Speaker
Y. CHENG
(COMPUTING CENTER,INSTITUTE OF HIGH ENERGY PHYSICS,CHINESE ACADEMY OF SCIENCES)
Description
With the development of Linux and improvement of PC's performance, PC cluster used
as high performance computing system is becoming much popular. The performance of
I/O subsystem and cluster file system is critical to a high performance computing
system. In this work the basic characteristics of cluster file systems and their
performance are reviewed. The performance of four distributed cluster file systems,
AFS, NFS, PVFS and CASTOR, were measured. The measurements were carried out on CERN
version RedHat 7.3.3 Linux using standard I/O performance benchmarks. Measurements
show that for single-server single client configuration, NFS, CASTOR and PVFS have
better performance and write rate slightly increases while the record length becomes
larger. CASTOR has the best throughput when the number of write processes increases.
PVFS and CASTOR are tested on multi-server and multi-client system. The two file
systems nicely distribute data I/O to all servers. CASTOR RFIO protocol shows the
best utilization of network bandwidth and optimized to large data size files. CASTOR
also has the better scalability as a cluster file system. Based on the test some
methods are proposed to improve the performance of cluster file system.
Primary authors
D. XU
(COMPUTING CENTER,INSTITUTE OF HIGH ENERGY PHYSICS,CHINESE ACADEMY OF SCIENCES)
G. CHEN
(COMPUTING CENTER,INSTITUTE OF HIGH ENERGY PHYSICS,CHINESE ACADEMY OF SCIENCES)
L. WANG
(COMPUTING CENTER,INSTITUTE OF HIGH ENERGY PHYSICS,CHINESE ACADEMY OF SCIENCES)
W. WU
(COMPUTING CENTER,INSTITUTE OF HIGH ENERGY PHYSICS,CHINESE ACADEMY OF SCIENCES)
Y. CHENG
(COMPUTING CENTER,INSTITUTE OF HIGH ENERGY PHYSICS,CHINESE ACADEMY OF SCIENCES)
