A Beowulf Parallel Computing System Scalability Research And Applications

In recent years, the high performance computing (HPC) technology vigorous development, more and more science computing problem could be solved through the parallel programming. But in domains and so on basic theory study, the question scale is getting bigger and bigger, need more computing resources, therefore in order to improve the handling ability, the parallel computing system must along with it expansion. However system's efficiency is not increases along with the node number presents the linearity grows, when the system arrives at the certain scale will present the efficiency to reduce, the execution time hard to forecast and so on questions, the question will be more serious in heterogeneous system. Therefore, deep research scalability will be helpful to solve the large-scale application question parallel computing system's performance to make the appropriate appraisal, is also beneficial in the parallel algorithm and parallel system's design and the improvement.Beowulf parallel computing system based on the MPI as the HPC's a branch, has inexpensive, easy to manage, the performance-to-price ratio higher numerous merit, the application is getting more and more widespread. This paper mainly analyzes heterogeneous Beowulf parallel system's scalability from the efficiency aspect, when the processor node increases, how should the question scale change can maintain the same efficiency, and predict system's extendibility.As heterogeneous systems handling capacity of each node differences, task allocation strategy will seriously affect the quality of the system scalability. This paper from research task allocation, load balancing starting, improved in the MPICH task allocation shortcoming, and constructed one based on OpenPBS and the MPICH load balancing model. Model using the concept of relative processing speed, the handling capacity of each node were quantified, scheduling nodes allocate task according to the capacity of each node weights in order to achieve the load balancing system as a whole. The experiment indicated that the model may carry out the task allocation to various nodes reasonably, has provided the good foundation for the analysis system's scalability as well as the scalability experiment.At present the scalability research mainly concentrates in the parallel algorithm and the parallel system's combination, that is, how to increase along with the node number expands the question scale, causes the execution time to be reasonable, and the efficiency is high.ISO-efficiency model has promulgated the computing performance under the parallel algorithm and the parallel system, but it mainly aims at the homogeneous system, has not considered each processing node the difference. Although heterogeneous system was already getting more and more common, but did not have the appropriate definition in the efficiency and the scalability concept aspect to study its characteristic. This paper improved under the homogeneous system's ISO-efficiency, proposed an efficiency's definition, enables it simultaneously to apply the homogeneous system and heterogeneous system, and constructed one to suit the homogeneous and heterogeneous system's ISO-efficiency models, found the necessary and sufficient condition of maintaining the same efficiency. From this we can analyze the system scale and the question scale should how to change, can cause the efficiency maintains consistent. Finally, this paper has taken a series of experiments to verify the above theory, results show that this method useful and effective, and better analysis of the homogeneous and heterogeneous system's physics and ability scalability, and be able to predict the system's scalability.