How does the Ring observe bad latency? If we compare it to crossbars, the two have the same latency cost and the Ring is cheaper to build.
ferozenaina
Why is constant bisection bandwidth an issue? For improved scalability, should it increase or decrease with more nodes?
Ring is bad for scaling as the bandwidth between two nodes is always constant. I'm not very clear on how bisection bandwidth differs from bandwidth.
cuiwei
Bandwidth is a measure of how much data you can transfer from a point A to another point B. However in a network, measuring bandwidth between any 2 specific node does not necessarily reflects how much data the entire network can carry and transmit, thus we need bisection bandwidth. This is analogous to the idea of maxflow-mincut theorem: the amount of flow you can put in the network is restricted by the minimum cut.
wxt
@acortes For two-way rings, it seems that the max latency is always N/2, which is indeed greater than constant since not all nodes are directly connected, but you are right in saying it doesn't seem terrible especially on low-core count processors, especially considering the cost-to-latency ratio.
My question is, how much more complex is a two-way ring than a one-way ring? I would imagine most modern architectures use a two-way ring for latency purposes.
How does the Ring observe bad latency? If we compare it to crossbars, the two have the same latency cost and the Ring is cheaper to build.
Why is constant bisection bandwidth an issue? For improved scalability, should it increase or decrease with more nodes?
Ring is bad for scaling as the bandwidth between two nodes is always constant. I'm not very clear on how bisection bandwidth differs from bandwidth.
Bandwidth is a measure of how much data you can transfer from a point A to another point B. However in a network, measuring bandwidth between any 2 specific node does not necessarily reflects how much data the entire network can carry and transmit, thus we need bisection bandwidth. This is analogous to the idea of maxflow-mincut theorem: the amount of flow you can put in the network is restricted by the minimum cut.
@acortes For two-way rings, it seems that the max latency is always N/2, which is indeed greater than constant since not all nodes are directly connected, but you are right in saying it doesn't seem terrible especially on low-core count processors, especially considering the cost-to-latency ratio.
My question is, how much more complex is a two-way ring than a one-way ring? I would imagine most modern architectures use a two-way ring for latency purposes.