<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
</head>
<body bgcolor="#ffffff" text="#000000">
Sept 4, 2008: HPCT: Summary of the issues<br>
<br>
1) We want to be able to reduce statistics, generated in mpi_profile.*
files, by some sort of MPI_Start_Statistics()/MPI_Stop_Statistics()
function calls. As we know from I-hsin,
MPI_trace_start()/MPI_trace_stop() calls control tracing part only.<br>
<br>
2) According to I-hsin, "mpi_profile.<rank number> are generated
per each MPI rank." We do not see this behavior: 4 files are generated
instead with minimal, maximal, mean communication time, and collective
over all tasks statistics file called mpi_profile.0. Please clarify 1)
what is printed at the end of the run, 2) how to print statistics from
a particular MPI task. Examples of the test run would be very useful to
have.<br>
<br>
3) What is IHPCT_BASE env. variable? When I should use it?<br>
<br>
4) peekperf does not show sources for Fortran program.<br>
<br>
5) Fortran instrumented program generates "Error getting the executable
name, line number information will not be available" message.<br>
<br>
6) peekview single_trace gives the node 0, 8, and 15. Example program
was sent to the list already and contains 2 Sends and 2 Recvs. Where
are the other communications? See attached screenshot.<br>
<br>
7) peekview on selective instrumentation does not show correct number
of calls. From 1 Send and 1 Recv, we see nothing on peekview screen.<br>
<br>
8) HPM part does not work: core file at runtime. Please, give us
example of 1) how to compile HPM-instrumented program, 2) what is the
result of the run.<br>
<br>
9) Xprofiler has 6 open issues:<br>
<ol>
<li> Load Files, Filter -> Uncluster functions, File -> Screen
Dump -> Select Target Window,
<ol>
<li> clicking now on the main window may occasionally cause crach
</li>
<li> resulting file Xprofiler_screenDump.ps.0 is a balack-box
Postscript
</li>
</ol>
</li>
<li> The meaning on exterior scroll bars in not clear, they never
scroll.
</li>
<li> <b>Removing Histogram functionality is unacceptable</b> A more
complicated multistep example
<ol>
<li> Load Files
</li>
<li> Report -> Flat Profile
</li>
<li> Code Display -> Show Source Code
</li>
<li> File -> Save as; File -> Close
</li>
<li> Code Display -> Show Histogram w/Source Code
</li>
<li> crash
</li>
</ol>
</li>
<li> Load Files, Report -> Flat Profile, Code Display -> Show
Source Code, File -> Save as
<ol>
<li> Close the window by clicking X (not File -> Close)
</li>
<li> Select another function in Flat Profile
</li>
<li> Code Display -> Show Source Code; File -> Save as;
crush
</li>
</ol>
</li>
<li> Load Files, Report -> Flat profile, Show Source Code, search
anything non-existent, press enter, crash
</li>
<li> Load Files, Report -> Flat Profile, try to search anything
like 'asfsafd', once not found press Enter, crash
</li>
</ol>
<br>
Vitali<br>
<br>
</body>
</html>