Abstract
The design of a power-constrained CubeSat is a complex problem involving several disciplines that are coupled. Algorithms and hardware for scientific computing have advanced to the point that recently, design and operations optimization of a CubeSat was successfully performed involving 7 disciplines, over 25,000 design variables, and roughly 2.2 million total variables modeled. This paper addresses the bottlenecks of this algorithm in computer memory costs and execution time. This is done through a new parallel computational modeling framework that automates many aspects of distributed-memory parallel computing, and an analytic approximation for the solar cell model that is significantly more efficient than the previous model. Results show improved scaling of execution time with the number of unknowns in the problem and nearly an order-of-magnitude improvement in gradient computation time.
Presentation
Efficient and Scalable Computational Design of a Small Satellite
The design of a power-constrained CubeSat is a complex problem involving several disciplines that are coupled. Algorithms and hardware for scientific computing have advanced to the point that recently, design and operations optimization of a CubeSat was successfully performed involving 7 disciplines, over 25,000 design variables, and roughly 2.2 million total variables modeled. This paper addresses the bottlenecks of this algorithm in computer memory costs and execution time. This is done through a new parallel computational modeling framework that automates many aspects of distributed-memory parallel computing, and an analytic approximation for the solar cell model that is significantly more efficient than the previous model. Results show improved scaling of execution time with the number of unknowns in the problem and nearly an order-of-magnitude improvement in gradient computation time.
