[Beowulf] Roadrunner shutdown
prentice.bisbal at rutgers.edu
Thu Apr 4 06:22:42 PDT 2013
On 04/03/2013 11:42 AM, Geoffrey Jacobs wrote:
> On 04/03/2013 08:27 AM, Bret Stouder wrote:
>> For the money you could buy a lot more real computational cycles that last a lot longer than anything proprietary from IBM. The reason the Origins have been out of production for a long time is the result of research moving to a commodity server platform. Taken a look at what Cray offers lately? Looks a lot like commodity. If you are going to take the time to port code, it seems that a reasonable argument could be made to stay away from Blue Gene and port to Cuda or Phi.
Actually, that's a naive argument. The Blue Gene architecture uses
standard MPI and support C, C++, Fortran, and Python, so any MPI
compliant program written in any of those languages, which will also run
just fine on a Blue Gene, with no porting necessary - just compile the
code for the BG processors using GCC or the IBM XL compilers. IBM
provides the ESSL and MASSV libaries that are optimized for the Blue
Gene, but those are specified at link time, so I wouldn't really call
that porting. Sure, there are tweaks you could use to really optimize
your code for BG, but just doing the above will get you very far.
In contrast, Cuda requires adding Cuda specific code to move data into
and out of the device, and the sections to be run on the cuda processors
needs to be rewritten in terms of cuda thread index, etc.
For a large code, this could take a significant amount of labor, enough
that it's conceivable in certain cases buying a Blue Gene could be
quicker and more cost-effective than spending all the money on the labor
to port your code to Cuda, test, optimize, test...
In short, it's a lot more effort to port MPI-compliant code for Cuda
than for a Blue Gene.
I specifically didn't mention Phi because I do not know the details of
porting/optimizing code for Phi.
More information about the Beowulf