Slashdot Mirror
TeraGrid v. Distributed Computing
Nevyan writes "After three years of development and nearly a hundred million dollars the TeraGrid has been running at or above most peoples expectations for such a daunting project. On January 23, 2004 the system came online and provided 4.5 teraflops of computing power to scientists across the country. However, the waiting list for TeraGrid is long, including a bidding process through the National Science Foundations (NSF's) Partnerships for Advanced Computational Infrastructure (PACI) and many scientists with little funding but bright ideas are being left behind. While the list of supercomputer sites and peak power is growing how is the world of Distributed Computing faring? "
Important to remember that the Grid is a _kind_ of distributed computing. But the main thing about The Grid (like The Internet, The Grid is basically TeraGrid in the US + European Data Grid) is that it is suitable for handing off parallel jobs with high intercommunication needs to (i.e. MPI jobs). Not necessarily because these jobs can run across different nodes of the grid (though they can with MPI/Nexus or whatever it's called), but because each "node" in the Grid network is a HUGE MOFO LINUX CLUSTER or similar. The grid gives lots of physicists access to computing resources for parallel processing jobs that would otherwise be sitting idle.
/.ers generally mean by distributed computing is a bit different - most apps there are "embarrassingly parallel" ones you can just farm out. They don't need to chatter to eachother, just process some data and send it back to Central.
What
Google's distributed OS has been discussed a lot on Slashdot, but it is more than just a search algorithm on their own servers:
Google Compute is a feature of the Google Toolbar that enables your computer to help solve challenging scientific problems when it would otherwise be idle. When you enable Google Compute, your computer will download a small piece of a large research project and perform calculations on it that will then be included with the calculations performed by thousands of other computers doing the same thing. This process is known as distributed computing.
The first beneficiary of this effort is Folding@home, a non-profit academic research project at Stanford University that is trying to understand the structure of proteins so they can develop better treatments for a number of illnesses. In the future Google Compute may allow you to also donate your computing time to other carefully selected worthwhile endeavors, including projects to improve Google and its services.
- The Google Compute Project
The BOINC platform (that seti@home is switching over to) has the ability to divide work between project as you suggest. Though I'm not really sure that there are very many other projects running on it.
As noted in earlier comments, the TeraGrid's individual nodes _are_ NUMA clusters. This allows large, non-parallel computations to be run without individual service level agreements, login coordination and scheduling issues gumming up the process. The TeraGrid is an effort to remove the administrative nightmare's keeping most clusters from being fully utilized and most small-time scientists work from being completed.
Care to cite a source?
When you apply to the PACI program you get a grant of Service Units -- i.e. time on the computers. You don't need huge amounts of funding. The requirements state that you need to be a researcher at a U.S. institution. It also helps if you can show that you actually need and can use that kind of computing power.
And, please, distributed computing and supercomputing are not synonymous in terms of what problems they address. Distributed computing cannot replace supercomputers in every case. DC is good for a limited set of problems.
Lastly, an example of Teragrid research: Ketchup on the Grid with Joysticks.
It is based on Apple's XGrid, and uses volunteers from the Mac community here at NCSU, as well as some of the lab macs, and soon we will hopefully have official Linux and Windows clients, maybe even Solaris, to run on more of the computers around campus.
There is even a really nice web interface that shows the active nodes and their status, as well as the aggregate power of the two clusters.
Its really nice, anyone who is part of the grid can just fire up the controller and submit a job, I am a part of the lower power grid since my TiBook is only a 667, but I was able to connect up and do the Mandelbrot Set thing that comes with XGrid at a level equal to around 7 or 8 GHz.
There are some screenshots here
e to the pi i plus one equals zero
This is exactly why there's BOINC, distributed.net, Grid.org, etc. that have multiple projects served to one user-installed program: Instead of projects having to compete for real-time resources, they run work units one after the other, and users can pick and choose which projects are run. This should prevent said burnout for most users.
One project on the Teragrid used 17 Teraflops. The poster hasn't done his/her research on the Teragrid.
I would encourage you to visit www.teragrid.org and read more than the front page to learn what the Teragrid is
There's this one, it's probably what you want:
http://www.stanford.edu/group/pandegroup/folding/
But I'm quite selfish (and actually interested in primes abd or at least know more about them than I do about protiens), and there are entities offering big prize money for big primes, and if one of my machines finds one, I'll get big bucks:
http://mersenne.org
Tag lost or not installed.
It's too bad that whoever modded this Insightful doesn't know much about parallel applications.
DC is fine and very cost-effective for its niche of applications, which is those that are "embarassingly parallel." This is (somewhat circularly) defined as being very easy to parallelize on a DC machine. What characterizes these apps is very low communications between different tasks, which works for DC because the high network latency doesn't get in the way.
I've love to see you try to put Conjugate Gradiant (CG) on a distributed system. It involves large matrix-vector multiplies that inherently require lots of vector fragments passing between the processors. CG is one of the 8 NAS Parallel Benchmarks, and if you look at Beowulf papers that use NAS, you'll see that they often leave out CG because performance is so bad. If it's low on a Beowulf, where the network is presumed to be local and dedicated, it will totally suck on anything with a typical high-latency/low-bandwidth network.
It's an all *nix environment presently totalling around 4200 CPUs of which 96 ( in a single cluster)
is AIX 5.2, 3128 (WOW!!) is on Tru64 (in 2 clusters) and the rest, distributed in 5 clusters
are some form of Linux.
Two of the clusters have a second phase which together will add 316 CPUs on Linux.
As of October 1 of this year, 5 clusters at 3 sites will be added with the OS / CPU breakdown as follows:
Linux : 1800 CPUs in 3 clusters
AIX 5.1 : 320 in 1 cluster
Solaris 9 : 256 in 1 cluster
That's an awful lot of Unix and a buttload of Tru64 and Linux
Pain is merely failure leaving the body