摘要：

普通PC机上搭载的GPU显卡存在功耗低、性能差的问题,不能满足大规模统一计算设备架构(CUDA)并行计算的需求。为此,将计算密集型任务转移到GPU集群中完成,设计并实现基于B/S模式的GPU集群管理系统,用户通过网页提交CUDA代码即可得到GPU集群计算后的结果。测试结果表明,该系统可使用户在任何能使用浏览器的电子设备上完成基于CUDA的计算密集型任务,不仅方便用户使用,而且能加速程序的运行过程,提高工作效率。

关键词: GPU集群, 统一计算设备架构, B/S模式, 资源管理, 任务调度

Abstract:

Due to the low power and poor performance on an ordinary PC equipped with the graphics card,it can not meet the needs of large-scale Compute Unified Device Architecture(CUDA) parallel computing.Aiming at these problems,the compute-intensive user tasks are transferred to the GPU cluster system,the GPU cluster management system is designed and implemented based on B/S mode.The user submits the CUDA code through the web and gets results from the GPU cluster management system.Test result shows that the compute-intensive tasks can be done on any browser-based electronic devices.It brings convenience for the users,accelerates the process of running the program,saves the users’ time and greatly improves the users’ efficiency.

Key words: GPU cluster, Compute Unified Device Architecture(CUDA), B/S mode, resource management, task scheduling

中图分类号: 

• TP311