Cisneros and Brackett’s Architecture

OPTOELECTRONIC PACKET SWITCHES 287 phase, the output ports read the tokens and tune their receivers to the appropriate input port wavelengths. Then, the cell transmission starts over the star transport network. The transmission and control cycles are over- lapped in time in order to increase throughput. Since each input has a unique wavelength and there is input᎐output port pair scheduling prior to transmis- sion, cells are transmitted simultaneously without causing contention. This architecture allows multicasting. However, the throughput of the switch may degrade as the number of multicasting connections increases, due to output port collisions in the first phase. It is shown that this problem can Ž be alleviated by call splitting i.e., allowing a multicast call to be completed in . multiple cell slots . This architecture can support different priority levels for the cell by adding minor tracks into the control network. However, in that case, the token should recirculate among the input ports more than once, depending on the number of priority levels. This will increase the length of the write phase and result in longer cell processing time. The main drawback of the switch is the sequential processing of the token by input and output ports. As a result, the time taken for the token to travel through the entire ring increases as the size of the switch increases. In the case of multiple priority levels, the recirculation period for the token be- comes even longer. Here, HOL blocking is another factor that degrades throughput.

11.2.3 Cisneros and Brackett’s Architecture

w x Cisneros and Brackett 17 proposed a large ATM switch architecture that is based on memory switch modules and optical star couplers. The architecture consists of input modules, output modules, optical star couplers, and a Ž . contention resolution de®ice CRD . Input and output modules are based on electronic shared memories. The architecture requires optical-to-electronic and electronic-to-optical conversions in some stages. Each output module has an associated unique wavelength. As shown in Figure 11.6, the input ports and output ports are put in groups of size n, and each group is connected to n = m or m = n memory switches, respectively. The interconnection be- tween the input and output modules is achieved by k optical star couplers. There are k tunable laser transmitters and k fixed-wavelength receivers connected to each optical star coupler. In the Figure 11.6, for simplicity, the optical transmitters and receivers are not shown. The cells transmitted through the switch are buffered at the input and output modules. In the proposed architecture, input and output lines transmit cells at the rate of 155.52 Mbitrs. The lines that interconnect the input modules to the optical stars, and the optical stars to the output modules, run at 2.5 Gbitrs. The values of n, k, N, and m are 128, 128, 16,384, and 8, respectively. The internal routing header of a cell is composed of two fields. One specifies the output module, and the other the port number in that output module. Each input module handles a single queue, in which the incoming OPTICAL PACKET SWITCHES 288 Ž . Fig. 11.6 The switch architecture proposed by Cisneros and Bracket. 䊚 1991 IEEE. cells are kept in sequence. The input modules, the optical stars, and the output modules are connected as in a three-stage Clos network. However, the working principle is not the same as in the Clos network. Here, each input module sends the output module request of its HOL cell to the CRD. The CRD examines the requests, chooses one cell for each output module, and responds. The cells that won the contention are routed through the first k = k optical star, and their HOL pointers are advanced. This process is repeated cyclically for each optical star. Cells at the output modules are kept in a sequence depending on which optical star they arrive in. In this architecture, all optical stars are kept busy if the CRD is m times faster than the time for cell transfer by the optical stars. The maximum number of optical couplers is determined from the time required to transfer a cell through an optical star and the time required by the CRD to resolve contention. In the architecture, the time required for optical-to-electronic conversion, electronic-to-optical conversion, and tuning optical laser transmitters is not considered. All the calculations are mainly based on the time required to transfer a cell through an optical star. The output port contention resolution scheme is very complex, and the electronic controller can become a bottle- neck. The switch does not have multicast capability, due to the fixed-wave- length receivers. Moreover, the maximum throughput of the switch is limited w x to 58 because of the HOL blocking 18 . OPTOELECTRONIC PACKET SWITCHES 289

11.2.4 BNR Switch