Slashdot Mirror
Low Energy Supercomputing
Faith Singer at TACC writes "The term 'supercomputing' usually evokes images of large, expensive computer systems that calculate unfathomable algorithms and run on enough energy to support a small city. Now, imagine a supercomputer, but run on the electrical equivalent of three standard-size coffee-makers. This year's international supercomputing conference, SC10, will feature the Student Cluster Competition that challenges students to build, maintain, and run the most-cutting edge, commercially available high-performance computing (HPC) architectures on just 26 amps."
Amps = current, not energy....
Can I use as many volts as I'd like?
I went to eat some animal crackers and the box said, "Do not eat if seal is broken." I opened the box and sure enough..
I don't know about computers, but you can get a lot of productivity out of humans with the power produced from three coffeemakers.
Gamingmuseum.com: Give your 3D accelerator a rest.
Try Joules (in context as a total), or watts (as a measure per unit time).
"The competition challenges students to build, maintain, and run a cutting edge, commercially available HPC architecture on just 26 amps of energy."
Only problem is that the Ampere is a unit of CURRENT, not energy. It's like saying someone weighs 686 Newtons.
While I understand that if the voltage is kept the same, then the amps are proportional to the energy involved per unit time because W = V x A. However 26 amps at 120 volts for 1 second is not the same energy as 26 amps at 5 million volts for 20 years.
Seven puppies were harmed during the making of this post.
Current * voltage = power. In the United States, alternating current from the wall is nominally 115 volts, and 115 V * 26 A = 2990 W. So I think the actual figure was supposed to be 3 kW of power. Run this for one eight-hour day* for 24 kJ of energy per session.
* This can be business hours (if interactive) or the most efficient cooling hours (if batch).
We call him "Mongo"
Why, without your clothes, you're naked, Miss Dudley!
This doesn't sound too difficult. The number one power-consumer is cooling. Distributing the same code over a larger surface area would allow you to reduce just how sophisticated and power-hungry your cooling needs to be. Any SIMD code will distribute just fine over such an architecture. If you're really clever, you'd design the cluster as a series of pentagons and hexagons, allowing you to build a geodesic. This would not only maximize the surface area but would also minimize the distance network traffic has to travel, networking being the biggest cause of latency in supercomputing. The really really clever geeks would then set up additional "regional" networks to allow for much higher performance when handling code that needed to talk much more locally, then distribute the code according to those regions. (Essentially, you then have a cluster of clusters.)
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
The computational hardware (processors, switch, storage, etc.) must fit into a single rack. All components associated with the system, and access to it, must be powered through the two 120-volt, 20-amp circuits, (each with a soft limit of 13 amps) for a total of 26 amps, provided by the conference. Power to each system will be provided via metered power distribution units The equipment rack must be able to physically hold these metering power strips.
This makes it even harder since theyir hardware has to be power balanced between the two power strips. They'll have to come up with some dynamic load balancing between cluster nodes based on power consumption. I guess dual power supplies might help (do dual power supplies draw perfectly balanced power between both power inputs?), but at a loss of power efficiency.
Having participated in the first of the Student Cluster Challenges at SC07 when I was still in undergrad, I can attest that there's far more to this than what the summary lets on. Not only are you limited to 26 Amps, which is the significant limiting factor, but you're located on the show floor and running your system for 36 hours straight in front of the conference attendees. Moreover, all hardware must be in production and unmodified and fit within a single rack. The Taiwanese team lucked out in this regard as they were using the (then new) 45nm Intel Xeons that were announced the day before the competition started. The only thing you can modify is the code for the programs you have to run (except for the HPC benchmarks).
Some of you might be thinking "pfff...I can stay awake for 36 hours, no problem". That's true, but you're not allowed to be in your booth for more than 12 hours straight and after you leave you must take an 8 hour break. Furthermore, the machines are firewalled from all incoming connections and do not share the same internet connection that the rest of the conference uses.
At SC07, there was a significant power failure on the second day of the competition which brought most teams to their knees. The applications we were running (GAMESS, POP, POV-RAY) are not designed to pick up from a power failure. While the Taiwanese had by far the most powerful system, they couldn't recover from the power failure that had corrupted their SAN in time to win.
To your point, I'm not sure you could get 158 Atoms in a set of off-the-shelf servers that would fit in a single rack to equal a cluster running the latest E series Xeons that perform at top clock but have a lower TDP.
I've always found electric heaters (including geysers, etc. but mostly environmental heating) a huge waste of low entropy. You can achieve the same goal by powering enough chips -- would work especially well for floor heating. Now, if you're not recycling old computers, it might cost some, but if our only constraint is energy, we can thus create a supercomputer that spends 0 energy "for itself", just by installing this system to a few buildings.
You could even communicate through the power line, thus eliminating the need for a separate network installation. "Buy our @home geyser, that pays for itself!", that sort of thing...
...Stony Brook University, and the piece that's missing from the summary is 26 amps@120v, (dual circuits, soft capped at draw of 13 each)
Links to more info from the conference: SC10 CC Page, rules, and app list.
The competition is harder than it sounds, you have to build a cluster from the ground up, fit it into the power requirement (which means stripping out redundancies among other things), strip down a distro (we used Debian as a starting point), get the apps optimized, and then run through the data sets. Your team needs to *understand* the apps, the OS, and the hardware in order to win. There are several people from various teams from past years who have moved on to doing their PhDs in comp sci based on work from this competition (At Carnegie Mellon, MIT, and UMich off the top of my head).
It's important too, in a few ways. For one I know I learned more about clusters the first day I started working on the team for this competition back in 2007 than I ever knew before. That knowledge has led to research fellowships, jobs, and knowledge of what I want to (biochemical modeling). It's an experiance that very few undergrads get, and I think that's a shame.
For the industry it's an important highlight of what can be done with a lot of dedication and a focus on wringing the most from your hardware and software. and in doing that we showcase a lot of work that people dont think about. For example our cluster last year ran off a single disk, plus a large ramdisk as scratch exported over QDR infiniband to the compute nodes. No, it's not new, but it was novel to a lot of people who dropped by our booth.
For another, the ASU team was the first time *I* and many others ever saw a windows cluster in the wild.
Competitions like this are important, they showcase technology and introduce it to undergrads early, with positive benefit!
"goodbye and hello, as always" ~Prince Corwin, from Zelazny's Amber series