There's an asymmetry in the average communication cost in this topology because the ones in the center are gonna get information quicker than those at the sides. If we know which ones are towards the center, do you think programs can be written to take advantage of that asymmetry? Or is it just not worth that amount of thought?
lusiliang93
Unlike torus, the link between two nodes are equal, I think the cost between two nodes should be equal as well. I am not sure if the inner nodes connected with 4 adjacent nodes could have less communication cost than corner/edge nodes because inner nodes get more information than edge/corner nodes.
caiqifang
Can someone explain why the latency is O(sqrt(n))?
cloudhary
@caiqifang, it's because the furthest distance to travel from one processor to another might be from the top right to the bottom left, which given this square mesh of total area N, would mean that it has to traverse through two sides of perimeter size sqrt N. Does that makes sense? Perhaps someone else can fact check me on this.
There's an asymmetry in the average communication cost in this topology because the ones in the center are gonna get information quicker than those at the sides. If we know which ones are towards the center, do you think programs can be written to take advantage of that asymmetry? Or is it just not worth that amount of thought?
Unlike torus, the link between two nodes are equal, I think the cost between two nodes should be equal as well. I am not sure if the inner nodes connected with 4 adjacent nodes could have less communication cost than corner/edge nodes because inner nodes get more information than edge/corner nodes.
Can someone explain why the latency is O(sqrt(n))?
@caiqifang, it's because the furthest distance to travel from one processor to another might be from the top right to the bottom left, which given this square mesh of total area N, would mean that it has to traverse through two sides of perimeter size sqrt N. Does that makes sense? Perhaps someone else can fact check me on this.