Beowulf Is Dead?

Relax. It is just a catchy headline to draw your attention to some classic insights into the HPC cluster market.

True confession time. I stole that headline from Beowulf pioneer Thomas Sterling. As Thomas had originally intended, I am sure it sparked some reaction. Of course, sensational headlines generate pageviews, but they also help drive home important points. As a writer, there is no better reward than hungry readers devouring every word in search of some truth or insight. The only problem is it is not my headline nor my insight. Indeed, this headline servers to introduce a new feature here at Linux Magazine.

As some of you may know, at one point I was editor of ClusterWorld Magazine. One of my goals with the magazine was to get the movers and shakers of the growing HPC Cluster community/market to share their wisdom and experience with the readership. To accomplish this, I reserved the last page of every issue for what I called the Head Node column and asked an community/industry luminary to comment on an important topic. To my delight, the back page was a huge success and managed to pull some very fine words out out of some very bright people.

Recently, I was reviewing some of these columns and I thought, much of this “stuff” is still relevant and probably would be of interest to the current crop of Linux Magazine HPC readers (and to those ClusterWorld readers that were not paying attention back in the day). Then, in a flash of insight, I thought “Why not post the Head Node columns on the web as they were only published in the paper version!” Yes, I can still amaze myself with obvious conclusions.

Continuing with my category 1 brainstorm, I thought, “How about if I post the columns and provide some commentary on the opinions, ideas, and predictions shared by the writers”. Great idea, I told myself. It has been over 5 years since the columns were originally published and the community/market has changed quite a bit. The past columns will certainly make for good discussions and I can posting these every few weeks as part of my weekly column (i.e. when I cannot think of anything else to write about). It is, therefore, with great pride, I present the first installment in the new HPC Master Series: Beowulf in Chrysalis By Thomas Sterling.

Go ahead read it. I’ll wait here until your are done. Good stuff don’t you think? After I read it, I wonder if Thomas could have seen the eventual disruption cause by Beowulf. Back when he wrote the column (Fall 2003) clusters composed 42% of the Top500 List. Today’s Top500 List (Fall of 2008) shows clusters hitting the 82% mark. The entire HPC industry has been taken over by clusters. In the early days, anyone in the trenches could have told you that clusters would dominate, those on higher ground were not so sure. Of course the people in the trenches had their sleeves rolled up and were pushing at the edges of the technology. As Sterling states, we were witnessing the transition from Do-It-Yourself with Off-The-Shelf parts not intended for HPC to a thriving HPC cluster market complete with cluster specific hardware and software companies. Perhaps a more fitting metric is how HPC is fairing in the the current economy. The HPC market is expected to see modest declines, (5.4% in 2009) and rebound to a projected $11.7 Billion market with a healthy growth rate of 9.6% by 2012. (Courtesy of the recent IDC Predictions 2009 Webinar). Nothing dead or dying about that prediction.

In my opinion, the robustness of HPC due to the fact that it delivers results. Many companies now consider HPC essential to their competitiveness and when these benefits are coupled with a low cost barrier to entry, the decision to spend money becomes not one of convenience but of necessity.

Beowulf, or what I like to call the concept and practice of employing low-cost off-the-shelf commodity hardware to solve big (HPC) problems, has certainly emerged from the chrysalis. It has enabled a demand for HPC products and fueled the growth of what was once a somewhat staid market. And, perhaps most importantly it has put HPC in the hands of almost anyone. Alive and well seems more like it, but that is a rather boring headline.

Comments on "Beowulf Is Dead?"


All this sounds so good. When are we going to see a good series of articles detailing the construction and provisioning of a “commodity” cluster so that the rest of us can build one too?

There was this nice article a couple of years back on the “Limulus” project http://limulus.basement-supercomputing.com/ that gave some details on the construction of a commodity cluster. The only mention of software in the article pretty much said that we would have to wait until they figured it out. We never heard any more about it and the web site doesn’t give us any clue either.


an updated howto would be nice


We have since 2001 installed approx. 16 HPC Linux clusters in our company. Build inhouse but with prof. server hardware. Some with infiniband ddr but most with std. GigE. Looking at 10GigE but it seems like IB are the best choice for latency aware applications.

Using Panasas Direct Flow on all clients for the large HPC environment. Could not survive without it.

