With the rapid developments of cluster systems,the problem of I/O bottleneck becomes more and moresevere. A good file system is crucial to the whole systems. Throughput,which means the jobs done per unit time ,is anim...With the rapid developments of cluster systems,the problem of I/O bottleneck becomes more and moresevere. A good file system is crucial to the whole systems. Throughput,which means the jobs done per unit time ,is animportant index of the systems' performance. In this paper,through testing,the authors find that the throughput ofPVFS(Parallel Virtual File Systems),which now is popular in the field of parallel processing,is abnormally low. Thispaper discusses the cause of this phenomenon and improves the throughput of PVFS drastically by eliminating thisbottleneck. The performance before and after this improvement has also been compared.展开更多
At present, there are many effective ways to achieve high performance in cluster system storage management, including server-end disk, server-end caching, local caching and cooperative caching. The cooperative caching...At present, there are many effective ways to achieve high performance in cluster system storage management, including server-end disk, server-end caching, local caching and cooperative caching. The cooperative caching mechanism shares caches among different clients so as to avoid expensive disk access costs and to improve overall throughput of cluster system. In this paper, a Single Copy Cooperative Cache model is proposed together with block lookup algorithm, block replacement algorithm and the consistency algorithm based on the model. Meanwhile, the prototype system of the model is implemented in PVFS file system. Finally, the performance of this system is tested in InfiniBand Framework, the result of which shows that in contrast to the original PVFS system, read performance of PVFS file system is improved by about two times, while write performance is reduced by nearly ten percent.展开更多
文摘With the rapid developments of cluster systems,the problem of I/O bottleneck becomes more and moresevere. A good file system is crucial to the whole systems. Throughput,which means the jobs done per unit time ,is animportant index of the systems' performance. In this paper,through testing,the authors find that the throughput ofPVFS(Parallel Virtual File Systems),which now is popular in the field of parallel processing,is abnormally low. Thispaper discusses the cause of this phenomenon and improves the throughput of PVFS drastically by eliminating thisbottleneck. The performance before and after this improvement has also been compared.
基金This work was supported by the National High Technology Development Program of China under Grant(No.2004AA111110,No.2006AA01A109)
文摘At present, there are many effective ways to achieve high performance in cluster system storage management, including server-end disk, server-end caching, local caching and cooperative caching. The cooperative caching mechanism shares caches among different clients so as to avoid expensive disk access costs and to improve overall throughput of cluster system. In this paper, a Single Copy Cooperative Cache model is proposed together with block lookup algorithm, block replacement algorithm and the consistency algorithm based on the model. Meanwhile, the prototype system of the model is implemented in PVFS file system. Finally, the performance of this system is tested in InfiniBand Framework, the result of which shows that in contrast to the original PVFS system, read performance of PVFS file system is improved by about two times, while write performance is reduced by nearly ten percent.
