100,000 processors is a big-sounding number, but it might be less interesting than the glue holding them together in this case. The dedicated 10Gbps pipe allows much more interesting problems to be solved than a traditional grid computer (unless the grid computer uses P2P).
One of the biggest problems that we always face with big grid computers is that we either need computations that can be broken up into very small parts and have very high compute/io ratios - or else we send bigger chunks of data with VERY long computations to the grid nodes.
> The dedicated 10Gbps pipe allows much more interesting problems to be solved than a traditional grid computer (unless the grid computer uses P2P).
Particularly as this article implies CERN is connected by dedicated 10Gbps links to sites in Canada, France, Germany, Italy, Netherlands, Norway, Spain, Taipei, UK and the USA.
There has never been a peer-reviewed comparison between the much hyped GRID ruled by CERN, and say more discrete distributed supercomputers, like the Storm or Kraken in terms of reliability, scalability (do you get spam?) configurability (how quickly does your new address get spam?). It would pail in comparison to the half dozen underground grids, not to mention a fair comparison with EC2, GOOG back-end etc.
Actually, most experiments, doing "serious" work opted to have their own farms, which they can handle as they please (configuration, versioning etc. -- most of the grid is still struggling with RHEL3 and gcc3, most of what sane people cannot put up with). Parallelization amounts to breaking up data to chunks and running the same dll in multiple instances -- no notion of threading, let alone locking (no dependencies).
One of the biggest problems that we always face with big grid computers is that we either need computations that can be broken up into very small parts and have very high compute/io ratios - or else we send bigger chunks of data with VERY long computations to the grid nodes.