Research: A guaranteed-no-cells-dropped buffer management scheme with selective blocking for cell-switching networks

• Authors: Qutaiba Razouqi;Tony Lee;Sumit Ghosh
• Venue: Computer Communications
• Year: 1998

Fundamentally, the inclusion of bursty traffic, the desire to achieve a cost-effective network design, i.e. to obtain maximum throughput in a given network while minimizing cell loss, and the limitation of maximum buffer size, derived from basic digital design principles, lends significant importance to the issue of buffer management in cell switching networks including ATM networks. The buffer management literature is rich and includes reports on improved network performance through priority schemes, fixed thresholds, admission control, buffer partitioning, fuzzy thresholds, fuzzy rule-based congestion and admission control, fuzzy policing mechanisms, and a shared memory buffer for all output ports of a switch subject to a delayed pushout scheme. While highly effective, these techniques suffer from two key weaknesses. First, they are susceptible to cell loss owing to buffer overflow. Second, the performance reported for all of these schemes have been obtained for a single switch and, consequently, the results are not necessarily applicable to a realistic cell-switching network consisting of multiple switches. This paper presents a novel scheme-guaranteed-no-cells-dropped (GNCD), that eliminates cell loss owing to buffer overflow. In this approach, the switching node first records the current buffer occupancy and then determines the absolute number of empty slots. It then admits from the input cell-burst from other switches an exact number of cells equal to the number of empty slots. The remainder of the cells are blocked at the sending switch. The computation involved in the decision to admit or refuse entry into the buffer is simple and fast-a key advantage for high-speed networks. Further, buffer overflow is eliminated implying no cell loss-a key potential advantage for specific traffic classes in ATM. Although the fuzzy thresholding-based buffer management scheme holds a significant potential, according to the literature, to effectively reduce cell loss owing to buffer overflow, GNCD achieves in eliminating cell loss owing to buffer overflow completely. Although the fate of a large fraction of the blocked cells, caused by excess traffic, is similar to that in the fuzzy scheme, it is logical to assume that the prevention of cell loss owing to buffer overflow manifests through increased blocking. GNCD includes an effort to reroute the cells, constituting the increased blocking, to their eventual destinations. Clearly, these rerouted cells may incur delays and cause other cells in the network to suffer additional delays. Thus, the absolute current buffer occupancy and the size of the incoming cell-burst constitute the basis for admitting or blocking cells into the buffer. The GNCD scheme is modeled for both: (1) a single, stand-alone switch to facilitate a meaningful comparison with other competing approaches in the literature; and (2) a 50-switch, representative cell-switching network to study its performance under realistic conditions. Both models are simulated, the first on a uniprocessor computer and the second on a testbed consisting of a network of 25 + Pentium workstations under linux, configured as a loosely coupled parallel processor. For the simulation of model 1, a representative traffic stimulus, characterized by a two-state, 'on-off' Markov chain model and exponentially distributed departures, is utilized. A comparative analysis of the results reveal that the GNCD scheme yields a slightly higher throughput than the fuzzy scheme while ensuring zero cell loss owing to buffer overflow, over a range of values of input cell arrival rates. Performance results also indicate that, contrary to the general expectation that large buffers will enhance throughput and mitigate congestion, the throughput and blocking behavior of the GNCD scheme are unaffected by increasing the buffer capacity beyond a 'critical buffer size'. For the simulation of model 2, the input traffic distribution consists of a total of 1.0-1.5 million cells that are asserted into the network for each experiment, with the video traffic synthesized from an actual movie in MPEG-1, the audio traffic obtained using an ON/OFF model utilizing parameters extracted from an actual voice segment, while the generator for the data traffic utilizes a Markov chain model, subject to traffic shaping. The results corroborate those in model 1 in that, even for a large-scale representative cell-switching network, the throughput is higher and blocking is lower under GNCD than for the fuzzy thresholding approach, over a range of input traffic distributions corresponding to throughputs of 73-91%, although the difference in the throughputs and blocking are negligibly small for both low and high values of input cell arrival rates. At either extremes of the input cell arrival rate values, the behaviors of both GNCD and fuzzy thresholding are similar. However, while the cell loss owing to buffer overflow is nil, the average delay incurred by the cells in the network are higher for GNCD than the fuzzy approach, for the same range of input traffic distributions, by a factor ranging from 1.43 for relatively light input traffic density to 1.92 for moderate input traffic to 0.93 for heavy traffic. This paper notes that GNCD incurs less blocking than the fuzzy scheme and that the fuzzy scheme incurs cell loss through buffer overflow, and argues conservatively that the cells lost through buffer overflow in the fuzzy scheme manifest through increased blocking at the sending switches and that effort must be expended to reroute them to their respective destinations. Under these circumstances, the average delay incurred by the cells in the network is higher for GNCD than the fuzzy scheme by a factor ranging from 2.5 for relatively light input traffic density to 5.72 for moderate input traffic to 1.6 for heavy traffic.