Resource Allocation with Persistent and Transient Flows

Abstract

Abstract. The flow control algorithms currently used in the Internet have been tailored to share bandwidth between users on the basis of the physical characteristics of the network links they use rather than the characteristics of their applications. This can result in a perception of poor quality of service by some users even when adequate bandwidth is potentially available, and is the motivation for seeking to provide differentiated services. In this paper, stimulated by current discussion on Web mice and elephants, we explore service differentiation between persistent and short-lived flows, and between file transfers of different sizes. In particular, we seek to achieve this using decentralized algorithms that can be implemented by end-systems without requiring the support of a complex network architecture. The algorithms we propose correspond to a form of weighted processor sharing and can be tailored to approximate the shortest remaining processing time service discipline.

Resource allocation between persistent and transient flows

“…For , we have (18) In Fig. 1, we plot the mean utility of a persistent flow and the mean number of transient flows in the system for (i) the static policy, as given by (17), (ii) weighted PS, as given by (18), (iii) naive PS, which gives an equal share of bandwidth to each flow, persistent or transient, and (iv) the optimal policy, obtained numerically using the techniques described in [6], [7]. If all flows use TCP, then case (iii) approximates the bandwidth shares they obtain.…”

“…Now, if the work backlogged at time is no smaller than , and if no flow contributes more than , then the number of backlogged flows must be at least . Thus, it is immediate from (4) and (5) that, given any and , it holds for all sufficiently large that (6) Therefore, by the Borel-Cantelli lemma, infinitely often, i.e., . Since this holds for arbitrarily large , the first claim of the claim is established.…”

“…It is difficult to determine the optimal policy in the general setting considered so far. When file sizes are exponentially distributed, [6] and [7] show how to calculate the optimal policy using value iteration, and describe some structural properties of the optimal policy.…”

Congestion control algorithms for background data transfers of minimal delay impact

“…Interestingly this policy has the same structure as the solution of a different optimization problem considered in [25] (see also Section 1.2), while a packet-level implementation is documented in [27]. The thesis considers a macroscopic implementation of this policy, where the targeted congestion-threshold is assumed to be identical with a targeted throughput per flow.…”

“…Closer to the methodology followed in this thesis is the work in [25]. The main difference is that the designer may control all the types of flows in the traffic mixture (background, persistent and short TCPs) and thus their numbers are considered to be known, a clearly unrealistic assumption for public environments.…”