Panda 2.1 Software Release

Panda 2.1 is the first Panda release specifically designed for people outside the Panda group to use. Compared to Panda 2.0, it offers a cleaner design, simpler interfaces, better documentation, and higher quality code, as well as several new features. Like release 2.0, release 2.1 is written in C++; is designed to support SPMD style application programs on distributed memory parallel architectures, networks of workstations and sequential workstations; and combines the following three techniques to support high performance array I/O operations:

Support for any HPF-style BLOCK and * based data distribution on disk and in memory
High-level interfaces for the application to use to input and output array data
Use of server-directed collective I/O to obtain high utilization of disk bandwidth

New features in Panda 2.1 include:

Part-time I/O processors, which allow a single node (processor) to alternate between running application code and acting as an I/O server. Part-time I/O processors are useful when processors are at a premium.
Shared I/O processors, which allow two jobs to share a set of I/O processors.
A reduced-messages implementation of server-directed I/O, which provides higher performance in cases where Panda's performance is limited by message latency. An implementation of the original server-directed I/O is also included in this release.

Panda 2.1 has been tested on sequential Unix workstations and on an IBM SP2. It currently runs on the UNIX file system using message passing library MPICH, and the NASA AMES IBM SP2's AIX file system using IBM's commercial version of the MPI message passing library. We have tested Panda 2.1 on the following platforms and configurations, using Panda's interface for C++ applications:

Sun Workstation with UNIX File System
HP Workstation with UNIX File System NASA Ames IBM SP2 with AIX 3.2.5 file system
Cornell Theory Center IBM SP2 with AIX 3.2.5 file system

You may obtain more information on Panda 2.1, including documentation and source code:

Panda 2.1 Reference Manual
Panda 2.1 source code (83 KB compressed)
Panda 2.1 Release Notes, including the installation procedure
MS thesis describing internals of Panda 2.1 and performance results

Panda Contributors

We thank our first Panda group member, former graduate student Kent Seamons, for his work as chief architect of Panda versions 1.0 and 2.0. We also thank former graduate students Phyllis Jones, John Jozwiak, and Mahesh Subramaniam (principal developer of release 2.1) for their involvement in the development and performance testing of Panda. We also appreciate the interest and support of Mike Folk's group at NCSA, our main avenue for technology transfer.

Last modified: July 8, 1996