Impact of fairness on Internet performance

Abstract

We discuss the relevance of fairness as a design objective for congestion control mechanisms in the Internet. Specifically, we consider a backbone network shared by a dynamic number of short-lived flows, and study the impact of bandwidth sharing on network performance. In particular, we prove that for a broad class of fair bandwidth allocations, the total number of flows in progress remains finite if the load of every link is less than one. We also show that provided the bandwidth allocation is "sufficiently" fair, performance is optimal in the sense that the throughput of the flows is mainly determined by their access rate. Neither property is guaranteed with unfair bandwidth allocations, when priority is given to one class of flow with respect to another. This suggests current proposals for a differentiated services Internet may lead to suboptimal utilization of network resources.

Stability of size-based scheduling disciplines in resource-sharing networks

“…, L, are necessary conditions for the various nodes i to be stable, and thus for stability of the entire system. It follows from the results in [9] that these conditions are in fact also sufficient for α-fair bandwidth-sharing policies. For conciseness, these conditions will often be referred to as the 'standard' conditions.…”

“…Instead, proper tuning of the parameters of so-called (weighted) α-fair bandwidth-sharing policies as introduced in [26] might provide a more promising approach for improving the delay performance. As is proved in [9], the usual necessary conditions are in fact sufficient for these policies to achieve stability. It is noteworthy that in single-link scenarios α-fair policies essentially reduce to DPS disciplines, which are known to cover the entire achievable mean-delay region in that case [17].…”

“…In DPS, jobs belong to one of several classes, and the total service rate is shared among competing requests in proportion to class-dependent weight factors. There are strong indications that DPS achieves a certain degree of isolation among classes and in particular immunity from highly variable service demands [6,9], in contrast to strict priority disciplines. Besides fairness aspects, a further major issue surrounding size-based scheduling in general and SRPT in particular, is that it relies on knowledge of (remaining) job sizes.…”

“…The performance gains from size-based scheduling disciplines have been thoroughly analyzed in the literature for single-server systems. In practice however, users commonly require service from several shared resources simultaneously, like traffic flows that traverse multiple links in bandwidth-sharing networks as considered in [9,25] for example. While single-server systems provide tractable results and useful insights, they do not exhibit the potential non work-conserving behavior that may arise in systems with concurrent resource possession.…”

Throughputs in processor sharing models for integrated stream and elastic traffic

“…In several papers [11,16,17] the timeaverage throughput, defined as the expected throughput the 'server' provides to an elastic call at an arbitrary (non-idle) time instant, is applied to approximate the call-average throughput. Many other papers use the ratio of the expected transfer volume and the expected sojourn time as an approximation [1,2,3,4,9,31]. In this paper we introduce the expected instantaneous throughput, i.e.…”