Previous | Next --- Slide 8 of 43
Back to Lecture Thumbnails
cwchang

Note on why we choose tile-major: we need to communicate on the boundary, and thus we use tile-major instead of row-major / column-major in order to minimize the number of the elements communicated.

ask

The arithmetic intensity is an indicator of the communication overheads in the given work and data distribution amongst the processors. Greater the arithmetic intensity, lesser is the relative communication overhead.

Metalbird

@cwchang, I do not think that the tile-major as opposed to row-major or column-major should impact the number of elements that we need to communicate across processors, that's determined by how we divide the grid up across the processors. What I think does matter is how efficiently we can use the cache is determined by the type of tile-major compared to row or column major.

fxffx

Small N yields low arithmetic intensity, because when the number of processors are the same, the communication overhead is the same, and smaller N means fewer arithmetic operations. Therefore, it yields low arithmetic intensity.