科研之友
微信
新浪微博
Facebook
分享链接
摘要
Load balance is critical for performance in parallel file systems. In those systems, a file is commonly placed in several storage nodes using replication or striping techniques for simultaneous file I/O operations. However, storage nodes may be removed or added as the result of node failure or upgrade. Files may also dynamically change with creating, deleting, and appending operations. All of these may result in load imbalance which means some storage nodes will be overloaded through receiving too many requests. Consequently, some nodes will become bottlenecks and degrade overall performance. In this study, we propose a novel load balancing algorithm for parallel file systems, called probability-based load balance (PBLB). The PBLB algorithm aims to achieve the optimal load balance with minimum communication overhead. We implement the PBLB algorithm in the pNFS file system and evaluate it with IOzone benchmark. Experimental results show that the PBLB algorithm is effective and stable, and it outperforms the random algorithm which is a classical load balancing algorithm by 16.2% in load balance and 16.3% in performance. Besides, the experimental results also show that the PBLB algorithm is able to deliver high scalability.
- 出版日期2015-9
- 单位华中科技大学; 中国电子科技集团公司第二十八研究所
