On Tue, Jul 31, 2012 at 7:17 AM, Matthew Knepley <[email protected]> wrote:

The problem with the programming model they propose, namely unroll the entire task graph and schedule, is that it has been proposed at least 20 times before by smart, well-funded people. These slides just tell us how the system is supposed to work, which is bullshit. They need to tell us how it is different from other similar systems which failed, and why that will make it succeed.

I feel like a lot of those projects don't seem to recognize that the granularity we work with is very small. For strong scaling, we need every grid cell to be processed once every 10-100 microseconds. There is a data dependence and ghost cell cost that needs to be amortized over a large enough patch. The task graph isn't a great model if you have only one task per thread. Also, it tends not to follow data locality well.