Date of Award

Fall 2007

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Electrical Engineering - (Ph.D.)

Department

Electrical and Computer Engineering

First Advisor

Roberto Rojas-Cessa

Second Advisor

Nirwan Ansari

Third Advisor

Sotirios Ziavras

Fourth Advisor

Jie Hu

Fifth Advisor

Aleksandar Kolarov

Abstract

Combined input crosspoint buffered (CICB) packet switches were introduced to relax inputoutput arbitration timing and provide high throughput under admissible traffic. However, the amount of memory required in the crossbar of an N x N switch is N2x k x L, where k is the crosspoint buffer size and needs to be of size RTT in cells, L is the packet size. RTT is the round-trip time which is defined by the distance between line cards and switch fabric. When the switch size is large or RTT is not negligible, the memory amount required makes the implementation costly or infeasible for buffered crossbar switches. To reduce the required memory amount, a family of shared memory combined-input crosspoint-buffered (SMCB) packet switches, where the crosspoint buffers are shared among inputs, are introduced in this thesis. One of the proposed switches uses a memory speedup of in and dynamic memory allocation, and the other switch avoids speedup by arbitrating the access of inputs to the crosspoint buffers. These two switches reduce the required memory of the buffered crossbar by 50% or more and achieve equivalent throughput under independent and identical traffic with uniform distributions when using random selections.

The proposed mSMCB switch is extended to support differentiated services and long RTT. To support P traffic classes with different priorities, CICB switches have been reported to use N2x k x L x P amount of memory to avoid blocking of high priority cells.The proposed SMCB switch with support for differentiated services requires 1/mP of the memory amount in the buffered crossbar and achieves similar throughput performance to that of a CICB switch with similar priority management, while using no speedup in the shared memory.

The throughput performance of SMCB switch with crosspoint buffers shared by inputs (I-SMCB) is studied under multicast traffic. An output-based shared-memory crosspoint buffered (O-SMCB) packet switch is proposed where the crosspoint buffers are shared by two outputs and use no speedup. The proposed O-SMCB switch provides high performance under admissible uniform and nonuniform multicast traffic models while using 50% of the memory used in CICB switches. Furthermore, the O-SMCB switch provides higher throughput than the I-SMCB switch.

As SMCB switches can efficiently support an RTT twice as long as that supported by CICB switches and as the performance of SMCB switches is bounded by a matching between inputs and crosspoint buffers, a new family of CICB switches with flexible access to crosspoint buffers are proposed to support longer RTTs than SMCB switches and to provide higher throughput under a wide variety of admissible traffic models. The CICB switches with flexible access allow an input to use any available crosspoint buffer at a given output. The proposed switches reduce the required crosspoint buffer size by a factor of N , keep the service of cells in sequence, and use no speedup. This new class of switches achieve higher throughput performance than CICB switches under a large variety of traffic models, while supporting long RTTs.

Crosspoint buffered switches that are implemented in single chips have limited scalability. To support a large number of ports in crosspoint buffered switches, memory-memory-memory (MMM) Clos-network switches are an alternative. The MMM switches that use minimum memory amount at the central module is studied. Although, this switch can provide a moderate throughput, MMM switch may serve cells out of sequence. As keeping cells in sequence in an MMM switch may require buffers be distributed per flow, an MMM with extended memory in the switch modules is studied. To solve the out of sequence problem in MMM switches, a queuing architecture is proposed for an MMM switch. The service of cells in sequence is analyzed.

Share

COinS