IP storage : a performance and security study, LDRD 04-1021.

Van Randwyk, Jamie A. ; Bierbaum, Neal Robert ; Chen, Helen Y. ; Bielecki, Frank T.

Sandia National Laboratories

关键词: 99 General And Miscellaneous//Mathematics, Computing, And Information Science;    Computer Security.;    Internet (Computer Network)-Security Measures.;    Data Transmission;    Memory Devices Internet (Computer Network)-Security Measures.;   

DOI  :  10.2172/921746
RP-ID  :  SAND2005-0698
RP-ID  :  AC04-94AL85000
RP-ID  :  921746

美国|英语

来源: UNT Digital Library

【 摘 要 】

Effective, high-performance, networked file systems and storage is needed to solve I/O bottlenecks between large compute platforms. Frequently, parallel techniques such as PFTP, are employed to overcome the adverse effect of TCP's congestion avoidance algorithm in order to achieve reasonable aggregate throughput. These techniques can suffer from end-system bottlenecks due to the protocol processing overhead and memory copies involved in moving large amounts of data during I/O. Moreover, transferring data using PFTP requires manual operation, lacking the transparency to allow for interactive visualization and computational steering of large-scale simulations from distributed locations. This paper evaluates the emerging Internet SCSI (iSCSI) protocol [2] as the file/data transport in order that remote clients can transparently access data through a distributed global file system available to local clients. We started our work characterizing the performance behavior of iSCSI in Local Area Networks (LANs). We then proceeded to study the effect of propagation delay on throughput using remote iSCSI storage and explored optimization techniques to mitigate the adverse effects of long delay in high-bandwidth Wide Area Networks (WANs). Lastly, we evaluated iSCSI in a Storage Area Network (SAN) for a Global Parallel Filesystem. We conducted our benchmark based on typical usage model of large-scale scientific applications at Sandia. We demonstrated the benefit of high-performance parallel VO to scientific applications at the IEEE 2004 Supercomputing Conference, using experiences and knowledge gained from this study.