Virtualizing Modern High-Speed Interconnection Networks with Performance and Scalability

Author(s):  
Bo Li ◽  
Zhigang Huo ◽  
Panyong Zhang ◽  
Dan Meng
2002 ◽  
Vol 03 (01n02) ◽  
pp. 49-65 ◽  
Author(s):  
NADER F. MIR

A thorough routing analysis of a switching network called the spherical switching network for high-speed applications is presented in this paper. The spherical switching network has a cyclic, regular, and highly expandable structure with a simple self-routing scheme. The network is constructed with fixed-size switch elements regardless of the size of the network. Each switch element consists of a carefully-selected sized 9 input/output crossbar and a local controller. One of the nine pairs of links is external and carries the external traffic, and the other eight pairs are internal. The contention resolution in each switch element is based on deflection of losing packets and incremental priority of packets. The switch elements do not utilize any buffering within the network. The analysis shows that this network clearly outperforms typical interconnection networks currently being deployed in practical switches and routers such as Banyan network. In order to keep the number of deflections low, each incoming external link is connected to a buffer with flow control capabilities. Due to the special arrangement of interconnections in the network, a much larger number of shortest paths between each pair of source/destination exists. The related analysis for finding the number of hops and shortest paths appear in this paper.


2006 ◽  
Vol 07 (04) ◽  
pp. 535-548 ◽  
Author(s):  
Shihang Yan ◽  
Geyong Min ◽  
Irfan Awan

Credit-based flow control scheme that can be used to support both end-to-end and link-level flow control is becoming increasingly popular in high speed system area networks (SAN), e.g. InfiniBand networks where multiple processor nodes and I/O devices are interconnected using switched point-to-point links. By virtue of such a scheme, the downstream node sends credits to the upstream node indicating the availability of buffer spaces. Upon receiving credits, the upstream node injects packets into the networks. Performance analysis of credit-based flow control scheme plays an important role for the design and optimization of InfiniBand interconnection networks which have been widely used in many high-performance cluster, Grid and P2P computing systems. This study develops a new queueing network model for performance evaluation of credit-based flow control in InfiniBand networks. The performance metrics to be derived include the mean queue length, throughput and response time of the system. Simulation experiments have been used to validate the accuracy of the queueing network model. Results obtained from the analytical model have showed that this model can effectively evaluate the performance of credit-based flow control in InfiniBand networks.


Sign in / Sign up

Export Citation Format

Share Document