Research On High Performance Parallel Processing For The Satellite Collision Prediction System

With the exploration to the outer space, the number of space crafts and space debris are quickly increasing. The probability of collision of them has increased too. Space crafts, especially manned spaceflight, are being menaced greatly. To avoid collision with these space objects, the collision of space objects must be analyzed and predicted. But the number of space objects is huge, more than 12,000, and the movement rule of these space objects is very complicated. It takes huge numbers of computations and large memory storage to calculate the flight relationship between main object and other objects. It is impossible for a single processor computer to accomplish this task in real time.This paper conducts research on satellites collision prediction system on parallel computer systems. It focused on load balancing and single processor performance optimization to improve the performance of the program. Based on the research result, a cluster based, efficient satellites collision parallel prediction system is developed. The work of this paper is summarized as follows:(1) A parallel method for the satellite collision prediction system is proposed. The parallel program is implemented in SPMD model based on MPI(Message Passing Interface).(2) By analyzing the merits and disadvantage of several load balancing models, this paper proposes centralized load balancing algorithm, which partially solve the problem of load unbalancing. In order to make further improvement, this paper proposes a light-load reporting dynamic load balancing algorithm based on data transferring techniques. Experimental result shows that the light load reporting dynamic load balancing algorithm outperforms the other algorithm significantly. When run on 32 processors, the application' s parallel efficiency has improved about 21%.(3) By profiling and hardware performance monitoring based measurement, the performance of program of satellites collision prediction is tested and analyzed, and the performance bottleneck of the program is fund. Then, optimization methods are proposed to improve the program' s performance. Efficient search for the program performance optimization factors is done based on Lega framework proposed by Yonggang Che. The experimental results on YINHE-Star parallel computer indicate that the program' s performance improved about 16% on single processor and about 9. 2% on 64 processors, after performance optimization.