An Efficient and Low Overhead Random Forwarding Table Construction Method for Deadlock-Free Routing Algorithms in InfiniBand Networks

碩士 === 國立清華大學 === 資訊工程學系 === 93 === The InfiniBand architecture (IBA) is an interconnect network standard designed for server I/O and inter-server communication. The IBA defines a switch-based network with point-to-point links to support arbitrary topology. Routing in an IBA network is based on th...

Full description

Bibliographic Details
Main Authors: Hsien-Ting Kuo, 郭顯廷
Other Authors: Yeh-Ching Chung
Format: Others
Language:zh-TW
Published: 2005
Online Access:http://ndltd.ncl.edu.tw/handle/07054885585097924860
Description
Summary:碩士 === 國立清華大學 === 資訊工程學系 === 93 === The InfiniBand architecture (IBA) is an interconnect network standard designed for server I/O and inter-server communication. The IBA defines a switch-based network with point-to-point links to support arbitrary topology. Routing in an IBA network is based on the forwarding table stored in each switch. The forwarding table forwards packets to output ports by only considering the destination addresses of packets. However, the forwarding tables computed by many irregular routing algorithms consider both the input port and the destination address of the packet. Therefore, these irregular routing algorithms can not be used in IBA network because they do not conform to the IBA forwarding table format. There are two types of IBA forwarding table: Linear forwarding table (LFT) and Random forwarding table (RFT). In this thesis, we propose a two-phase method, partial renaming method, to construct the deadlock-free RFT according to the paths computed by any irregular deadlock-free routing algorithm which considers both the input port and the destination address. The partial renaming is an efficient and low overhead forwarding table construction method for IBA. The forwarding tables constructed by the partial renaming method do not modify the original paths computed by the routing algorithm before. This reserves the characteristic of the original routing algorithm. To evaluate the proposed method, we compare our method with the renaming scheme on a simulator. We evaluate the forwarding table overhead and the computation time of both methods in different network size and different switch settings. The simulation results show that the computation time and the size of forwarding table introduced by our method are much less than those introduced by the renaming scheme in all test cases