Benrchmark on shared memory system

[fvella]

Perform experiments by varying the following parameters [To Be Completed]. @SalvatoreDiGirolamo please feel free to comment.

• Graph size and structure:

1. Check how to pass them via command line
2. S states for scale, it generates an R-MAT with 2^S vertices.
3. E states for Edge Factor, it generates an R-MAT with an avg. degree of E

Possible values for S are 10,14, 18, 20
Possible values for E are 4, 16, 32

• Scalability over the number of threads.
  Do not exceed the physical cores. You can set up the number of omp threads as follows:
  export OMP_NUM_THREADS=<number of threads to use>.

• OpenMP Loop Scheduling and chunk size.
  it would be great to understand:

1. what are the best scheduling policy and chunk?
2. there is any relation between the chunk size and the edge factor of the graph? Any already published work on this aspect?
   Possible policies for the scheduler are [dynamic, guided, static]
   Possible values for the chunk are [2, 4, 8, 16]
   Clauses for the chunk size are [simd, monotonic, non-monotonic].
   Other question:
3. a possible ordering by out-degree may affect scheduling policy adopted or chunk size and clauses for the chunk size?
4. any published work on user-defined schedulers for irregular applications like graph analytics on ARM v8?

• Cache Measurement use PAPI @SalvatoreDiGirolamo

[fvella]

@strausza I strongly recommend modifying the code in order to automatize experiments.

[SalvatoreDiGirolamo]

For the PAPI measurements, I'd keep using libLSB. It is also already integrated in CLaMPI (configure with --with-liblsb), so you can performance statistics for free.

To use libLSB (https://github.com/spcl/liblsb/, use testing branch) with PAPI and MPI, you will need to configure it with:
--with-papi=<papi path> --with-mpi (check the configure log to see if it actually found them). Once you have it compiled, compile CLaMPI with --with-liblsb<path to liblsb build>. With this configuration, you can specify the PAPI counters you want to observe via environment variables (e.g., LSB_PAPI1=<PAPI COUNTER>; type papi_avail to see the available PAPI counters). Once you run with this configuration, you will get one output file per MPI rank, containing the measurements.

Let me know if you have issues with this. ;)