Research And Implementation Of Task Scheduling Mechanism On A Parallel Computing System

The compute-intensive computing tasks have the following characteristics:1. large-scale,long computation time;2. Different types of computing tasks, its computational complexity isnot the same, To deal with these computational tasks required for system reliability, faulttolerance and scalability. On the other hand, in the cluster system, there can be many differenttypes of nodes, thus, the system should be able to organize the compute nodes to collaborativecomplete computing tasks.Nowadays, the Map/Reduce or MPI is commonly used in the parallel computing system,but they can not solve the heterogeneous nodes, dynamic load balancing, node dropped andother issues.This paper built up a dedicated GPU cluster system to solve the enormous amount ofcomputation for compute-intensive applications. The system consists of three parts:management node, scheduler node and the compute nodes. The management node is mainlyresponsible for the management and monitoring of tasks. Scheduling node is responsible fortask scheduling and distribution of computing task. Computing node processes tasks with thepowerful computing capabilities of the GPU. This paper mainly studies computing powerallocation algorithm of multi-task, compute node capacity estimation algorithm, taskassignment algorithm of the compute node and the re-calculation method of the missed task,so that to achieve dynamic load balancing for the cluster, the fault tolerance and scalability arealso guaranteed.Finally, this paper tests the availability and performance of our system in the large-scalecluster and small-scale cluster environment. The results show that, in the large-scale cluster,the efficiency of our system can reach more than90%, besides, the system can effectivelybalance the load and can support heterogeneous computing nodes.