Multi-stage Switching Fabric In The Case Of Link Failure Traffic Distribution Method

At present, many high performance routers including Cisco CRS-1 series carrier router and Dune Network SAND switch architecture are all using multi-plane multi-stage switch fabric. This switch fabric is more scalable and reliable, but it still could not be avoid of congestion because many data cells may compete one output port. To this point, the actually switch fabric usually use multi-plane multi-stage buffered switch fabric.Because of high performance router always running 24 hours every day, it requests a perfect scheme to detect and deal with the failure when switch fabric having some fault links. This paper pay attention to failure detection and propose a algorithm which can distribute traffic according to each switch plane's capacity, then we built two different Opnet simulation models according to push scheduling scheme and pull scheduling scheme to validate the algorithm.First we will introduce Cisco CRS-1 router's push scheduling scheme and Dune Network SAND switch architecture's pull scheduling scheme. In push scheduling scheme mode, when one IM had some data cells, it can send it to switch fabric immediately except that IM received control message from switch fabric. In pull scheduling scheme mode, when one IM had some data cells but it had no credit, it will generate queue control cells and send them to OM which keeps a credit generator to get credits, then it sends data to switch fabric.Next we will introduce the failure detection of the multi-plane multi-stage buffered switch fabric. To detect the fault links, we need a kind of cell called"empty cell". In switch fabric, if there is no data, IM will generate empty cells and send them to switch planes. In each switch plane, SE3 will receive empty cell from SE2, SE2 will receive empty cell from SE1, and SE1 will receive empty cell from IM. If one SE received data cells, it will forward them according their ID; if one SE received empty cells, it will discard them. Besides, if one SE such as SE2 do not received any cells at a special time from one input port connecting to SE1; it shows there are failure links between SE2 and SE1. For multi-plane multi-stage buffered packet switch system, the switch capability will be influenced when one or more switch planes having some failure links. If we do nothing to this point, the switch plane containing failure links may be congested when input traffic become larger. To solve this problem, we propose a flow distribution algorithm and introduce it combine with 3 stage clos switch fabric. Flow distribution algorithm'basic idea: each IM keeps one route table and two kinds of marking field, one field record the location of failure links between SE1 and SE2 and the other field record the failure links between SE2 and SE3. Once an IM received a control cell containing a fault indication and the fault occurred in switch fabric, it will record the location of the failure and distribute traffic according to switch plane's capacity.At last, to test the performance of the flow distribution algorithm, we built two different Opnet simulation models according to Cisco CRS-1 router's push scheduling scheme and Dune Network SAND switch architecture's pull scheduling scheme. The simulation shows that the algorithm can reduce the average delay and the buffer size of switch fabric, and the merit become more and more significant with the input load increasing.