摘要
访存带宽是限制众核处理器性能提升的关键,将片上最后一级Cache设计为所有处理器核共享是必要的.在共享Cache中隔离放置冲突的数据,是提高共享Cache性能的关键.文中提出了缓存块链接的硬件方法,用于隔离共享Cache中不同线程之间的数据.文中基于时钟精准的片上众核结构模拟器,使用Splash2程序组和生物信息学中的任务,对所提机制进行了评估.实验结果表明,与传统共享Cache相比,使用缓存块链接机制时,使得共享Cache的冲突性缺失率降低约20%,而使得IPC平均提高了约10%.
Memory bandwidth is critical to overall performance, especially for on-chip many-core architecture. It may be necessary to design a shared last level on-chip cache, to eliminate capacity wasted by multiple copies of one data block in private caches. However, when it comes to on-chip architecture, the conflict in shared cache becomes more serious than traditional single processor architecture. It is crucial to isolate conflicting data blocks in shared cache. This paper proposes a novel hardware approach, that is, block agglutinating, to isolate blocks of different threads in shared cache. Extensive analysis of the proposed scheme with Splash2 benchmarks and Bioinfor- matics workloads is performed using a cycle accurate many-core processor simulator. Experimen- tal results show that when using block agglutinating, it makes an average reduction by about 20% in conflict miss rate of shared cache compared to the traditional shared cache, and it makes IPC improved by about 10%.
出处
《计算机学报》
EI
CSCD
北大核心
2009年第10期1896-1904,共9页
Chinese Journal of Computers
基金
国家自然科学基金重点项目(60736012)
国家"九七三"重点基础研究发展规划项目基金(2005CB321600)
国家"八六三"高技术研究发展计划项目基金(2009AA01Z103)
北京市自然科学基金(4092044)资助
关键词
众核
共享CACHE
数据冲突
资源隔离
容量划分
many-core architecture
shared cache
conflict
resource isolation
cache partition
