Slashdot Mirror
Europe Plans Exascale Funding Above U.S. Levels
dcblogs writes "The European Commission last week said it is doubling its multi-year investment in the push for exascale computing from €630 million to €1.2 billion (or the equivalent of $1.58 billion). They are making this a priority even as austerity measures are imposed to prevent defaults. China, meanwhile, has a five-year plan to deliver exascale computing between 2016-20 (PDF). The Europeans announced the plan the same week the White House released its fiscal year 2013 budget, which envisions a third year of anemic funding to develop exascale technologies. Last year, the U.S. Department of Energy science budget asked for nearly $91 million in funding for the efforts in the current fiscal year; it received $73.4 million. DOE science is trying for about $90 million for exascale for 2013. There's more funding tucked in military and security budgets. The U.S. wants exascale around 2018, but it has yet to deliver a plan or the money for it."
I didn't know what it was, I don't follow supercomputing very closely. I looked it up. From http://en.wikipedia.org/wiki/Exascale_computing
"Exascale computing refers to computing capabilities beyond the currently existing petascale. If achieved, it would represent a thousandfold increase over that scale."
To define Petascale:
"In computing, petascale refers to a computer system capable of reaching performance in excess of one petaflops, i.e. one quadrillion floating point operations per second." http://en.wikipedia.org/wiki/Petascale
A Petascale computer, the Cray XT5 Jaguar can do 1.75 petaflops. To reach an exaflop, it would require almost 6000 installations of this supercomputer.
So yeah, Exaflop is pretty big. http://en.wikipedia.org/wiki/Orders_of_magnitude_(computing)
Nobodies Prefect
Tidbits for Techs Technology Blog
One big reason why an exa-scale installation is generally better than an exa-scale distributed project is that of Data Transfer.
Distributed computing is plagued by Data Transfer bottlenecks. If it's an internet project, the cumulative effect of combined bandwidth does add up. But serving out project segments at exa-scale levels is very expensive, and equally expensive receiving the solution chunks. There's also the problem of "internet climatology" (I'm not sure what it's really called) where the connections aren't uniform. While the internet does "self-heal" it takes time, and that adds up as well.
Basically, when you scale up the computing power on a distributed project, the problems scale too. Out of order processing of problem chunks also causes problems when peers join and drop out in unpredictable ways. Often the same chunk has to spend many times more cycles than actually required, due to peers getting bored with the work, or just testing out the system and dropping the piece they're working on.
An exa-scale supercomputer would remove the problem of collaboration overhead, or at least significantly reduce it. Scheduling is much more efficient, and in the end FLOPS doesn't measure performance in any reliable qualitative way. A distributed project can run at an exaFLOPS rate and still do no productive work, if the participants never finish any of the work they are tasked with.
Hectice, baby, Mercator says hello to you
Some examples of wasteful, government-enforced research: man in space, GPS, the Internet.