Dax TCon 2011-07-07
- Super fun times with memory - Bobasaur
No meeting next week. Ken and Utkarsh will be at the SciDAC conference.
Super fun times with memory
Spent most of last week trying to figure out what the performance issues are with dax and how they could be improved.
With current implementation of Dax, with staged evaluation, there is already pretty good use of L1 cache on the card.
The only current problem is that there appear to be more memory access than need be. There are about twice as many in Dax than in a hand tuned version.
Utkarsh is keen to go back to lazy evaluation (a pull model) because it is easier to implement. He poses if when requesting a computed value, it would generate it and then store it somewhere so that the next time it is requested it can be found.
To avoid some confusion, one thing we might try is to not consider a single worklet thread or the entire thing, but rather some friendly block size. Dax could load a block of some size, and given the worklet types would be able to know how much the extent is valid in the output. It could then shift to the next block and so on.