A Cluster Load Balancing System Based On The Total Processes Remaining Time

Cluster computing is always the focus of a great deal of research in the computer field. Cluster system can not only make full use of computing resources, but also gain high performance in computer system with lower hardware and software cost. With the rapid development of micro-processor and high performance network, cluster computing gradually has becoming a parallel/distributed computing resources.Load balancing is important technology in the cluster system. It improves the system performance by balancing the load between nodes. The previous research proved that load balancing could greatly improve the cluster systems' performance. Load balancing algorithms can be generally characterized as dynamic and static. Static load balancing algorithms use only static information when distributing load to nodes. It can be easily analyzed in theory and easily designed in implementation. The static load balancing algorithms do not consider current load status of nodes when making decisions of distributing load. That leads to worse performance because these algorithms can't make full use of cluster resources. Dynamic load balancing make distributing load decisions according to the system-state information. So it can get better performance.The Linux Cluster Load Balancing System is a dynamic load balancing system using a new load index in the resources heterogeneous cluster environment. The new load index is the total processes remaining time in CPU queue. It takes into account not only the status resources but also the received tasks info. So it presents cluster system processing ability and the load being deal with well. Otherwise, in the heterogeneous cluster system, considering CPU and memory factors when deciding the threshold of each node. The load balancing system was managed centrally. All nodes collect load information and report it to the manage node periodically. The manage node choose the receiver using random selecting algorithms. Preemptive process migration is adopted to balance the load in system dynamically. The system is based on Linux operating system, mainly use Linux kernel programming and network programming in user level. Although the system design deals with operating system kernel—process operation, the software is easy to setup and upgrade because it take use of module programming rather than modify the kernel itself. Experimental results show that the new load index and the new way to consider threshold decrease the executing time of tasks and improve the cluster's performance. The new way to decided threshold improves cluster performance by 26% compared to using same threshold in system. Using new load index improved system performance approaching 30% compared to using length of CPU queue as load index.