Bridging the Gap between Big Data System Software Stack and Appl

来自台湾成功大学的 Hung-chang Hsiao 带来的有关 HBase 应用于半导体晶圆制造行业的例子。他们提供了一种融合多种不同存储系统的方案,解决了小文件问题,以及设计一个统一的协议使不同的存储系统之间和兼容以及透明传输,此外他们还设计和实现了一个负载均衡系统,并发表在 IEEE 上。



2. Bridging the Gap between Big Data System Software Stack and Applications: The Case of Distributed Storage Service for Semiconductor Wafer Fabrication Foundries Huan-Ping Su (蘇桓平), Yi-Sheng Lien (連奕盛) National Cheng Kung University

3.Agenda •Introduction •Background •Goal •Design •Performance •Summary

4.Intro In the semiconductor manufacturing industry, data volume increases exponentially during the manufacturing process, which greatly helps in monitoring and improving production quality.

5.Background •Heterogeneous storages (such as FTP, SQL Server, HDFS, HBASE...) •Data transfer between storages (moving, coping, ETL...) •Learning curve for the sophisticated storage

6.Goal •For easy administration between storages •Compatible with underlying storages (We can communicate with different storages in one protocol ) •Intuitive operation


8.Design •HDFS Interface

9.Design •HttpServer (Usage Pattern) http://<hds_host>/access?from=smb://user/

10.Design •Transparency by Mixing HDFS and HBase

11.Design •Compliance with HDFS Interfaces, and thus Hadoop Ecosystem

12.Design •Load Balancing

13.Design •Load Balancing

14.Experimental Setup •Server Spec CPU : Intel Xeon E7-8850 @2GHz (80 cores) Mem : 512GB Disk : 750GB * 16 •Virtual Machine Spec (16 nodes) CPU : 80 * 2GHz Mem : 32GB Disk : 750GB

15.Experimental Setup •Cluster Settings Hadoop 2.6.0-cdh5.10.0 HBase 1.2.0-cdh5.10.0 ZooKeeper 3.4.5-cdh5.10.0 Yarn 2.6.0-cdh5.10.0 Hive 1.1.0-cdh5.10.0

16.Performance Results (Transparency) Hive and Spark are studied over our HDS •Hive (r,s) : read small files with hive (SELECT queries over 30000 files, each with 0.001 MBytes) •Hive (r,l) : read large files with hive (SELECT queries over 12 files, each with 16000 MBytes) •Hive (w) : write with hive (Hive consequently generates one 32 GBytes file)

17.Performance Results (Transparency)

18.Performance Results (Load balancing)

19.Performance Results (Overheads) Overhead by HDS, compared with the native HDFS. The workloads are represented by (1,10000), (100,100), (1000,10), and (10000,1), where (x,y) denotes the replication of y files of each in x MBytes from an FTP server to the HDS cluster.

20.Performance Results (Overheads)

21.Summary •Solve small file problem in HDFS •Transparency to different storages •Compatible with hadoop-eco project •Improve 1% yield rate of semiconductor manufacturing