Actually, network bandwidth growth has stayed at 3x Moore's law. Also, the gap between your "typical" user and the needs of HPC other "power users" is considerable. I mean, how many "typical" users really need a 2Gigahertz Intel processor? How many wouldn't be able to tell if it was actually a 400MHz cheapo Cyrix processor???
Carriers use SONET and POS. Really, ATM ceased to be relevant quite a while ago. SONET will continue as it offers quite a lot of features that are interesting to telcos and that IP people seem to underestimate.
So there is a WAN-PHY std. for 10GigE which will be interesting for people lighting up their own dark fiber, but I don't expect telcos to be migrating from POS/SONET solutions any time soon.
10GigE is going to be a very competative LAN solution primarily and a WAN solution for some gung-ho dark fiber regional loops.
In a typical application scenario, the clusters would not be of the same size (this was only so due to a limited number of machines for the testbed). Indeed, LBL/NERSC's primary supercomputing resource is a 3k processor 5Teraflop IBM SP2. The Power3 chip that forms the heart of the SP2 is a floating point powerhouse, but is terrible at the integer calculations that dominate visualization processing. The "visualization server/cluster" will likely be a much smaller cluster of machines with much less expensive processors for integer-intensive work.
Also, the comment presupposes that pipelined parallelism is somehow less valuable than time-slicing. An important aspect of the application is that the visualization component of the calculation does *not* run in lockstep with the simulation code. Timeslicing this on the same set of processors in order to maintain responsiveness is far less efficient than feeding the latest data to an independent resource that can react asynchronously to user requests for interactive graphics.
Clearly it is late at night (have a good sleep).
As stated in the article, the application was binary black-hole coalescence (a core Cactus "Application"). Indeed, without the *application* there wouldn't have been any pretty pictures to show.
Indeed, for this demo, two 10GigE channels were bonded together to form a 20Gigabit link. This made it easier to use the line monitoring equipment to determine if the application had really achieved 10Gigabits+ performance.
In the lab, Force10 has demonstrated 280Gigabits/sec throughput, so you can expect by Sept. this demonstration will resurface with 28-bonded 10Gigabit links.
Slashdot Mirror
Actually, network bandwidth growth has stayed at 3x Moore's law. Also, the gap between your "typical" user and the needs of HPC other "power users" is considerable. I mean, how many "typical" users really need a 2Gigahertz Intel processor? How many wouldn't be able to tell if it was actually a 400MHz cheapo Cyrix processor???
Carriers use SONET and POS. Really, ATM ceased to be relevant quite a while ago. SONET will continue as it offers quite a lot of features that are interesting to telcos and that IP people seem to underestimate. So there is a WAN-PHY std. for 10GigE which will be interesting for people lighting up their own dark fiber, but I don't expect telcos to be migrating from POS/SONET solutions any time soon. 10GigE is going to be a very competative LAN solution primarily and a WAN solution for some gung-ho dark fiber regional loops.
Also, the comment presupposes that pipelined parallelism is somehow less valuable than time-slicing. An important aspect of the application is that the visualization component of the calculation does *not* run in lockstep with the simulation code. Timeslicing this on the same set of processors in order to maintain responsiveness is far less efficient than feeding the latest data to an independent resource that can react asynchronously to user requests for interactive graphics.
Clearly it is late at night (have a good sleep). As stated in the article, the application was binary black-hole coalescence (a core Cactus "Application"). Indeed, without the *application* there wouldn't have been any pretty pictures to show.
They'd love to use a SAN. Show me a SAN that can sustain 10Gigabits (even with Bonded 1-GigE) and we'd be happy to use it.
Indeed, for this demo, two 10GigE channels were bonded together to form a 20Gigabit link. This made it easier to use the line monitoring equipment to determine if the application had really achieved 10Gigabits+ performance. In the lab, Force10 has demonstrated 280Gigabits/sec throughput, so you can expect by Sept. this demonstration will resurface with 28-bonded 10Gigabit links.