Dynamically-Allocated Multi-Queue Buffers for VLSI Communication Switches
IEEE Transactions on Computers
ISCA '90 Proceedings of the 17th annual international symposium on Computer Architecture
Tiny Tera: A Packet Switch Core
IEEE Micro
Queueing in high-performance packet switching
IEEE Journal on Selected Areas in Communications
Destination-Based HoL Blocking Elimination
ICPADS '06 Proceedings of the 12th International Conference on Parallel and Distributed Systems - Volume 1
Cost / performance trade-offs and fairness evaluation of queue mapping policies
Euro-Par'05 Proceedings of the 11th international Euro-Par conference on Parallel Processing
| Hi-index | 0.00 |
Head-of-line blocking is one of the main problems arising in input-buffered switches. The best-known solution to this problem consists of using Virtual Output Queues (VOQs). However this strategy is not scalable. Its implementation cost increases quadratically with the number of ports in the switch. Taking into account current trends, the demand for larger number of ports in high-performance switches is likely to increase very rapidly in the future. Therefore, a scalable and cost-effective solution is required. In this paper we propose an efficient and cost-effective strategy (belonging to a family of strategies previously proposed, referred to as Destination-Based Buffer Management (DBBM)), to reduce HOL blocking in single-stage and multistage networks. The proposed strategy is based on allowing certain destinations to share the same queue. Its main purpose is to maximize network throughput whereas keeping HOL blocking to negligible values. In this paper, we apply the strategy at every switch included in a bidirectional multistage network (BMIN). We have evaluated DBBM, VOQ, and alternative strategies in different BMIN sizes and with different traffic conditions (synthetic traffic, IP traces, and I/O traces). Results show that DBBM with a reduced number of queues at each switch obtains roughly the same throughput as the VOQ mechanism. Moreover, VOQ at the switch level (as many queues as output ports at every switch) has also been analyzed. Results demonstrate that it does not scale. As the number of stages in the network increases, the VOQ solution at the switch level introduces more HOL blocking that leads to a severe degradation in network throughput. With the DBBM using 16 queues, maximum network throughput is sustained for all the traffic cases analyzed. Moreover, as the network size increases (up to a 2048脳2048BMIN), DBBM keeps roughly the same performance with the same number of queues.