Admission Control for Statistical QoS: Theory and Practice Edward W. Knightly Ness B. Shroff Rice University Purdue University knightly@ece.rice.edu shroff@ecn.purdue.edu To appear in IEEE Network, March 1999. Abstract In networks that support Quality of Service (QoS), an admission control algorithm determines whether or not a new traffic flow can be admitted to the network such that all users will receive their required performance. Such an algorithm is a key component of future multi-service networks as it determines the extent to which network resources are utilized and whether the promised QoS parameters are actually delivered. Our goals in this paper are threefold. First, we describe and classify a broad set of proposed admission control algorithms. Second, we evaluate the accuracy of these algorithms via experiments using both on-off sources and long traces of compressed video; we compare the admissible regions and QoS parameters predicted by our implementations of the algorithms with those obtained from trace-driven simulations. Finally, we identify the key aspects of an admission control algorithm necessary for achieving a high degree of accuracy and hence a high statistical multiplexing gain. 1 Introduction Provisioning network resources to meet the Quality of Service (QoS) demands of bursty traffic sources is a key issue for future multi-service networks. Such resource provisioning may be realized by an admission control algorithm, which has the function of limiting the number of traffic flows in a class such that the required QoS constraints can be satisfied. The design of admission control algorithms has important consequences for network performance, as an algorithm that unnec- essarily denies access to flows that could have been successfully admitted will under-utilize network resources; similarly, an algorithm that incorrectly admits too many flows will induce QoS violations. Unlike a deterministic service [51], a statistical or soft real-time service associates a small violation probability with delay and throughput bounds, as needed to obtain a utilization gain over a purely worst case approach. Developing re- source allocation schemes for a statistical service has proven particularly challenging due to both the multiple-time-scale characteristics of many multimedia applications, e.g., [19, 34, 52], as well as potential intractabilities arising from complex interactions among traffic flows and the shared multiplexer. Our goals in this paper are threefold. First, we describe a broad set of admission control algorithms from the literature which we divide into the following five classes: (1) tests based on average and peak rate combinatorics [17, 35], (2) tests based on additive effective bandwidths [11, 15, 23, 26], (3) tests based on engineering the “loss curve” [2, 9, 14, 45], (4) tests based on maximum variance approaches [7, 27, 30], and (5) tests based on refinements of effective bandwidths using large deviations theory. Second, we perform a large number of experiments to evaluate the accuracy and effectiveness of these admission control algorithms under realistic workloads, namely, thirty-minute traces of variable-rate MPEG-compressed video and exponen- tial on-off sources commonly used to model voice traffic. To achieve this, we first implement a number of algorithms from the aforementioned classes and determine their respective admissible regions for various traffic mixes and QoS parameters. We then simulate a 45 Mbps multiplexer servicing the same traffic mix, with each flow’s arrival sequence given by either a video trace with a random start time, or an on-off source. For each combination of traffic flows and a particular buffer size, we measure the flows’ resulting performance parameters. By comparing the measured admissible regions with those Excerpts of this paper appear in [29]. The research of E. Knightly is supported by NSF CAREER Award ANI-9733610, NSF Grant ANI-9730104, Nokia Corporation, the Texas Advanced Technology Program, and Texas Instruments. The research of N. Shroff is supported by the NSF CAREER Award NCR-9624525, and NSF Grants ANI-9805441, CDA-9422250, and CDA 96- 17388. 1