Previous | Next --- Slide 25 of 49
Back to Lecture Thumbnails
yeq

Worst-case latency: 2N

bochet

@yeq worst case 2 * sqrt(N)?

pdp

One potential problem with Mesh topology as discussed in class: Nodes at the ends will have higher latency in comparison to nodes at the center of the network. So, work distribution decisions will play a significant role in such a network.

khans

I didn't understand why Kayvon said the worst-case was 2N. Shouldn't the worst case be 2 sqrt(N) as bochet said? The greatest latency is when a message needs to be sent from opposite corners and the length of the side is sqrt(N).

o_o

Pros: Easy chip layout as all links are equidistant from each other; ability for multiple paths between nodes so there will be fewer blocking messages; good for grids

Cons: Latency between nodes is inconsistent depending on where the nodes are; can be inefficient if work is not scheduled well

sherwood

I agreed that the worst case should be 2sqrt(N)

jedi

Since we consider a mesh as a blocking network, do we basically mean that any network without dedicated end-to-end links between each pair of nodes is blocking (i.e. only crossbar-like networks are non-blocking)?