To build the Pintool, you need to first download and unpack Intel's PIN.
Then, configure the build:
cmake -DCMAKE_INSTALL_PREFIX=<install-dir> -DPIN_ROOT=<pin-dir> ..
Then build and install the tool.
First, you need to specify instrumentation regions in your code:
#include <memreuse.hpp>
__memreuse_roi_begin("foo");
__memreuse_roi_end("foo");See the example in examples/test.cpp.
Extend your compilation flags with -I<install-dir>/include and linking flags with -L<install-dir>/lib -lmemreuse.
Run your application with:
<build-dir>/bin/memtracer -o tracer.out <your-app-with-arguments>
The files tracer.out.{region}.{counter} will contain the statistics on memory accesses for entered regions.
To run an MPI application with OpenMPI, use:
mpirun -np 2 <mpiflags> <build-dir>/bin/memtracer_ompi -o tracer.out <your-app-with-arguments>
The output for each MPI process will be in tracer.{rank}.out.{region}.{counter}.
Install Python dependencies:
pip install -r tools/requirements.txt
Then run the the postprocessing analysis by pointing to the directory and naming schemes of produced files:
tools/postprocessing.py <dir>/tracer.out <out-file>
After a while, the results will be written to the output file. For each iteration, we compute the number of new memory regions that have been written or read:
,iteration,read_bytes,new_read_bytes,write_bytes,new_write_bytes
0,0,5052433,5052433,0,0
1,1,3371595,3371595,3177832,3177832
2,2,473686,15020,140083,21080