Linux is a flexible, robust node operating system for Beowulf computational clusters. Stability and adaptability set it apart from the legacy operating systems that dominate desktop environments. While not a "cancer" like some detractors have labeled Linux, it has spread quickly from its humble beginnings as a student's hobby project to a full-featured server operating system with advanced features and legendary stability. And while almost any Linux distribution will perform adequately as a Beowulf node operating system, a little tuning and trimming will skinny down the already lean Linux kernel, leaving more compute resources for scientific applications. If this chapter seems a little overwhelming, we note that there are companies that will completely configure and deliver Beowulf systems, including all the aforementioned tweaks and modifications to the kernel. There are also revolutionary systems such as the Beowulf software from Scyld Computing Corporation (www.sycld.com). The software from Scyld combines a custom Linux kernel and distribution with a complete environment for submitting jobs and administering the cluster. With its extremely simple single-system image approach to management, the Scyld software can make Beowulfs very easy indeed. Chapter 18 is devoted to a discussion of the Scyld approach.
One final reminder is in order. Many Beowulf builders became acquainted with Linux purely out of necessity. They started constructing their Beowulf saying, "Every OS is pretty much like every other, and Linux is free... free is good, right?". On the back of restaurant napkins, they sketched out their improved price/performance ratios. After the hardware arrived, the obligatory LINPACK report was sent to the Top500 list, and the real scientific application ran endlessly on the new Beowulf. Then it happened. Scientists using Linux purely as a tool stopped and peered inquisitively at the tool. They read the source code for the kernel. Suddenly, the simulation of the impending collision of the Andromeda galaxy with our own Milky Way seemed less interesting. Even though the two galaxies are closing at a rate of 300,000 miles per hour and we have only 5 billion years to wait, the simulation simply seemed less exciting than improving the virtual memory paging algorithm in the kernel source, sending Linus Torvalds the patch, and reading all the kernel mailing list traffic. Beware. Even the shortest of peeks down the rabbit's hole can sometimes lead to a wonderland much more interesting than your own.