From Tau Wiki
Jump to: navigation, search

A short Demo of TAU for use with the POINT Live DVD version 2, built 5/18/09.

Instrumenting NPB

open a terminal window
cd workshop-point/NPB3.1
vi config/make.def

Notice that the MPIF77 and FLINK variable is set to This will enable TAU's automatic instrumentation with PDT.

Setting the TAU makefile

close vi.
setenv TAU_MAKEFILE $TAU/Makefile.tau-mpi-pdt

This tells TAU to perform a basic instrumentation using PDT and the TAU MPI wrapper library. Now build the BT example program:

make bt CLASS=W NPROCS=16

Running the NPB example

cd bin
mpirun -np 16 ./bt.W.16

TAU Profiles will automatically be generated in the current directory, one profile file per thread.

bt.W.16         profile.11.0.0  profile.15.0.0  profile.5.0.0  profile.9.0.0
profile.0.0.0   profile.12.0.0  profile.2.0.0   profile.6.0.0
profile.1.0.0   profile.13.0.0  profile.3.0.0   profile.7.0.0
profile.10.0.0  profile.14.0.0  profile.4.0.0   profile.8.0.0

Viewing TAU profiles

To get a simple summary of the TAU profiles type:


This gives you a basic idea of how much time was spent in different NPB routines.


Let's view this profile in TAU's ParaProf profile viewer


Paraprof will load the profile and show a single bar representing Node 0. Each colored subsection represents a different routine in NPB program. The length of a subsection is proportional to the exclusive time spent in that routine.

Right click on the "Node 0" label
Select "Show Thread Bar Chart"

A new window will pop up ordering each routine by the amount of exclusive time.

Click "Windows" -> "Group Legend"
Right click on "MPI" select "Hide This Group"

Now the MPI routines are excluded from all profile views.

close Group Legend window
Click "Options" -> "Select Metric..." -> "Inclusive"

Now the bars are ordered by inclusive time.

Selective Instrumentation

Click on the "TAU: ParaProf Manager" window
Right click on the trial name: "bin/NPB3.1/...."
Select "Create Selective Instrumentation File"

A window will pop up showing a number of routines. These routine have been flagged by TAU as lightweight routines. Lightweight routines are defined as routines that have less than 10 microseconds per call and are called more than 100,000 times (these parameters can be changed--see the form above). Excluding these routines from instrumentation will help lower the instrumentation overhead.

Click "Save"
Close ParaProf

Tell TAU to use this newly created selective instrumentation file:

setenv TAU_OPTIONS -optTauSelectFile=`pwd`/select.tau

Now rebuild the BT program.

cd ..
make clean bt CLASS=W NPROCS=16

This time let's run the program with callpath profiling enabled:

cd bin
mpirun -np 16 ./bt.W.16

Open ParaProf:

Right click on 'Node 0'
Select "Show Thread Call graph"

close window
Right click on 'Node 0'
Select "Show Thread Statistics Table"
double click on "MPBT"

Here we can see that the MPBT routine calls INITIALIZE twice.

Comparing Trials

ParaProf is also very useful for comparing different runs of the same application. So far we have been compiling the NPB applications with a high level of optimization (-O3), can we quantify the performance benefit of doing so?

close ParaProf
cd ..
vi config/make.def
Comment out line 50, "FFLAGS = -O3"
write file and close vi
make clean bt CLASS=W NPROCS=16

Before running the experiment let's package the performance data we have already gathered:

cd bin
paraprof --pack bt-O3.ppk

mpirun -np 16 ./bt.W.16
paraprof --pack bt.ppk

Now open both profiles in ParaProf

paraprof *.ppk
Click on 'TAU: ParaProf Manager'
Right click on the 'bt.ppk' trial
Select 'Add Mean to Comparison Window'
Click on 'TAU: ParaProf Manager'
Right click on the 'bt-O3.ppk' trial
Select 'Add Mean to Comparison Window'

In the Comparison Window we can see a comparison between these two runs. Each bar shows the exclusive time for each routine. Some routine show little variation (MPI_Init) while some show a huge speedup when compiled with -O3 (X_BACKSUBSTITUTE).


close paraprof

We need to change the TAU makefile to one built with PAPI

setenv TAU_MAKEFILE $TAU/Makefile.tau-papi-mpi-pdt
cd ..
make bt CLASS=W NPROCS=16
cd bin

We set PAPI counters using environment variables, find out which ones are available


Let's select FP instructions, total instructions and papi wall clock time:

mpirun -np 16 ./bt.W.16

Remove old profiles:

rm profile.*
Click on "TAU: ParaProf Manager" window
Click "Options" -> "Show Derived Metric Panel"
click on "P_WALL_CLOCK_TIME"
click on "PAPI_FP_INS"
select "Divide" 
click "Apply operation"

Right click on "Mean", select "Show Mean Bar Chart"

Routines are now sorted my FP instructions per seconds.