Is it possible to have a hybrid direct/indirect network? Ie, one where some endpoints are internal and others are external?
I suppose those are probably uncommon, practically speaking. Indirect networks have switches that do nothing but switch, while direct networks have switches that are tied to endpoints. Mixing and matching the two types might be more effort than using all of one or the otter type. But I'm completely making this up.
This comment was marked helpful 0 times.
pwei
http://web.mit.edu/6.173/www/currentsemester/handouts/L14-interconnect-1.pdf. Both direct and indirect networks have their advantages as well as their disadvantages. From my understanding, direct networks are good in that it is difficult to remove any node from the graph simply by breaking a connection, as there are many connections emanating from every node. In the event that the node itself fails, only one node is removed from the graph, and all other endpoints are still connected to the graph through another path. The disadvantage is that latency may not be quick for endpoints on opposite ends of the graph. Indirect networks are good in that they are better at relaying data, since the latency grows logarithmically rather than linearly for direct networks. However, having a node fail could remove many endpoints from the graph (single point of failure). Probably to have advantages of both and reduce disadvantages of either, one can create a hybrid network that is more efficient than either a fully direct or fully indirect graph.
This comment was marked helpful 0 times.
bwasti
@tcz @pwei
I am not sure what you guys mean by hybrid networks, but if you are talking about adding an indirect network layer atop the direct network layer, I'd agree it would work well. However, I think the problem is a hardware/price situation, where redundancies are expensive. This paper shows that added routes to a direct network (Torus routing) can bring down network latency significantly http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6182372&tag=1. Quite clearly this would be extremely expensive to implement (direct networks themselves are really only used in supercomputers and high end routers, apparently).
If you guys were talking about removing the directness from the network or adding routes the the indirect network proposed without achieving a full direct network, I think you'd only be creating a network specific to certain situations. It might seem reasonable to say you are simply taking the best of both worlds, but you are also eliminating some of the best of both of those types of networks: it being no longer direct and there being no longer a guaranteed logarithmic time connection.
Is it possible to have a hybrid direct/indirect network? Ie, one where some endpoints are internal and others are external?
I suppose those are probably uncommon, practically speaking. Indirect networks have switches that do nothing but switch, while direct networks have switches that are tied to endpoints. Mixing and matching the two types might be more effort than using all of one or the otter type. But I'm completely making this up.
This comment was marked helpful 0 times.
http://web.mit.edu/6.173/www/currentsemester/handouts/L14-interconnect-1.pdf. Both direct and indirect networks have their advantages as well as their disadvantages. From my understanding, direct networks are good in that it is difficult to remove any node from the graph simply by breaking a connection, as there are many connections emanating from every node. In the event that the node itself fails, only one node is removed from the graph, and all other endpoints are still connected to the graph through another path. The disadvantage is that latency may not be quick for endpoints on opposite ends of the graph. Indirect networks are good in that they are better at relaying data, since the latency grows logarithmically rather than linearly for direct networks. However, having a node fail could remove many endpoints from the graph (single point of failure). Probably to have advantages of both and reduce disadvantages of either, one can create a hybrid network that is more efficient than either a fully direct or fully indirect graph.
This comment was marked helpful 0 times.
@tcz @pwei
I am not sure what you guys mean by hybrid networks, but if you are talking about adding an indirect network layer atop the direct network layer, I'd agree it would work well. However, I think the problem is a hardware/price situation, where redundancies are expensive. This paper shows that added routes to a direct network (Torus routing) can bring down network latency significantly http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6182372&tag=1. Quite clearly this would be extremely expensive to implement (direct networks themselves are really only used in supercomputers and high end routers, apparently).
If you guys were talking about removing the directness from the network or adding routes the the indirect network proposed without achieving a full direct network, I think you'd only be creating a network specific to certain situations. It might seem reasonable to say you are simply taking the best of both worlds, but you are also eliminating some of the best of both of those types of networks: it being no longer direct and there being no longer a guaranteed logarithmic time connection.
This comment was marked helpful 0 times.