Profiling with IPM and Vampir - Revision history

S.wales at 03:54, 12 December 2019

2019-12-12T03:54:45Z

S.wales at 05:21, 18 March 2019

2019-03-18T05:21:00Z

ScottWales: Imported from Wikispaces

2012-08-27T05:31:45Z

Imported from Wikispaces

New page

NCI has a variety of profiling tools that can be helpful to use when developing. IPM is a low-overhead way to collect messaging information, while Vampir/Trace Analyser is a more in-depth profiler.

To use these with the UM some hand-edits are provided. You don't need to recompile when you add IPM, but you will with Vampir and Trace Analyser.

=IPM=
==To run==
Add this handedit to Input/Output Control->User hand edit files in the UMUI
<syntaxhighlight>
~access/umui_jobs/hand_edits/profiling/ipm.sh
</syntaxhighlight>

==To view==
===OpenMPI===
When using OpenMPI IPM creates an xml file in the UM output directory, who's name consists of your user name plus a timestamp. ipm_view is used to display the file, showing a number of graphs with the call time and load balance.
<syntaxhighlight>
$ module load ipm
$ ipm_view $DATAOUTPUT/$USER/saatd/saw562.1346035173.302475.0
</syntaxhighlight>
[[File:ipm.png]]
[http://nf.nci.org.au/wiki/Profiling/MPI#IPM | NCI IPM notes]

===Intel MPI===
Intel MPI's IPM produces a text file with communications information
<syntaxhighlight>
$ cat $DATAOUTPUT/$USER/saatc/stats.txt
</syntaxhighlight>
<syntaxhighlight>
Intel(R) MPI Library Version 4.0 Update 3
<u>__ MPI Communication Statistics __</u>

Stats level: 3
P2P scope:< FULL >
Collectives scope:< FULL >

[[user:ScottWales | 1346044901]] Process 0 of 4 on node vayu2 lifetime = 394501.92

Data Transfers
Src Dst Amount(MB) Transfers
-----------------------------------------
000 --> 000 0.000000e+00 0
000 --> 001 5.019608e-01 6
000 --> 002 1.968384e-03 2
000 --> 003 0.000000e+00 0
=========================================
Totals 5.039291e-01 8

Communication Activity
Operation Volume(MB) Calls
-----------------------------------------
P2P
Csend 3.929138e-03 4
Send 5.000000e-01 4
Bsend 0.000000e+00 0
Rsend 0.000000e+00 0
</syntaxhighlight>
[http://nf.nci.org.au/facilities/software/intel-mpi/4.0.3.008/doc/Reference_Manual.pdf | Intel-MPI docs]

=Vampir=
==To run==
Add this handedit to Input/Output Control->User hand edit files in the UMUI
<syntaxhighlight>
~access/umui_jobs/hand_edits/profiling/vampir.sh
</syntaxhighlight>

==To view==
===OpenMPI===
When using OpenMPI the vampir utility is used to view the output files and graph the call timeline
<syntaxhighlight>
$ module load vampir
$ vampir $DATAOUTPUT/$USER/saatd/saatd.otf
</syntaxhighlight>
[[File:vampir.png]]
[http://nf.nci.org.au/facilities/software/Vampir/7.4.0/Manual.pdf | Vampir docs]

===Intel MPI===
Intel has its own version of Vampir, called Trace Analyser. It has a similar function, to view graphs of call timelines etc. use the Charts menu.
<syntaxhighlight>
$ module load intel-itac
$ traceanalyser $DATAOUTPUT/$USER/saatc/saatc.stf
</syntaxhighlight>
[[File:itac.png|1024x664px]]
[http://nf.nci.org.au/facilities/software/intel-itac/8.0.1.009/doc/ITC_Reference_Guide.pdf | Trace Collector docs]
[http://nf.nci.org.au/facilities/software/intel-itac/8.0.1.009/doc/ITA_Reference_Guide.pdf | Trace Analyser docs]

← Older revision		Revision as of 05:21, 18 March 2019
Line 1:		Line 1:
		+	{{Pre UM10}}
		+	[[Category:Unified Model]]
		+
	NCI has a variety of profiling tools that can be helpful to use when developing. IPM is a low-overhead way to collect messaging information, while Vampir/Trace Analyser is a more in-depth profiler.		NCI has a variety of profiling tools that can be helpful to use when developing. IPM is a low-overhead way to collect messaging information, while Vampir/Trace Analyser is a more in-depth profiler.

@@ Line 9: / Line 9: @@
 ==To run==
 Add this handedit to Input/Output Control->User hand edit files in the UMUI
-<syntaxhighlight>
+<syntaxhighlight lang=text>
 ~access/umui_jobs/hand_edits/profiling/ipm.sh
 </syntaxhighlight>
@@ Line 16: / Line 16: @@
 ===OpenMPI===
 When using OpenMPI IPM creates an xml file in the UM output directory, who's name consists of your user name plus a timestamp. ipm_view is used to display the file, showing a number of graphs with the call time and load balance.
-<syntaxhighlight>
+<syntaxhighlight lang=text>
 $ module load ipm
 $ ipm_view $DATAOUTPUT/$USER/saatd/saw562.1346035173.302475.0
@@ Line 25: / Line 25: @@
 ===Intel MPI===
 Intel MPI's IPM produces a text file with communications information
-<syntaxhighlight>
+<syntaxhighlight lang=text>
 $ cat $DATAOUTPUT/$USER/saatc/stats.txt
 </syntaxhighlight>
-<syntaxhighlight>
+<syntaxhighlight lang=text>
 Intel(R) MPI Library Version 4.0 Update 3
 <u>__ MPI Communication Statistics __</u>
@@ Line 62: / Line 62: @@
 ==To run==
 Add this handedit to Input/Output Control->User hand edit files in the UMUI
-<syntaxhighlight>
+<syntaxhighlight lang=text>
 ~access/umui_jobs/hand_edits/profiling/vampir.sh
 </syntaxhighlight>
@@ Line 69: / Line 69: @@
 ===OpenMPI===
 When using OpenMPI the vampir utility is used to view the output files and graph the call timeline
-<syntaxhighlight>
+<syntaxhighlight lang=text>
 $ module load vampir
 $ vampir $DATAOUTPUT/$USER/saatd/saatd.otf
@@ Line 78: / Line 78: @@
 ===Intel MPI===
 Intel has its own version of Vampir, called Trace Analyser. It has a similar function, to view graphs of call timelines etc. use the Charts menu.
-<syntaxhighlight>
+<syntaxhighlight lang=text>
 $ module load intel-itac
 $ traceanalyser $DATAOUTPUT/$USER/saatc/saatc.stf