Amarnath Mukherjee

On resource management and QoS guarantees for long range dependent traffic

The effect of long-memory processes on queue length statistics of a single queue system is studied through a controlled fractionally differenced ARIMA (1,d,0) input process. This process has two parameters /spl phi//sub 1/ and d representing an auto-regressive component and a long-range dependent component, respectively. Results show that the queue length statistics studied (mean, variance and the 0.999 quantile) are proportional to e(c/sup c/spl phi/1/) e(c/sub 2/d), where (c/sub 1/, c/sub 2/) are positive constants, and c/sub 2/>c/sub 1/. The effect of the auto-correlation structure on queue length statistics is also dependent on the tail of the distribution of the input process. Therefore, multiplexing gains can be achieved for long-range dependent processes because the resulting process has a tighter tail distribution. Given these observations, and the need to provide guaranteed maximum-delay, delay-jitter bounds, and low cell-loss probabilities, the solution appears to be (a) allocating higher rates if the distribution tails are large, (b) using multiplexing to narrow these tails, (c) applying a per circuit frame-clock in conjunction with an active cell-discard strategy to deal with long-range dependence.

Analysis of dynamic congestion control protocols: a Fokker-Planck approximation

We present an approximate analysis of a queue with dynamically changing input rates that are based on implicit or explicit feedback. This is motivated by recent proposals for adaptive congestion control algorithms [RaJa 88, Jac 88], where the senders window size at the transport level is adjusted based on perceived congestion level of a bottleneck node. We develop an analysis methodology for a simplified system; yet it is powerful enough to answer the important questions regarding stability, convergence (or oscillations), fairness and the significant effect that delayed feedback plays on performance. Specifically, we find that, in the absence of feedback delay, the linear increase/exponential decrease algorithm of Jacobson and Ramakrishnan-Jain [ Jac 88, RaJa 88] is provably stable and fair. Delayed feedback on the other hand, introduces oscillations for every individual user as well as unfairness across those competing for the same resource. While the simulation study of Zhang [Zha 89] and the fluid-approximation study of Bolot and Shanker [BoSh 90] have observed the oscillations in cumulative queue length and measurements by Jacobson [ Jac 88] have revealed some of the unfairness properties, the reasons for these have not been identified. We identify quantitatively the cause of these effects, via-a-vis the systems parameters and properties of the algorithm used. The model presented is fairly general and can be applied to evaluate the performance of a wide range of feedback control schemes. It is an extension of the classical Fokker-Planck equation. Therefore, it addresses traffic viability (to some extent) that fluid approximation techniques do not address. Comments University of Pennsylvania Department of Computer and Information Science Technical Report No. MSCIS-91-18. This technical report is available at ScholarlyCommons: http://repository.upenn.edu/cis_reports/385 Analysis Of Dynamic Congestion Control Protocols: A Fokker-Planck Approximation MS-CIS-91-18 DISTRIBUTED SYSTEMS LAB 5 Amarnath Mukherjee John C. Strikwerda Department of Computer and Information Science School of Engineering and Applied Science University of Pennsylvania Philadelphia, PA 19104-6389

Dynamic Time Windows: packet admission control with feedback

We present a feedback congestion control method, Dynamic Time Windows, for use in high speed wide area networks based on controlling source variance. It is part of the two-level integrated congestion control system introduced in our earlier work[1]. The method consists of a packet admission control system and a feedback system to dynamically control source burstiness. Source throughput is not modulated as with traditional packet windows, allowing system throughput to remain high while avoiding congestion. Furthermore, the admission control bounds congestion times in the network, allowing feedback to be effective in the face of large bandwidth delay products. The basic control mechanisms are analogs to traditional packet windows applied to controlling time windows - a new mechanism which allows switches to modulate source variances. The proposed system is simulated, and the results reported and analyzed. Enhancements to the basic system are also proposed and analyzed. We wish to stress that the system described here is the second level of a two-level congestion control. Previous work[1] concentrated on the switch queueing mechanism, Pulse, while this work is a detailed examination of the feedback system used to adjust time windows to changing network load.

