Forgot your password?

typodupeerror
Supercomputing The Internet

BOINC Exceeds 2 Petaflop/s Barrier 114

Posted by kdawson
from the faster-faster-for-science dept.
Myrrh writes "Though an official announcement has not yet been made, it would appear that the BOINC project as a whole has exceeded two petaflop/s performance. The top page features this legend: '24-hour average: 2,793.53 TeraFLOPS.' According to last month's Top500 list of supercomputers, BOINC's performance is now beating that of the fastest supercomputer, RoadRunner, by more than a factor of two (with the caveat that BOINC has not been benchmarked on Linpack)."
This discussion has been archived. No new comments can be posted.

BOINC Exceeds 2 Petaflop/s Barrier

Comments Filter:
  • Nothing. BOINC requires no CO2 to operate.

    It could just as easily be run on computers powered by nuclear or solar power, producing no CO2 (past initial construction).

    Why does CO2 have to be the end-all-be-all of everything? Why not ask how much coal dust or mercury is now in the atmosphere thanks to the plants that power most of those computers.

  • Since they know what CPUs are running on every BOINC client and the thermal power of them are generally known, it should be possible to calculate...

    That only counts CPU usage. It doesn't count I/O, which would at least include memory I/O, disk I/O, network I/O.

  • by Chabo (880571) on Tuesday July 14 2009, @03:19PM (#28694859) Homepage Journal

    I can guarantee that several orders of magnitude more kWh are consumed by computers that are needlessly on and idle.

    Running BOINC on a computer that's sitting idle helps improve its energy efficiency. It may be consuming electricity, but at least then it's doing something.

  • I'm aware. My point is that I'm tired of "but how much CO2 does it generate?" being tacked on to everything because it's the current fad question.

    The coming ice-age was a science disaster fad. So was the coming overpopulation and world famine. And the ozone holes that would cause everyone to get skin cancer. And....

    There are more important questions. Much of this energy would be used anyway, but it would be in centralized supercomputers. This way though it's cheaper for the scientists so we can get more research done, even though it's slightly less efficient.

    I'm just really tired about CO2 being discussed attached to everything. "Should I buy new shoes?" "Well, the CO2 produced from rubber is... and.... but...".

  • by Sycraft-fu (314770) on Tuesday July 14 2009, @03:51PM (#28695187)

    This is NOT a supercomputer. This is a cluster, and a very slow cluster at that. It seems like people think that anything fast is a "supercomputer" and as techies, we ought to know better.

    What makes a supercomputer "super" is its internode communication. You have extremely fast links so that, in theory, any node can access the memory from any other node as it would its own local memory. Now in reality there are some performance penalties, but still. Basically you really have created one large computer, rather than tons of small ones.

    This is a cluster, which is as the name implies just a bunch of little computers networked in some fashion working on the same problem. That's great, but not the same thing. The nodes do not have high speed communication, some may even be on modems and only connected occasionally.

    Now, why does this matter? Well it depends on the problem you are trying to solve. Some problems need very little communication. A good example would be cracking cryptography. You just divide up the keyspace among all your nodes. There's also very little data to send back and forth. You send you the problem, consisting of the encrypted message to the nodes, and then all the communication from this on is:

    Node: Didn't find the key.
    Controller: Ok try this range.
    Node: Ok.

    As such link speed of the cluster can be very slow. Well other problems still work in a clustered environment, but need higher link speeds like gig Ethernet. 3D rendering would be an example. All the nodes can act independent, they are just divided up on frames to render, or parts of a frame or whatever. However since the problem and results are much larger in this case, they need faster communication to make it practical. A modem won't cut it for transferring images that are 50MB each when you are rendering thousands.

    However, there are other problems where there is heavy inter node communication. A particle simulation would be like this. Since what happens with one particle affects all others, nodes have to chat continuously. For this, you need a supercomputer. The bandwidth of links must be extremely high and the latency must be extremely low, or else processor power will be wasted just waiting on getting the data that is needed.

    So just because something has a lot of CPUs and can crunch a lot of numbers, doesn't make it a supercomputer.

  • by Dirtside (91468) on Tuesday July 14 2009, @04:03PM (#28695365) Journal

    So just because something has a lot of CPUs and can crunch a lot of numbers, doesn't make it a supercomputer.

    There's no reason "supercomputer" needs to only refer to monolithic machines with high-speed interprocess communication, merely because it has primarily meant that in the past.

  • 82% solution (Score:2, Insightful)

    by symbolset (646467) on Tuesday July 14 2009, @04:16PM (#28695551) Homepage Journal
    Strangely enough 410 of the supercomputers listed in the top500, or 82% are of architecture type "cluster".
  • by dtfusion (658871) on Tuesday July 14 2009, @04:28PM (#28695715)
    Well, you're both wrong. You can't simply redefine the terms to win your argument and the term supercomputer doesn't necessarily refer to computer cores networked by a high speed interconnect. Come to think of it, the original post is absurd, because there is no way BOINC could run LINPACK which is the measure of the TOP500 rankings anyway. LINPACK stresses communication performance as well as scalar processor performance. BOINC would probably be slower than my desktop for that purpose. Like a lot of these silly comparisons on slashdot ("My hammer is better than your screwdriver!") it comes down to using the right tool for the right job. If you're not in a hurry and your job is "a bag of jobs" type problem, use BOINC or some other distributed/cloud computing approach. If you're trying to solve some type of PDE use a purpose built system like Roadrunner. Using something like Roadrunner for the type of jobs that BOINC is good at is just a waste of resources - those networks aren't cheap.
  • Barrier? (Score:4, Insightful)

    by necro81 (917438) on Tuesday July 14 2009, @05:27PM (#28696601) Journal
    Was there some sort of fundamental, theoretical limit that could have made getting to 2 petaflop difficult or impossible? Did a graph of BOINC computer power vs time ramp up from zero, stall around 2 PFLOP, and only now punch through? Did the administrators have to come up with some sort of breakthrough or new insight to reach this mark? Two PFLOP is just a round number - is it really any different from 1.9 or 2.1?

    I think not: 2 petaflops is just a matter of recruiting enough computers and having them running BOINC at the same time. If it has achieved this mark, then it couldn't have been that much of a barrier, could it?

A CONS is an object which cares. -- Bernie Greenberg.

Working...