CUDA内存资源池设计避免访问冲突  

Memory Resource Pool Devisal in CUDA to Avoid Accessing Collision

作　　者：于玉龙[1] 王宇新[2] 郭忠明[1] 唐堃[1] 郭禾[1] 

机构地区：[1]大连理工大学软件学院,辽宁大连116621 [2]大连理工大学计算机科学与技术学院,辽宁大连116024

出　　处：《小型微型计算机系统》2013年第11期2583-2587,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61033012;61173179)资助

摘　　要：CUDA内存资源往往存在静态申请与访问的约束,多CPU线程并发执行时会产生资源访问冲突.经建模分析发现,消除资源与设备核函数之间的静态关系可避免该问题.基于资源池模式设计CUDA内存资源池,整合多份资源副本,增加相应调度管理机制,达到了消除资源访问冲突的目的.以纹理内存为案例,在确认纹理内存线程间访问冲突存在及发生频率的基础上,设计并实现纹理内存资源池.大量实验证明资源池不但能够保证程序正确性,避免访问冲突,而且具有较小的性能损耗与负载开销,较低的代码维护成本,同时适用于多GPU系统.There are usually static allocation and accessing restrictions in using the CUDA device memories, which encourage the colli-sion between multiple CPU threads instances when concurrent accessing. A method to avoid this issue is found by modeling and ana- lyzing to decouple the static relationship between the resource and the device kernel. The solution is proposed based on Resource Pool Pattern with multiple resource copies and the management mechanism to remove the accessing collision. Texture memory is selected for case study. The texture pool is devised and implemented after the confirmation of existence and frequency of the texture memory accessing collision between multiple threads. The experimental results illustrate that the resource pool brings accuracy, better code ma-intainability with little overhead on performance and load, and can be used in multi-GPU systems.

关 键 词：CUDA 多线程 纹理内存 访问冲突 内存资源池 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]