It was 23 years ago this November when a small NASA project, whimsically called “Beowulf,” was initiated to find a new way of achieving order of magnitude improvement in performance to cost. The goal was to provide a sustained performance greater than 1 Gigaflops on real applications for under $50,000, in order to allow single computational scientists access to dedicated data analysis resources. There were several opportunities, including reduction of contention for and reduced latency of mass storage of scientific data and its usage.
The Beowulf project, funded under the NASA HPCC ESS program and guided by the Goddard Space Flight Center, undertook this path after a year of exploring alternative methods available through commercial products. The idea of clusters was hardly new and extended back almost a decade. But the formula of hardware and software synthesized to develop low-cost scalable computing from consumer-grade COTS products was unique in its form and valuable in its accessibility, flexibility of configuration, low cost, and ultimately its empowerment of many contributors (within the US and internationally).
I did the math that demonstrated the opportunity and viable trajectory. And, yes, I named the project and then hired the team. Today, more than 80 percent of the systems of the Top500 list are commodity clusters, and more than 95 percent of all supercomputers on that list employ one or another variation of the Linux operating system, which we introduced to the HPC community through the Beowulf project. Every year there are student cluster competitions at the SC and international ISC conferences, as well as many others specific to different nations.
But back in 2002, I made a mistake. Beowulf clusters had become an effective means of low cost medium-scale computing in academia, industry, and national labs, such that vendors started repackaging nodes for rack-mounting, changing their new COTS computing nodes to better serve the computing markets. Networking evolved to serve commodity clusters as well. Beowulf-class systems no longer had to be assembled from the scrap heap of whatever was available — they could now be assembled from subsystems designed for the purpose of building Beowulf-class systems.
Indeed, system integrators such as Dell, Penguin, and others began shipping fully integrated commodity clusters to end-user data centers with software already installed. In an unthinking moment when talking with someone from the press, I quipped that “Beowulf is dead.” Of course this showed up in one or more journals (a lesson there, I think) and I was forced to respond to this article in a keynote address I was about to give at ISC in Heidelberg.
I explained that Beowulf was not dead but in chrysalis, transforming to a new class of user-driven clusters in partnership with vendors in synergy. But what I had not understood, even then, was that metamorphosis was part of Beowulf, not just one point of singularity. Beowulf is never the same, it keeps changing — and for the better. Today we see emerging an entirely new form of Beowulf with the same cultural and performance opportunities as the original generation two dozen years ago, but with entirely new capabilities for students, experimentalists, and OEMs.
A plethora of very inexpensive processing cards are now available from multiple sources and can be acquired for ridiculously little money and assembled with, if anything, greater flexibility than ever before. Examples include the Nvidia Jetson TK-1, the Raspberry PI-3, and the Digilent Zed Boards, all of which are small, single PCB with SOCs, including, but not limited to, multi-core ARM processors.
The Jetson includes the TEGRA K1 GPU; the Zed Board has a Zilinx Aynq SoC with FPGA. And the Raspberry PI-3 at $40 boasts a 64-bit ARM with Fast Ethernet, 1 GigaByte of DRAM, USB and HDMI interfaces, and other elements. These readily available components are firing up the imagination of many and expanding the role of Beowulf clusters from just number and data crunching to embedded parallel systems, and even robotics.
Once again, this Monday, November 14th starting at 9 pm, the Beowulf Bash will be thrown by gracious sponsors, this time at the Discovery Gateway Children’s Museum. There is no other event like it at SC, and I am sure that it will attract many new and younger attendees. But this is not just a party of older nostalgia-seeking gray-beards reminiscing about Don Becker’s cool Ethernet Driver software of the mid-90s or the other events that launched this new era. Rather, I expect it will be filled with people talking about Parallella, Raspberry, Jetson, and Zed, and how new thinkers are clustering these in a new generation of Beowulf clusters for innovative purposes.
I look forward to this year’s Beowulf Bash and hearing about what you are doing. I was wrong before. Beowulf will never die, and I hope the Bash lives on as well.