Improving Student Outcomes with Final Parallel Program Mastery Approach for Numerical Methods

Abstract Student projects for "Numerical and Parallel Programming", a course taught at California State University Chico, stress ability to put theory into practice. The course normally includes a final project, allowing students to assess their success with verifiable numerical methods (e.g., using MATLAB) and speed-up compared to Amdahl's law. The pandemic presented new challenges for projects, and this course had the extra challenge of proprietary cluster software, now only remotely accessible. The course has a reputation for being difficult based upon pre-course surveys. Given the limitations, it would have been easiest to simply eliminate the project and focus on exercises and assessments. However, based upon student mid-term surveys, it was clear that students preferred a project over online assessment, and the new remote constraints presented an opportunity for change, with the added goal to allow all students to share their work with their cohort in a new way. The projects for parallel programming focus on numerical methods commonly used in science and engineering, traditionally summarized and/or presented one-on-one with faculty as individual efforts. Students are expected to use divide-and-conquer approaches to design parallel programs for speed-up using well-known numerical methods from calculus combined with algorithms learned in class. To maintain course learning objectives and improve upon them while overcoming the new pandemic limitations, three specific project modifications were made: 1) All projects could be completed on a home system using equipment with cost less than one hundred dollars or could be completed on the existing CSU cluster. 2) Projects could focus on mastery of a prior problem given as an exercise with emphasis on a detailed code walk-through for parallel design or could be more creative with instructor approval. 3) All projects were required to include a brief report, but also a 20-to-30-minute video of the build, run, and code design completed with a detailed walk-through shareable with the CSU learning management system. Longer term, CSU Chico is investigating a remotely accessible cluster built using the same at-home hardware but scaled to 48 nodes. The goal is to support not only POSIX threads and OpenMP shared memory scaling, but also distributed memory MPI (Message Passing Interface) and shared memory CUDA (Compute Unified Device Architecture). In general, the at-home use of single node versions showed no new issues compared to use of the proprietary software. The project video walk-through appears to have positively motivated students and gave them multiple options to engage and complete the course. While not as ideal as all students presenting to each other, the new approach scaled well for large classes and allowed students to share their experience. The pedagogical experiment introduced by this new project sharing approach and low-cost hardware scaling shows promise, but results are preliminary. The fall 2020 results from pre-course, mid-course, and post-course surveys are summarized in this paper along with instructor lessons learned and plans to repeat in spring 2021.