一种透明访问远程内存系统的设计和实现
发布时间：2018-10-30 18:25
【摘要】：随着计算机技术的不断发展，CPU和系统主存的性能得到了极大的提高，但依然不能满足应用程序日益增长的内存需求，特别是I/O设备的发展相对滞后，磁盘性能成为影响计算机整体性能的重要瓶颈。 为了提高局域网中内存利用率和系统I/O性能，利用高速局域网中高带宽、低延迟的传输特点，在现有研究基础上，本文通过扩充系统交换区容量，提高交换区换页速度，设计和实现了I/O密集型应用和内存密集型应用透明访问远程内存系统(NS-DSM)。该系统实现了在局域网范围内对远程内存的透明共享。本文主要工作有以下几点： 首先，为了能够访问其他节点的空闲内存，本文在客户端节点创建网络块设备来接收响应访问远程内存的操作，，然后通过基数树结构查找、插入、删除操作把系统内存页映射到远程内存，从而实现对远程内存的访问。 其次，NS-DSM系统通过创建和激活交换空间的方法把网络块设备挂载到客户端节点，从而使应用程序能够透明的访问远程内存。为了使内存页能够在最大传输单元为1500KB的以太网中分片、重组后快速传输，系统通过实现轻量级的传输协议——访问网络内存协议(ANMP)来对需要传送的页面进行分片、重组以及数据包传输。 最后，本文设计和实现了客户端-服务端服务加载和转换策略。该策略根据系统内存使用情况的不同而加载不同的状态，并在系统运行时定期检测结点的空闲内存并做相应的服务转换。该策略使得系统中各结点尽可能的负载均衡，从而使整个系统处于一个资源利用的“平衡”状态。 NS-DSM系统实现了局域网内内存的透明共享，并且不用对系统内核作任何修改，也不需要任何特殊的API或者函数库。它对应用程序完全透明，同时不用进行繁杂的数据一致性检验和完整性检验。实验表明，对于大量磁盘访问和内存密集型的应用，该系统能有效地提高系统整体性能和大型应用程序的执行速度。
[Abstract]:With the continuous development of computer technology, the performance of CPU and system main memory has been greatly improved, but still can not meet the growing memory requirements of applications, especially the development of I / O devices is lagging behind. Disk performance has become an important bottleneck affecting the overall performance of computers. In order to improve the utilization ratio of memory and the performance of system I / O in LAN, taking advantage of the transmission characteristics of high bandwidth and low delay in high speed local area network, based on the existing research, this paper expands the capacity of system switching area and improves the speed of switching area. I / O-intensive applications and memory-intensive applications are designed and implemented to transparently access remote memory systems (NS-DSM). The system realizes the transparent sharing of remote memory in LAN. The main work of this paper is as follows: first, in order to be able to access the free memory of other nodes, this paper creates a network block device in the client node to receive the operation of accessing remote memory in response, and then looks it up through the cardinality tree structure. Insert, delete operations map system memory pages to remote memory, thus enabling access to remote memory. Secondly, the NS-DSM system mounts the network block device to the client node by creating and activating the swap space, so that the application can access the remote memory transparently. In order to divide the memory pages into pieces in Ethernet, where the maximum transport unit is 1500KB, and transfer quickly after reorganization, the system implements a lightweight transport protocol, (ANMP), to slice the pages that need to be transmitted. Reorganization and packet transmission. Finally, this paper designs and implements the client-server service loading and transformation strategy. The strategy loads different states according to the system memory usage and periodically detects the free memory of the node and makes the corresponding service transformation while the system is running. The strategy makes every node in the system load balance as much as possible, thus making the whole system in a "balanced" state of resource utilization. The NS-DSM system realizes the transparent sharing of memory in the LAN, and does not need any modification to the kernel of the system, nor any special API or function library. It is completely transparent to the application without complicated data consistency checking and integrity checking. Experiments show that the system can effectively improve the overall performance of the system and the execution speed of large applications for a large number of disk access and memory intensive applications.
【学位授予单位】：湖南大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP333.1
本文编号：2300807
[Abstract]:With the continuous development of computer technology, the performance of CPU and system main memory has been greatly improved, but still can not meet the growing memory requirements of applications, especially the development of I / O devices is lagging behind. Disk performance has become an important bottleneck affecting the overall performance of computers. In order to improve the utilization ratio of memory and the performance of system I / O in LAN, taking advantage of the transmission characteristics of high bandwidth and low delay in high speed local area network, based on the existing research, this paper expands the capacity of system switching area and improves the speed of switching area. I / O-intensive applications and memory-intensive applications are designed and implemented to transparently access remote memory systems (NS-DSM). The system realizes the transparent sharing of remote memory in LAN. The main work of this paper is as follows: first, in order to be able to access the free memory of other nodes, this paper creates a network block device in the client node to receive the operation of accessing remote memory in response, and then looks it up through the cardinality tree structure. Insert, delete operations map system memory pages to remote memory, thus enabling access to remote memory. Secondly, the NS-DSM system mounts the network block device to the client node by creating and activating the swap space, so that the application can access the remote memory transparently. In order to divide the memory pages into pieces in Ethernet, where the maximum transport unit is 1500KB, and transfer quickly after reorganization, the system implements a lightweight transport protocol, (ANMP), to slice the pages that need to be transmitted. Reorganization and packet transmission. Finally, this paper designs and implements the client-server service loading and transformation strategy. The strategy loads different states according to the system memory usage and periodically detects the free memory of the node and makes the corresponding service transformation while the system is running. The strategy makes every node in the system load balance as much as possible, thus making the whole system in a "balanced" state of resource utilization. The NS-DSM system realizes the transparent sharing of memory in the LAN, and does not need any modification to the kernel of the system, nor any special API or function library. It is completely transparent to the application without complicated data consistency checking and integrity checking. Experiments show that the system can effectively improve the overall performance of the system and the execution speed of large applications for a large number of disk access and memory intensive applications.
【学位授予单位】：湖南大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP333.1
【参考文献】
相关期刊论文 前1条
1 褚瑞;肖侬;卢锡城;;一种基于内存服务的内存共享网格系统[J];计算机学报;2006年07期
相关硕士学位论文 前1条
1 陈丹丹;用软件实现局域网中的内存共享[D];南京航空航天大学;2002年
本文编号：2300807
本文链接：https://www.wllwen.com/kejilunwen/jisuanjikexuelunwen/2300807.html
