Linux Project Report

[attachment=16414]

1 Project Overview
The goal of the Linux Project at LLNL is to field production Linux clusters providing Livermore
Computing (LC) users with the Livermore Model programming environment, described in
Sec. 1.1. In the past, this was accomplished using variants of the UNIX operating system on
proprietary hardware. By using Linux on near-commodity hardware, we believed that we could
procure more cost-effective clusters, offer better support to our users, and create a platform from
which open source system software could be developed jointly with strategic partners and
leveraged by the high-performance computing (HPC) community at large


The Livermore Model
The LC scalable system strategy, known as the Livermore Model (Seager, 2002) and depicted in
Fig. 1, is to provide a common application environment on all LC production clusters. This
allows highly complex scientific simulation applications to be portable across generations of
hardware architectures and between currently available operating systems. The strategy has
provided a stable applications development environment since about 1992, when the Meiko CS-2
MPP was introduced at LLNL.


Choosing Linux
In the past, the Livermore Model was implemented on proprietary UNIX platforms including
Solaris, AIX, and Tru64. While this approach was successful, it had a few disadvantages. First,
highly specialized proprietary hardware is expensive to purchase and maintain. Second, the HPC
market has dwindled in comparison to the commodity market in recent years, making it
increasingly difficult to convince proprietary vendors to implement HPC-specific features


Scalable Parallel File System
As part of the Compaq partnership just described, Livermore and Compaq ported the
Petal/Frangipani (Lee, 1996 and Thekkath, 1997) research file system to Linux and improved
performance and stability, utilizing QsNet for transporting file system information at high
bandwidth. As performance and stability drew near targets, this project was aborted because of
business changes within Compaq and an end to the Livermore partnership


Cluster Administration Tools
Specialized tools are required to effectively administer large clusters of nodes. As the node count
rises, it becomes increasingly advantageous to treat a set of nodes as a logical unit and to act on
those nodes in parallel. Although open source solutions exist for some problems, they generally
do not scale to the size of the clusters we deploy. We have therefore developed several tools to
augment those already available in the open source community.