It may be putting two hot topics togeher (MapReduce and GPGPU), but they sure don't make it easy. One of the strong sells of the MapReduce feature was how it reduced the amount of code something as complex as PageRank was written in and made most programs fairly straitforward.
There is a lot of challenges of programming CUDA (disclosure: this is what my ongoing MS thesis work is on), but their code examples show they add no abstraction layer and because of the generality they force on their kernels (all must have the same signature) will be near impossible to reach full GPU performance.
tl;dr Unlike MapReduce, this does not make programming GPUs easier or accessing their hardware potential easier.
Wow, totally read the title as GNU Map Reduce. This is an Nvidia CUDA implementation of Map Reduce, hence the GPU, not a project by the FSF for GNU or anything.
the text in section 4.2 is kind of greyed (hatched) out (and unreadable). very odd. i'm using okular on opensuse / kde - is it odd an acrobat reader too?
There is a lot of challenges of programming CUDA (disclosure: this is what my ongoing MS thesis work is on), but their code examples show they add no abstraction layer and because of the generality they force on their kernels (all must have the same signature) will be near impossible to reach full GPU performance.
tl;dr Unlike MapReduce, this does not make programming GPUs easier or accessing their hardware potential easier.