Dynamic time windows and generalized virtual clock: combined closed-loop/open-loop congestion control

The authors present mechanisms for congestion control of data traffic in high-speed wide area networks. The network model assumes reservation of resources based on average requirements. The key ideas involve separation of different sources of network congestion, short-term bursts and medium-term load, and using separate mechanisms to address them. Thus, dynamic time window (DTW) admission control is proposed as a mechanism to limit traffic burstiness from sources as a function of the medium-term load on the system, while a new fairness criterion for short-term congestion (Pulse) is proposed as a mechanism for dealing with fair scheduling by switches of short-term bursts. The model of the network is presented. The DTW and Pulse mechanisms are discussed. A detailed analytical and simulation study is presented for static time-windows and various scheduling algorithms. Preliminary results on DTW are discussed.<<ETX>>

Evaluation of retransmission strategies in a local area network environment

We present an evaluation of retransmission strategies over local area networks. Expressions are derived for the expectation and the variance of the transmission time of the go-back-n and the selective repeat protocols in the presence of errors. These are compared to the expressions for blast with full retransmission on error (BFRE) derived by Zwaenepoel [Zwa 85]. We conclude that go-back-n performs almost as well as selective repeat and is very much simpler to implement while BFRE is stable only for a limited range of messages sizes and error rates. We also present a variant of BFRE which optimally checkpoints the transmission of a large message. This is shown to overcome the instability of ordinary BFRE. It has a simple state machine and seems to take full advantage of the low error rates of local area networks. We further investigate go-back-n by generalizing the analysis to an upper layer transport protocol, which is likely to encounter among other things, variable delays due to protocol overhead, multiple connections, process switches and operating system scheduling priorities.

Simultaneous analysis of flow and error control strategies with congestion-dependent errors

We investigate the performance of flow control and error control protocols and their role in controlling congestion. We address two kinds of packet errors: (a) independent errors and (b) dependent errors. The latter kind are those which are caused by congestion in the system. We consider the go-back-n and the selective repeat protocols for error recovery. The flow control strategy that we study is the sliding-window protocol where we vary the window size as the control parameter. Our performance measure is the expected time and the standard deviation of the time to transmit a large message, consisting of N packets. Previous studies have focussed on either flow control or error control strategies, see for example [Mor 88, MLS 89, TW 79 and Zwa 851. The complexity of analyses has usually precluded simultaneous study of both. One of the main results in this paper shows that under some circumstances the two issues are quasi-independent. For example, to study throughput versus window size for the sliding window flow control protocol using the go-back-n error control strategy, we can study the flow control issue using window models of varying complexity and then combine these results with the term representing the cost of errors due to go-back-n. We next develop a framework to evaluate the two retransmission strategies in presence of windows when packet errors are congestion-dependent. Earlier work on retransmission strategies, for example [MLS 89, TW 79, Zwa 851, have assumed the independence of errors. If the cause of packet errors is random noise in the communications channel, then this is a reasonable assumption. However, in most networks, such random errors are extremely infrequent as compared to packet failures due to lack of availability of buffers because of congestion [Jac 881. This introduces complications in that the premise of independent packet failures is no longer valid. In fact, it is more likely for a failure to occur when one has already occurred than when none has occurred. In our study, we assume an error function p(o), where o is some relevant ‘congestion information’. We then compare the two retransmission strategies for different error functions in the presence of window flow control. In particular, we show when an increase in window size can cause a sharp degradation in performance, and how the two retransmission strategies perform in such a case.

Providing heterogeneous quality of service bounds for correlated video traffic at a multiplexor

Abstract Network support for variable bit-rate video needs to consider (i) properties of workload induced (e.g., significant auto-correlations into far lags and heterogeneous marginal distributions), and (ii) application specific bounds on delay-jitter and statistical cell-loss probabilities. The objective of this paper is to present a quality-of-service solution for such traffic at each multiplexing point in a network. Heterogeneity in both offered workload and quality-of-service requirements are addressed. A per-virtual-circuit framing structure and a pseudo earliest-due-date cell dispatcher are introduced to provide guaranteed delay-jitter bounds. Heterogeneous jitter-bounds are supported through software controlled frame-sizes which may be independently set for each virtual-circuit. The framing structure is a generalization of per-link framing introduced by Golestani. The proposed framing structure eliminates correcting for phase mismatches between incoming frames and outgoing frames, necessary in per-link framing. This results in reduction in end-to-end delay bound and buffer requirements, and a simpler implementation. Strong auto-correlations typically seen in video traffic make equivalent bandwidth computations for heterogeneous cell-loss bounds intractable. To address this, the framing strategy is combined with an active cell-discard mechanism with prioritized cell-dropping, the latter utilizing the history of dropped cells and target cell-loss bounds for each virtual circuit. Upper bounds on the equivalent bandwidth needed to support a given workload with a target quality of service are developed. These are validated through numerical and simulation results from variable bit-rate MPEG-I video traces.