some of
the states displayed in the "Preview" seem useless or confusing, it
can be easily removed. First highlight the state through click on
the name in the "Legend"; then click on Select/Deselect button to
remove them from "Preview". In Figure [fig: preview deselected few states],
all MPI and system-related states are deselected in "Preview" to highlight
the user-defined states "layout", "setup", "bdrys" and "glbl". It
becomes apparent that all the communications between the user MPI
processes are done in regular interval. Also, all communications(i.e.
arrows) are within the user-defined state "bdrys". Assume you are
interested in the behavior of the program between the last two instances
of MPI_Allreduce. Then you can simply click on None to deselect all
the states and arrows, highlight MPI_Allreduce, then click Select/Deselect
to display only MPI_Allreduce in the "Preview". Now click on any region
between the last two MPI_Allreduce's in the "Preview". Then the frame
index is changed from 0 to 2, and a red line is positioned in the
middle of the frame 2 and is also between the last two instances of