Parallel Processing with the Little Giant

During the fall semester of 1998 we built a 16 processor " Beowulf " parallel supercomputer for carrying out molecular modeling studies of biological membranes. In the fall of 1999 and the summers of 2001 and 2003 it was expanded to its current size of > 100 processors. These simulations represent each atom in the system as a particle, providing extremely high spatial and temporal resolution for the study of molecular structure and dynamics. Unfortunately, such a detailed model requires large amounts of computer time (typically we follow 10⁴ atoms for 10¹⁰ time steps). In the past, the solution to this problem was to obtain very fast (and very expensive) mainframes or high-end workstations from companies such as Cray, Silicon Graphics, Sun, etc. Recent advances in the performance of commodity-type computers (e.g., Intel Pentium processor-based) and fast ethernet networking equipment have led to a more economical alternative. Using a cluster of Pentium-based computers, connected by a high speed network and running Linux (a type of Unix operating system), a large computational task is divided among the processors. Parallel computers employing this new strategy have been constructed at several national laboratories (e.g., Los Alamos and NIH), research universities (the nearest example is Purdue), and a small but growing number of corporations.

At Wabash we originally combined standard desktop computers from Dell Corporation through a 100 Mbit/s full-duplex ethernet switch from Cisco Systems (more recently we have employed myrinet connections to allow greater scalability). Funds for this project were provided by Wabash College and a National Science Foundation Research at Undergraduate Institutions grant to Scott Feller.

We have benchmarked the Little Giant and found that sixteen processors achieve greater than 1 GFLOP (1 billion floating point operations per second) performance on the standard NAS parallel benchmark with near linear scaling (see graph). Using the parallel linback benchmark that forms the basis for the top500 supercomputer ranking, 5.7 GFLOP performance was obtained with 32 processors. During the summer of 99, we installed compilers for the fortran77, fortran90, C, and C++ programming languages. We have also installed libraries of subroutines for parallel programming and scientific computing. Applications programs for carrying out simulations using classical or quantum mechanical methods are also running on the Little Giant. Several students are using this facility in conjunction with independent research projects under the supervision of professor Feller. Additionally, a course in parallel programming was offered by computer science professor David Maharry where students used the Little Giant to run parallel programs that they wrote as class assignments.

This system is based on work by colleagues at the National Institute of Health. For more information see www.lobos.nih.gov