高级检索

    基于冲突检测的高吞吐NVM存储系统

    A High Throughput NVM Storage System Based on Access Request Conflict Detection

    • 摘要: 非易失性存储器(non-volatile memory, NVM)是解决计算机系统存储墙问题的有效手段,但现有的I/O栈缺乏相应的适应和优化机制,特别是基于文件系统的锁机制是影响NVM存储系统性能的重要因素.将存储系统访问请求的管理嵌入到存储设备中,利用NVM存储设备自主管理访问请求,使用基于冲突检测的方法代替基于文件系统的现有锁机制,提高操作系统执行访问请求的并发度,缓解设备接口带来的瓶颈问题.给出了高吞吐NVM存储系统的结构,设计了一种基于二维链表的访问请求管理方法减少访问请求管理中的冲突,一种基于冲突检测的访问请求调度算法用于管理NVM存储设备中共享的访问请求,并给出了新访问请求提交和已有访问请求释放流程,并基于Intel的开源NVM存储设备模拟器PMEM实现了HTPM的原型系统,使用Filebench中的多种负载和Fio测试了HTPM的吞吐率和I/O性能,实验结果表明:HTPM相比PMEM最大能提高31.9%的IOPS值和21.4%的I/O性能.

       

      Abstract: The NVM storage is a useful way to improve the efficiency of storage system in computer. However, there is a lack of adaptation and optimization mechanisms for NVM storage devices in I/O stack of operating system. Especially, the file system-based lock mechanism becomes an important factor affecting the efficiency of NVM storage systems. In this paper, we embed the management function of the access request for storage system in the NVM storage device. In order to improve the access request concurrency of operating system and alleviate the performance bottleneck caused by the device interface, we remove the existing lock mechanism in file system and use the conflict detection algorithm in NVM storage devices. Firstly, the structure of the high-throughput NVM storage system is given. Then, an access request management method based on two-dimensional linked list is design to change the structure of access request management and reduce the conflict of management. And a conflict detection algorithm is design to manage the access requests for the sharing data stored in NVM storage device. The new submission and release process for access requests are designed. Finally, the prototype of high throughput NVM storage system named HTPM is implemented based on PMEM, which is the open source NVM storage device simulator from Intel. The I/O performance and throughput of HTPM are tested by Fio and Filebench. The results show that HTPM can improve the IOPS by 31.9% and the I/O performance by 21.4% compared with PMEM.

       

    /

    返回文章
    返回