Xiao Zhong

Correlation Based File Prefetching Approach for Hadoop

Hadoop Distributed File System (HDFS) has been widely adopted to support Internet applications because of its reliable, scalable and low-cost storage capability. Blue Sky, one of the most popular e-Learning resource sharing systems in China, is utilizing HDFS to store massive courseware. However, due to the inefficient access mechanism of HDFS, access latency of reading files from HDFS significantly impacts the performance of processing user requests. This paper introduces a two-level correlation based file prefetching approach, taking the characteristics of HDFS into consideration, to improve performance by reducing access latency. Four placement patterns to store prefetched data are presented, with policies to achieve trade-off between performance and efficiency of HDFS prefetching. Moreover, a dynamic replica selection algorithm is investigated to improve the efficiency of HDFS prefetching. The proposed prefetching approach has been implemented in Blue Sky, and experimental results prove that correlation based file prefetching can significantly reduce access latency therefore improve performance of Hadoop-based Internet applications.

Performance Implications of Next-Generation Multi-processing Platforms on e-Business Server Applications

When running multiple e-Business server applications simultaneously on the same hardware, inappropriate CPU sharing may endanger the performance stability for individual applications. Robustness and manageability are critical for Java server applications on emerging multi-processing hardware platforms. This paper investigates the performance implications of multiprocessing (including SMP and CMP) systems on highly parallel Java server applications. It also evaluates the feasibility and benefits of partitioning the processor cores among multiple applications. By experimenting the JGF benchmark, a comparison is provided to the synchronization behaviors in different primitives, on different hardware configurations. Cache contention, thread migration are all factors to be considered when perform thread affinity setting. The feasibility of enhancing dependability of each server application is experimented with a pseudoJBB benchmark. The performance evaluation and analysis show that it is practical of avoiding performance interferences among highly parallel, CPU-intensive server applications by restricting thread migration. Effective performance isolation can be achieved along with performance gain.