Performance characterization of PSU new implementation

Weiming-Hu commented 6 years ago

A summary of reasons for program refactoring:

Scientific reason: Separating similarity computation and the analog selection makes the program more flexible in providing detailed information, for example, the similarity metric, the mapping mechanism from forecast time to observation time. This change might lead to a larger memory requirement.
Performance reseason: Redesigning several costly loops by using lookup tables and removing branching leads to a significant reduction (60%) in execution time.
Coding style: Refactoring also focuses on producing readable codes with a consistent coding style. Modulization and encapsulation help the program to be more extensible.

Weiming-Hu commented 6 years ago

Ongoing work: implementing python scripts and C++ executables on Cheyenne to lay the ground for coming experiments.

Weiming-Hu commented 5 years ago

C++ executables ready. Python scripts on-going.

mturilli commented 5 years ago

Ongoing.

Weiming-Hu commented 5 years ago

vivek-bala commented 5 years ago

Simulation kernel is profiled above.

TODO:

Update wiki with profiles and experiment setup
Experiments to be done with varying search space (currently fixed to 31 days)
Use case document to be updated with details of the simulation kernel and how data is moved from anen kernels to the simulation kernel

mturilli commented 5 years ago

Duplicate #67

radical-collaboration / hpc-workflows