NUMA架构内多个节点间访存延时平衡的内存分配策略

发布时间：2018-03-11 09:21

本文选题：NUMA架构　切入点：内存分配策略　出处：《计算机学报》2017年09期 　论文类型：期刊论文

【摘要】：随着多核架构的发展和普及,NUMA多核架构凭借其本地访存低延时的优势,被各大商业数据中心以及科学计算集群广泛采用.NUMA架构通过增加多个内存控制器,缓解了多核架构下对同一个内存控制器的争用问题,但同时也增加了内存管理的负担.Linux的系统开发者为了实现充分利用NUMA本地访存低延时的特点,在为进程分配内存时,选择进程当前正在运行的NUMA节点作为分配内存的目标节点.这种分配会导致进/线程之间共享内存的不公平.例如,一个在当前本地节点被分配很多内存的进程,可能被调度到远端节点运行,这样会导致进程的性能波动.针对这一问题,该文设计了一种保证NUMA架构内各内存节点间访存延时平衡的内存分配策略,并在Linux系统中实现和验证.延时的获取方法依赖平台,但是系统内核的策略是通用的.实验结果表明,与Linux默认的内存分配策略相比,进/线程间的不公平性平均降低了16%(最多34%),并且各进/线程的性能没有较大抖动.
[Abstract]:With the development of multi-core architecture and the popularization of Numa multi-core architecture, with the advantage of local access and low latency, it has been widely used in various commercial data centers and scientific computing clusters by adding multiple memory controllers. It alleviates the contention for the same memory controller under the multi-core architecture, but also increases the burden of memory management. In order to make full use of the low latency of local memory access of NUMA, the system developer allocates memory for the process. Select the NUMA node currently running by the process as the target node for allocating memory. This allocation can result in an unfair sharing of memory between threads. For example, a process where a large amount of memory is allocated at the current local node, This paper proposes a memory allocation strategy to ensure the balance of memory access delay between memory nodes in NUMA architecture. It is implemented and verified in Linux system. The method of obtaining delay depends on platform, but the strategy of system kernel is universal. The experimental results show that compared with the default memory allocation strategy of Linux, the method of obtaining delay depends on the platform. On average, the inequity between feed / thread reduces 16% (up to 34%), and the performance of each entry / thread has no significant jitter.
【作者单位】：北京航空航天大学计算机学院中德联合软件研究所;
【基金】：国家“八六三”高技术研究发展计划项目基金(2012AA01A302) 国家自然科学基金(61133004,61361126011,61502019,91530324)资助~~
【分类号】：TP333.1

【相似文献】