Scalable network architectures for providing per-flow service guarantees
Jasleen Sahni, Harrick M. Vin · 2002
Two requirements drive the design of next-generation networks: (1) the need for richer service semantics to support emerging network applications with timeliness constraints; and (2) the need to operate at rapidly-increasing link speeds. Unfortunately, these two requirements are often conflicting. Proposals to provide per-flow service guarantees require the use of complex router mechanisms; whereas the ability to operate at high link speeds mandates router-simplification. The goal of this dissertation is to design network architectures that meet the above requirements of scalability and providing per-flow service guarantees simultaneously. Past efforts design network architectures that are either scalable or rich in their service offerings, but not both. Conventional network architectures use the First-in-First-Out (FIFO) link scheduler in routers which, although scalable, fails to provide service guarantees in the presence of bursty traffic. The Integrated Services (IntServ) network architecture, in contrast, enables a network to provide per-flow service guarantees by requiring all routers to employ sophisticated scheduling algorithms. These scheduling algorithms, however, require routers to perform per-flow packet classification and maintain per-flow scheduling state, which limits their scalability, especially in the core of networks that carry large number of flows. To achieve simultaneously our objectives of scalability and richness, we ask two questions: (1) Can traffic conditioning mechanisms that prevent bursty traffic from entering the network enable FIFO networks to provide per-flow service guarantees? (2) Is it possible to eliminate complexity from IntServ mechanisms while retaining their strong service semantics? In this dissertation, we answer both of these questions. First, we evaluate the effect of constant-bit-rate (CBR) traffic conditioning in FIFO networks. Our results indicate that CBR shaping is effective in providing per-flow guarantees only in environments where the total premium traffic occupies a small fraction of link capacities. Second, we develop a network architecture that provides per-flow service guarantees similar to IntServ networks, but without, requiring per-flow state or per-flow packet classification in the core routers. We do this in two steps: (1) we understand the end-to-end guarantees of core-stateful networks, and (2) we design core-stateless networks that provide similar guarantees. We instantiate router prototypes of our core-stateless architectures, and find that they operate at speeds similar to routers in current FIFO networks.