Design And Implementation Of Network And Data Transmission In Supercomputer Center

High performance computing (HPC) is a major branch of information technology. HPC level has become a key index of a country's competitive ability. Supercomputer center is an important HPC provider. The reliability and availability of network system in supercomputer center have become more important than ever.This paper has designed and implemented a network system in supercomputer center. During this period, we researched data transmission of supercomputer center,and finally we found the reason for unstable data transmission is load unbalancing.Especially for supercomputers,static scheduling to shared resources causes severe load unbanlancing. So, this psper is focused on designing network system in supercomputer center. Layer seven switches and redundancy structure are respectively used to solve load unblancing in multi-internet links and internal network.The other important part of this paper is implementing a dynamic feedback scheduling algorithm as an extension for LVS. At present, the LVS software has eight scheduling algorithms, but all of them belong to static scheduling algorithm. As we all know, due to the variation of the tasks'attributes and the dynamic environment, the static scheduling algorithms, which depends greatly on the tasks' or servers' static attributes, can't distribute a workload evenly amongst a cluster of back-end servers. In order to improve the request allocation decisions made, we presents the dynamic feedback scheduling algorithm for flexible workload. Dynamic feedback scheduling requires mechanisms for monitoring and evaluating the current load on each server, gathering the results, combining them, and taking real-time decisions.In order to test the effectiveness of the dynamic feedback system, we set up an experimental environment to do the bench test, with and without dynamic feedback in place. To simulate an unbalanced server load, one of the servers was given an additional I/O intensive task. The results of the experiment show that the dynamic feedback scheduling does better load balance job than original LVS schedulings.Especially, the system is scalable. Since load balancer gathers the information of back-end servers regularly, if one node failed, all the new tasks can be handled by remaining nodes. So that, as far as the total performance of the system is concerned, single server failure can hardly lead to termination of the whole system, but only a little decrease of its performance.