Cycle Speed

  One of the most significant advantages of scheduled routing is that it allows all routing decisions to be made offline, and thus out of the critical path at run time. Consider a typical routing decision made by a dynamic router: it must read a word from another node, examine the header of the node for its destination (and possibly adaptive-routing information), and choose a neighbor node (or virtual channel) to send it to. These decisions, which are in the critical path for message latency, take time. Pipelining techniques can be applied to this decision-making to reduce cycle times and increase bandwidth, but the latency for message transfer remains the same.

Furthermore, as is discussed in Section 1.4, a scheduled router can avoid all round-trip handshakes when doing flow-controlled data transfers. This allows scheduled routing to perform flow control within a single inter-node transfer time. Larger buffers and asynchronous flow-control signals can be used to avoid round-trip handshakes for dynamic routers as well, but this is costly in terms of pin resources and/or buffer management.

In [2], a comparison of dynamic routers is given, breaking down all the components of the cost. A simple deterministic router's cycle time is quoted at 9.8 ns, with a simple planar-adaptive router taking 11.4 ns. In [56], Shoemaker uses these component estimates to derive an estimate for the cycle time of a scheduled router using the same technology. He finds that the NuMesh scheduled router's cycle time is constrained only by the time for a single cross-node transfer, 4.9 ns. This is 50% of the deterministic router's clock period, and 43% of the simple adaptive router's. Improved signaling techniques or pipelining the cross-node transfer itself would decrease this cycle time even further with only minimal effect on the techniques used for scheduled routing.