Fast and Efficient Packet Schedulers for Relative Differentiated Services in High-Speed Networks

• Main Authors: Chin-Chi Wu, 吳晉吉
• Other Authors: 林偉
• Format: Others
• Language: en_US
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/77568429222855804759

博士 === 國立中興大學 === 資訊科學系所 === 95 === Recently much research effort has proposed various packet scheduling models or algorithms for the provisioning of relative differentiated services (RDS), such as proportional delay differentiation (PDD), over the Internet. The service differentiation between classes can be simply adjusted with RDS parameters. Because of significant advances in optical technologies and interconnection networks, the increasing gap between link and processor or between link and memory speeds makes the traditional schedulers face a strict challenge. The key issue of prior research which may result in a bottleneck for schedulers arises from either the computational overhead for service rate adjustment or memory latency for the access of packet timestamp. In this dissertation, we propose four fast and efficient schedulers intended for four RDS objectives and with different design considerations. However, they are based on the same assumption that packets come from e links and are classified into N classes. These objectives are proportional delay differentiation, high resilience for traffic bursts, average delay reduction, support of multimedia applications with multiple streams. The first proposed scheduler is called efficient and robust proportional dynamic deficit round-robin (ERPDDRR) which aims at providing PDD between classes. The second one is two-class hierarchical dynamic deficit round-robin (HDDRR) which can provide high resilience for low-priority traffic bursts and achieve average delay reduction. The third and fourth schedulers, multi-level (multi-class) dynamic deficit round-robin (MLDDRR) and multi-level hierarchical dynamic deficit round-robin (MLHDDRR), are designed to support real-time video/multimedia applications with multiple streams. However, the two schedulers, MLDDRR and MLHDDRR, are designed for high- and low-capacity routers/switches, respectively. With small packet weights defined for high-priority packets and the use of token queues to save packet weights, ERPDDRR is an efficient and desirable PDD scheduler if Internet service providers (ISPs) intend to deliver PDD service. The time complexity of ERPDDRR is O(K), where and K is a constant representing the number of token queues. However, because the deployment of delay-sensitive audio, video and multimedia applications increases recently, provision of small delays for short packets and reduction of remarkable delays resulting from traffic bursts become increasingly important issues for RDS schedulers. Hence, different architectures and definitions of packet weights are introduced to HDDRR, MLDDRR, and MLHDDRR. HDDRR is a preliminary scheduler whose architecture can assure quality of service for high-priority traffic when low-priority traffic bursts occur, and reduce average packet delays via an SJF-like scheduling algorithm. The time complexity of HDDRR is O(e). In order to support real-time video or multimedia applications with multiple streams, we modify HDDRR and propose MLDDRR and MLHDDRR with an emphasis on achieving RDS for multiple traffic classes. The architecture and algorithm of MLDDRR is simpler relative to HDDRR. Nevertheless, the time complexity of MLDDRR is O(eN), higher than HDDRR. MLDDRR is intended for backbone high-capacity routers/switches with high-speed links because link utilization is rarely high under ISPs’ control, in such a case, the computational overhead with MLDDRR is lower than MLHDDRR. MLHDDRR is considerably effective in reducing complexity of scheduling time to O(e+N), when high-priority traffic of real-time applications comes in bursts and drastically increases link utilization. And MLHDDRR is extremely suitable for the existing low-capacity edge routers/switches with lower link speeds. Extensive simulations are conducted to demonstrate the performance of our proposed schedulers. Some experiments are carried out with other approaches for comparison. Experiments show that ERPDDRR can achieve stable PDD, especially when a link reaches high utilization and the number of token queues increases. Simulation results also demonstrate that HDDRR can achieve RDS between two classes, deliver small delays for short packets, and provide Premium-like service for the high-priority traffic even though a large amount of low-priority traffic arrives. In the simulations of MLDDRR, we apply three classes of traffic. In addition to achieving RDS among three classes, simulation results show that MLDDRR outperforms the advanced WTP (AWTP) scheduler in delivering small delays for short packets. Finally, the simulations for MLHDDRR confirm that the computational overhead for scheduling decisions can be reduced as the loads of high-priority classes and link utilization increase. The memory occupancy of token queues for each scheduling scheme is so small that token queues can be implemented with high-speed cache to increase scheduling efficiency. Through extensive experiments and analyses, we have shown that our proposed schedulers are effective in delivering relative differentiated services in high-speed networks.