I am having some trouble understanding memory-constrained scaling -
1) is N(P^0.5) x N(P^0.5) gird size and arbitrary selection?
2) Does this also mean that there are a total of (N^2)*P elements in the grid?
3) Does memory constrained mean N can't increase because the available memory remains the same? Considering that execution time is independent of N, I get this idea. However, how does this make it different from problem-constrained scaling?
fire
how do the implications came out? Is it because MC scaling does not have issues of increasing "comm-to-comp" ratio, TC scaling has slight increase on Comm-to-comp ratio and PC has the biggest increase on comm-to-comp as we scale up P?
flxpox
Why is the communication per processor O(1/P^ 1/6)? So elements per processor is K/P^ 1/2 = (NP^ 1/3) / (P^ 1/2) = N/(P^ 1/6). And the communication needed to be done is proportional to the size of edge, which should be O(P^ 1/12)?
I am having some trouble understanding memory-constrained scaling -
1) is N(P^0.5) x N(P^0.5) gird size and arbitrary selection?
2) Does this also mean that there are a total of (N^2)*P elements in the grid?
3) Does memory constrained mean N can't increase because the available memory remains the same? Considering that execution time is independent of N, I get this idea. However, how does this make it different from problem-constrained scaling?
how do the implications came out? Is it because MC scaling does not have issues of increasing "comm-to-comp" ratio, TC scaling has slight increase on Comm-to-comp ratio and PC has the biggest increase on comm-to-comp as we scale up P?
Why is the communication per processor O(1/P^ 1/6)? So elements per processor is K/P^ 1/2 = (NP^ 1/3) / (P^ 1/2) = N/(P^ 1/6). And the communication needed to be done is proportional to the size of edge, which should be O(P^ 1/12)?