Conjugate Gradient Benchmark

CMake flags needed to build the benchmark:

• -D Trilinos_ENABLE_Tpetra=ON
• -D Trilinos_TEST_CATEGORIES=PERFORMANCE

Path to benchmark source:

• Trilinos/packages/tpetra/core/test/PerformanceCGSolve/

Path to benchmark executable:

• $BUILD/packages/tpetra/core/test/PerformanceCGSolve/TpetraCore_Performance-CGSolve.exe

Command line for small test:

• mpirun -np 4 TpetraCore_Performance-CGSolve.exe --size=128

Suggested strong scaling study:

• mpirun -np 1 TpetraCore_Performance-CGSolve.exe --size=200
• mpirun -np 2 TpetraCore_Performance-CGSolve.exe --size=200
• mpirun -np 4 TpetraCore_Performance-CGSolve.exe --size=200
• mpirun -np 8 TpetraCore_Performance-CGSolve.exe --size=200
• mpirun -np 16 TpetraCore_Performance-CGSolve.exe --size=200

Suggested weak scaling study

Since the number of elements is the cube of --size, it's not possible to exactly double the total number of elements.

However, the sizes below keep the elements per rank at 1 million, with less than 0.5% deviation.

• mpirun -np 1 TpetraCore_Performance-CGSolve.exe --size=100
• mpirun -np 2 TpetraCore_Performance-CGSolve.exe --size=126
• mpirun -np 4 TpetraCore_Performance-CGSolve.exe --size=159
• mpirun -np 8 TpetraCore_Performance-CGSolve.exe --size=200
• mpirun -np 16 TpetraCore_Performance-CGSolve.exe --size=252
• mpirun -np 32 TpetraCore_Performance-CGSolve.exe --size=317
• mpirun -np 64 TpetraCore_Performance-CGSolve.exe --size=400

Preliminary results:

• Platform used:
• Summary or screenshot: