| As Internet traffic is explosively increasing every year, Internet capacity should take rank of these traffic growths. Router/Switch is one of the core components of the Internet and its technologies are outpaced by Internet traffic evolution. Currently known switch architectures are suffered from complex arbitration process to route cells through the switch, memory speed limitation, scalability and etc. As a response for these needs, this dissertation finds ways to improve them by introducing new switch architectures. The major contribution of this dissertation is to present several high performance switch architectures that don't require any centralized arbitration process while achieving high performance.;As an improvement of input queued switch, a contention-tolerant crossbar (CTC(N)) switch was proposed in the previous study. To improve the throughput, this dissertation presents Multi-layer CTC(N) (MCTC(N)) switch architecture. MCTC(N) switch consists of k number of CTC(N) fabrics and the traffic is evenly distributed over these layers. MCTC( N) switch with 2 layers was proven to achieve 100% throughput under Bernoulli i.i.d. uniform traffic. Due to the architectural merit of CTC(N) switch, we can construct large N x N switches with the number of switching elements (SEs) significantly smaller than that of CTC( N). In addition, we present several of its variations, SCTC (N, n) (sparse CTC), SMCTC( N, n) (sparse multi-layer CTC) and MSCTC( N, n) (multi-layer sparse CTC). These variations maintain the same performance of their counterparts CTC(N) and MCTC(N), while reducing the cost form O(N2) to O(N log N).;Although CTC(N) and its variations can achieve good performance in terms of throughput, it has cell out-of-order problem. Therefore, Mesh of Trees based switch architecture, denoted as MOTS(N), is introduced. MOTS( N) ensures QoS like FIFO output queued switch and it's shown to achieve 100% throughput by theoretical analysis. However, the assumption made for MOTS(N) switch is that size of all internal buffers is unlimited. To overcome this impracticality, we present one of its improved (practical) variations, IMOTS(N) switch. For IMOTS(N) switch architecture, size of the buffers in the fabric is limited by a certain amount. To more improve IMOTS(N), two more its variations are also presented; IMOTS(N) with cut-through and advanced IMOTS(N) which are denoted as IMOTS-CT( N) and AIMOTS(N), respectively. |
