SC²S Colloquium - October 05, 2017

Date:	October 5, 2017
Room:	02.07.023
Time:	5:00 pm, s.t.

Vincent Bode: Parallelization of a Sparse Grids Batch Classifier

The parallelization of a sparse grids batch classifier is evaluated, implemented, and analyzed. Several potential methods of distributing the workload are identified and evaluated for their feasibility, including class-based, spatial, and batch-based methods. From this analysis, a parallelization scheme that distributes work by batches of training data is chosen due to the flexibility it offers for less than ideal datasets, even though it involves more overhead. A design for the overall structure of the program, that makes use of a master/worker role distribution, and a communication plan are created and subsequently implemented. The semantics of the implementation are shown using software architectural techniques and several challenges are presented. Next, a guide for tuning the performance of the new learner is introduced, showing which variables need to be adjusted to decrease the time to solution on multi-node systems. Using strong and weak scaling tests, the implementation was then tested for consistency and efficiency. Results that scale almost linearly, or in one case super-linearly, were observed when comparing the old sequential to the new parallel implementation.