freebsd cluster

Building a High-performance Computing Cluster Using FreeBSD
Brooks Davis, Michael AuYeung, Gary Green, Craig Lee
The Aerospace Corporation
El Segundo, CA
{brooks,lee,mauyeung} at aero.org, Gary.B.Green at notes.aero.org© 2003 The Aerospace Corporation
Abstract
In this paper we discuss the design and implementation of Fellowship, a 300+ CPU, general use computing cluster based on FreeBSD. We address the design features including configuration management, network booting of nodes, and scheduling which make this cluster unique and how FreeBSD helped (and hindered) our efforts to make this design a reality.
1 Introduction
For most of the last decade the primary thrust of high performance computing (HPC) development has been in the direction of commodity clusters, commonly known as Beowulf clusters [
Becker
]. These clusters combine commercial off-the-shelf hardware to create systems which rival or exceed the performance of traditional supercomputers in many applications while costing as much as a factor of ten less. Not all applications are suitable for clusters, but a signification portion of interesting scientific applications can be adapted to them.
In 2001, driven by a number of separate users with supercomputing needs, The Aerospace Corporation (a non-profit, federally funded research and development center) decided to build a corporate computing cluster (eventually named Fellowship
1
) as an alternative to continuing to buy small clusters and SMP systems on an ad-hoc basis. This decision was motivated by a desire to use computing resources more efficiently as well as reducing administrative costs. The diverse set of user requirements in our environment led us to a design which differs significantly from most clusters we have seen elsewhere. This is especially true in the areas of operating system choice (FreeBSD) and configuration management (fully network booted nodes).
Fellowship is operational and being used to solve significant real world problems. Our best benchmark run so far has achieved 183 GFlops of floating point performance which would place us in the top 100 on the 2002 TOP500 clusters list.
In this paper, we first give an overview of the cluster's configuration. We cover the basic hardware and software, the physical and logical layout of the systems, and basic operations. Second, we discuss in detail the major design issues we faced when designing the cluster, how we chose to resolve them, and discuss the results of these choices. In this section, we focus particularly on issues related to our use of FreeBSD. Third, we discuss lessons learned as well as lessons we wish the wider parallel computing community would learn. Fourth, we talk about future directions for the community to explore either in incremental improvements or researching new paradigms in cluster computing. Finally, we sum up where we are and where we are going. Table

标签：

版权申明：本站文章部分自网络，如有侵权，请联系：west999com@outlook.com
特别注意：本站所有转载文章言论不代表本站观点，本站所提供的摄影照片，插画，设计作品，如需使用，请与原作者联系，版权归原作者所有