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.

MPI_Allreduce. Now select the properties of the frame to be viewed.
Most end users who are interested in the performance of the code sppm
should select the Connected States option with either MPI-Process
or Thread view. If you are interested in how the operating system
dispatches threads among CPUs allocated for the job, you may want
to select the Disconnected States option in the Processor view.([fig] 		<fig: preview deselected few states>All
MPI and OS-related states are deselected in the "Preview" and "Legend"
to highlight the user-defined states and the communication among the
processes.) 

 Statistics Viewer([fig] 		<fig: statistics viewer>The Statistics Viewer
and its "Legend" windows.) 

The Statistics Viewer can be invoked from the "Preview" through the
button sequence File->View Statistics. Once the viewer is up, click
on File to select a statistics file([footnote] In AIX's UTE environment,
a program like "utestat" is used to generate the statistics file.
In MPICH's MPE profiling environment, there is no tool which can generate
a separate statistics file yet.)  for processing. As shown in Figure
[fig: statistics viewer], the statistics file of sppm, sppm.stats.0301,
is displayed in the Statistics Viewer and its corresponding "Legend"
window. The "Legend" window works like that of "Preview". The title
in the Statistics Viewer provides crucial information regarding the
label of the y-axis in the viewer and that of the "Legend window".
Usually the title has the form, Viewer_Yaxis_Label / Legend_Label
/ Name_Of_Statistics. In Figure [fig: statistics viewer], the viewer's
y-axis label is "type"(i.e. MPI and user-defined states) and the label
for entities in the "Legend" window is "node" ID. The name of the
statistics is called "sum(duration)". The statistics indicate that
the state MPI_Wait takes the most time in the run and consumes an
equal amount of time in each node. The second most time-consuming
state is User Marker Interval, and again each node uses similar amount
of time in the User Marker Interval.

The statistics file usually contains more than one set of statistics.
Another set of statistics can be selected through the Graph menu tab.
Clicking on the menu tab will pull down a menu with all available
statistics data in the file. As soon as a different set of statistics
is selected, both the viewer and the "Legend" windows will be updated.
When you want to view another statistics file, be sure to close the
current opened file before opening another file. Otherwise, all newly
opened statistics will be added to the existing ones under the Graph
menu tab.

 Time Lines Window ([fig] <fig: TimeLine in Thread with all>Timeline
window in Thread view) 

Let's assume that you select Connected States in Thread view in "Preview",
when you click on the Display buttons in the "View & Frame Selector"
window. A Time Lines window as shown in Figure [fig: TimeLine in Thread with all]
will pop up. The Time Lines window provides a detailed display of
the sppm trace data as a GANTT chart with the x-axis as time and the
y-axis as thread ID. The control buttons in the top panel provide
zoom IN and OUT operations around the zoom focus, which could be set
by putting the cursor at the point of interest on the GANNT chart
and pressing the key "z". In Figure [fig: TimeLine in Thread with all],
the zoom focus is marked by a white line drawn from the top to the
botton of the diagram and is labeled as Zoom Lock in red.([fig] 		<fig: TimeLine in Thread view all zoomed in>A
zoomed in view of the Time Lines window.)  

After being zoomed in several times, the Time Lines window looks like
Figure [fig: TimeLine in Thread view all zoomed in]. At this resolution,
many more details are exposed. For instance, the MPI_Isend states
that are in navy blue in the figure become noticeable. Clicking on
any rectangle in the Time Lines canvas will pop up a "Rectangle Info"
box which contains various information regarding the subroutine call(e.g.
start and end time of the call), and various call arguments and instruction
address(es) if there are any. Clicking on the "Rectangle Info" box
again will remove the box from the screen. Also, clicking on the red
circle at the end of the arrow will pop up a "Arrow Info" box, which
provides a function similar to that of "Rectangle Info". In Figure
[fig: TimeLine in Thread view all zoomed in], MPI_Isend, MPI_Irecv,
and MPI_Wait are all nested within the user-defined state "bdrys".
This suggests that "bdrys", which is a user-defined subroutine call,
makes all these MPI calls. Also, the regularity of the pattern of
arrows can provide insight about how data are exchanged.

Several tricks and hidden operations are worth mentioning. First, it
is the trick in locating small rectangles: since all the rectangles
have a white border surrounding them, when many small rectangles are
next to each other in very low resolution, they form a completely
white rectangle. A totally white rectangle usually means a lot more
details are hidden inside. Second, a trick about scrolling: you can
drag on the scroll tab to advance to later time, but scrolling becomes
slow when there are many threads with a lot of objects. In this case,
click on the white space between the scroll tab and the end arrow
tab in the direction that you want to advance to. The operation will
allow the next time frame in the canvas to be redrawn immediately.
Third, the label of the y-axis in Thread view is (MPI-rank, local
thread ID), but it is different in different views. Because of the
limited space on the canvas, the label of the y-axis is actually written
in the tooltip of the two vertical y-axis label. The tooltip can be
activated by simply putting the cursor over any y-axis integer doublet
for few seconds. Fourth, doubly clicking on any integer doublet label
will invoke the "Time Lines Manipulation" window as shown in Figure
[fig: TimeLines Manipulation]. This window provides various operations
on the selected time line, for instance, the adjustment of time line
by changing the offset of the time line for alignment of rectangles,
or the swapping of different time lines for organizational purposes.
([fig] <fig: TimeLines Manipulation>The
Time Lines Manipulation window.) Fifth, there is a trick about the Definitions
window, which is located at the bottom of the Time Lines window. The
Definitions windows consists of a collection of definition panels,
each has a checkbox and a definition button corresponding to each
state/arrow in the Time Lines canvas. The Definitions window can be
torn out of the Time Lines window as a standalone window, or it can
be re-attached to any one of the four sides of the Time Lines window.
This allows the user to change the arrangement of the state and arrow
definitions to optimize the use of the Time Lines canvas. All states
and arrows definitions are checked by default. Unchecking any of definitions
will make the corresponding states or arrows invisible. The goal is
to help users highlight what they are interesed in by hiding the uninteresting
ones. Clicking on any of the definition buttons will bring up the
histogram window of the corresponding state/arrow. For example, the
forward arrow button will invoke a window as shown in Figure [fig: histogram, forward arrow].
([fig] <fig: histogram, forward arrow>The
histogram for the forward arrow in the frame.) The Histogram window
provides basic statistics and statistical distribution of the duration
of the selected state. The "Blink States" button enables the selected
state/arrow in the Time Lines window to flash. Again the goal is to
help highlight the objects. Finally, we note that the memory demand
for Time Lines window is very high, especially on Linux box with JDK-1.1(running
green thread). Our experience indicates that the amount of X server
memory needed to have one instance of Time Lines window running is
proportional to the size of the Time Lines window and the color depth
of the X server. If one plans to run multiple instances of Time Lines
windows at the same time, be sure to have enough physical memory for
the X server and Java Virtual Machine.

References

Anthony Chan, William Gropp, Ewing Lusk, Anthony Bomarcich, Thedore
Hoover, Yarsun Hsu, Marc Snir, and Eric Wu, Scalable Log File Format
Specification: Concepts, DRAFT on 1/25/1999.

Anthony Chan, William Gropp, and Ewing Lusk, Scalable Log Files for
Parallel Program Trace Data, DRAFT on 3/20/2000.