PVFS Porting Project

PVFS Porting Project

Background

In recent years the disparity between I/O performance and processor performance has led to I/O bottlenecks in many applications, especially those using large data sets. A popular approach for alleviating this kind of bottleneck is the use of parallel file systems, which can be found on many commercial parallel computers. The goal of the Parallel Virtual File System (PVFS) Project is to explore the design, implementation, and uses of parallel I/O. PVFS serves as both a platform for parallel I/O research as well as a production file system for the cluster computing community. PVFS is currently targeted at Linux clusters of workstations, or Beowulfs.

Feature

Figure 1 shows how nodes might be assigned for use with PVFS. Nodes are divided into compute nodes, on which applications are run, a management node which handles metadata operations, and I/O nodes which store file data for PVFS file systems. The management and I/O nodes might be used for computation as well. It's up to the administrators. 

Figure 2 shows the file access flow of the PVFS file system. To access a file in the parallel file system, we need to install three components in the cluster: mgr, iod and libpvfs. The application on the compute node first acquires the description information of the accessed file from the mgr to find the file size, location and etc. Then it makes connection with the iod daemons and gets the real data from the I/O nodes using the interface provided by libpvfs. 

Porting

In our project, we port the PVFS from Linux to Solaris operating system. The porting is based on PVFS-1.6.3.tar.gz software package. According to the difference between Linux and Solaris 11 operating system, we made slight change in the source code. The code has been compiled and run on the Solaris 11 operating system successfully. Meanwhile, we have acquired 95MB/s aggregate bandwidth on a 5-node Solaris11 based cluster, which is separated into one management node, two I/O nodes and two compute nodes.