Teaching and public dissemination of HPC-related topics can be a difficult task. First, there is a number of sophisticated hardware and software components, each complex on their own and often showing non-intuitive interaction. Second, we consider education in HPC among the more difficult topics in computer science since larger distributed memory systems are ubiquitous yet inaccessible and intangible to students and the general public. By design, most of these systems are physically and logically encapsulated to provide a layer of abstraction that facilitates their use, simultaneously hindering on-hands education.
In this talk, we discuss our Cluster Coffer, a mobile miniature cluster computer that reduces the entry barrier to HPC in teaching and public outreach through direct access to all major components of a distributed memory system. It consists of an aluminium case with 16 NanoPI M4 ARM boards that form the compute nodes, while a NanoPC T4 board serves as the login node. All boards are equipped with an actively-cooled Rockchip RK3399 SoC with two higher-power Cortex-A72 and four lower-power Cortex-A53 cores and use Gigabit Ethernet as the interconnect. All components are low-cost and easy to maintain to minimize operational cost. Preliminary peak performance is 101 GFLOPS (HPL Rmax), total system power draw is less than 200 W.
We employ the Cluster Coffer for two main use cases. First, we address teaching students in multi-objective optimization and Pareto optimality by comparing execution times of HPL or HPCG against energy consumption for varying cores or DVFS settings. Since the Cluster Coffer is equipped with high-resolution per-node power instrumentation, students can observe the trade-off of e.g.\ wall time and energy consumption for various points in the problem space.
The second use case demonstrates the possibilities of HPC to the public at the example of a space-weather prediction proxy app, simulating solar winds (originating from the H2020 project AllScale). However, experience has shown that high levels of audience engagement are crucial for maximizing public outreach success. Therefore, the proxy app was enriched with an interactive element by including live input from participating audience through a motion-detecting Kinect camera. It is attached to a host PC, which converts motion information into charged particles and transfers them to the Cluster Coffer, to be incorporated into the continuously running simulation, while simultaneously all current particle data is sent back to the host PC for visualization. This continuous loop allows the audience to directly influence the simulation and observe effects in real time. We have successfully demonstrated the Cluster Coffer at numerous outreach activities such as science fairs and open-day events. CAD blueprints, a bill of materials, and setup software is available online for reproduction.