HPC Linux clusters are a the most important tool for our core business regarding depth imaging etc. Could not have done it without
HPC. Thx to Linux and the Beowulf concept we are more than capable to do the work.

Biggest issues now are power, cooling and space.

Beowulf is absolutely alive and kicking. Totally agree with the article writer on that.


One good howto is worth a thousand general stories.


With the 4 GPGPU you can get 1TFLOP (double floats) or 4 TFLOPS (single floats) for scientific computing on a PC with 4 graphic cards for well under 10K. Well I guess you could cluster Playstation 3 (cell) for about a quarter of the FLOPS for about what 1200 bucks?


Bait and Switch headlines betray the reader’s confidence, causes their eyes to glaze over and move quickly to another webizine.


The Limulus project is still moving forward albeit slowly. Since the article was written, there has been a lot of work on packaging. Some of those involved, including me, believe that there needs to be a low cost, high performance, and power efficient “personal HPC workstation” reference hardware platform before we can develop a portable software platform. The good news is we have a working solution that will provide 8-16 cores using four motherboards, use a single power supply, actively manage power usage, and run quietly next to a desk. Once the hardware is in place, then the software will soon follow. There should be an announcement in the May/June time frame. One thing to remember this is not a “cluster solution” but rather a high performance workstation.


Indeed. The problem is, however, the array of hardware and software has become so vast and fluid that trying to capture it all would require a good sized book. Thus the need for a hardware reference (see above) where basic concepts can be understood and enhanced.


Darn. You got me.


Now that s interesting … Cluster including the cell processor (PS 3 ) .. The one thing that strikes me immediately is how can one classify the workload and process them accordingly ? say more graphics intensive by the Cell Proc …

Note : I am very new to this field ..pl correct if am thinking the other way ..


I recently noticed this blog.

I think that the very first Beowulf mini supercomputer was built by Max Gilliland of Denelco of Denver, which was the Heterogeneous Element Processor (HEP) with PDP/11′s 50 CPUs stacked up in a metal box. This was built in early 1970s. (I mention of this fact though I recognize Tom Sterling’s first naming of Beowulf computer in 2001 .)

BTW, Max designed the world largest hybrid computer at Beckman Instrument Co., a decade before of this, which was used by Boeing engineers to design space shuttle, and by M.I.T. scientist to simulate Armstrong’s lunar landing, and by me at Mobil Oil for extraction of Shale oil from Shale rock out of Rocky mountains.

I introduced this technology to NEC, which then produced The Earth Simulator with US$350 million and four tennis court size air conditioned room. This was to simulate environment of the entire earth with the use of real-life climate data from satellites and ocean buoys. Japanese scientists have already completed a forecast of global ocean temperatures for the next 50 years, and a full set of climate predictions was ready by the end of 2002. “Soon, instead of speculating about the possible environmental impact of, say, the Kyoto accord, policymakers will be able to plug its parameters into the virtual Earth, then skip ahead 1,000 years to get a handle on what effect those policies might have. That kind of concrete data could revolutionize environmental science. By digitally cloning the Earth, we might just be able to save it.” (TIME.com, “Best Inventions, 2002”) This Earth Simulator was once the world fastest supercomputer, as the TIME magazine hailed.

Alas, a recent Japanese newspaper reported that NEC decided to terminate their supercomputer business. Therefore, I was very glad to read your article saying that the Beowulf mini supercomputer has now 82% market share. This in turn may mean that “Super computer in a single metal box is DEAD!!”

When the HEP concept was initiated in early 1970s, I had another idea of spreading CPUs around the world and interconnecting them via data telecom network (i.e., ARPANET at that time, which can be Internet nowadays) — see and .

Best, Tak Utsumi

Mr.Utsumi ppl like u made our life easy and happy. Thanks for your great work. You leave a mark on humanity through your work :).

Appreciate you sharing, great blog.Really thank you! Much obliged.

You could certainly see your expertise in the paintings you write. The arena hopes for even more passionate writers like you who are not afraid to say how they believe. At all times follow your heart.

I just want to tell you that I’m newbie to blogs and honestly savored you’re web site. Almost certainly I’m likely to bookmark your site . You definitely have awesome stories. Bless you for sharing your web site.

I just want to say I am just new to blogging and honestly savored your blog site. Most likely I’m likely to bookmark your blog post . You amazingly have really good posts. Kudos for revealing your web page.

Leave a Reply