Express Virtual Channel Flow Control

A. Kumar, L.-S. Peh, P. Kundu, and N.K. Jha. "Express Virtual Channels: Towards the Ideal Interconnection Fabric." International Symposium on Computer Architecture (ISCA), June 2007.

This recent paper proposes a technique to bypass some of the router pipeline stages in the traditional multi-stage virtual-channel router used in many on-chip networks. The technique essentially virtualizes express channels by reusing the standard short, local channels to create long, low-latency global channels. This high-level idea has several subtle implementation issues discussed in the paper related to buffering requirements, starvation, and flexible channel management. Students should skim Section 2.1, since we have not discussed these more sophisticated microarchitectural techniques; the key point is simply that state-of-the-art routers can use complex speculation to significantly reduce the zero-load latency through a single router stage. What do students think about the equations presented in Section 2.2 that attempt to capture the latency and energy of an ideal interconnection fabric? What are the inherent trade-offs between static and dynamic express virtual channels? Why does the energy increase with offered load in Figure 4b? Express virtual channels are always given highest priority so that they have guaranteed access to the switch and desired output port. How do the authors ensure that there are never conflicts between flits arriving from two different express virtual channels at the same time? Students should be able to follow Section 4 based on the high-level description of a basic router pipeline presented in lecture. What do students think about the author's claim that their technique reduces the latency of one of the SPLASH benchmarks by over 80%? Consider the author's qualitative comparison of physical express channels to virtual express channels. How does this comparison relate to the more qualitative analysis provided in the related CMPMSI'09 paper by many of the same authors? Students may also be interested in skimming the related MICRO'08 paper on token flow control by the same authors. Token flow control attempts to achieve a similar effect as express virtual channels but with a more general mechanism.