authors: yang xu, zhaobo liu, zhuoyuan zhang, h. jonathan chao conf. : ancs 2009
DESCRIPTION
An ultra high throughput and memory efficient pipeline architecture for multi-match packet classification without TCAMs. Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009 Presenter : JHAO-YAN JIAN Date : 2011/12/14. INTRODUCTION. - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/1.jpg)
Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan ChaoConf. : ANCS 2009Presenter : JHAO-YAN JIANDate : 2011/12/14
1
An ultra high throughput and memory efficient pipeline architecture for multi-
match packet classification without TCAMs
![Page 2: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/2.jpg)
INTRODUCTION decompose the operation of multi-match
packet classification from the complicated multi-dimensional search to several single-dimensional searches
present an asynchronous pipeline architecture based on a signature tree structure to combine the intermediate results returned from single-dimensional searches.
propose two edge-grouping algorithms to partition the hash table at each stage of the pipeline into multiple hash tables.
2
![Page 3: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/3.jpg)
PROBLEM STATEMENT
3
Coming packet
Return by single-dimensional searches
![Page 4: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/4.jpg)
SIGNATURE TREE
4
![Page 5: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/5.jpg)
ASYNCHRONOUS PIPELINE ARCHITECTURE
5
![Page 6: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/6.jpg)
Character-Based Edge-Grouping
reserve the first log2 (M +1) bit of each character to be the locating prefix, whose value is between 0 and M.
Given the M+1 edge sets after the edge-grouping, we could store edges of each independent edge set into an individual hash table, while duplicate edges of the residual edge set into all M hash tables.
Given M work-conserving hash tables and y matching characters, the processing of each active node ID can be finished within y /M parallel hash accesses.
6
![Page 7: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/7.jpg)
Character-Based Edge-Grouping
7
![Page 8: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/8.jpg)
Node-Character-Based Edge-Grouping
NCB_EG scheme stores the grouping information of each edge in both the encoding of the edge’s associated character, and the ID of the edge’s source nodelocating prefix(sft) :first log2 (M +1) bit of each
charactershifting prefix(loc) :first log2 M bits of each node
hash table indexed by (sft+loc-1) mod M +1If loc equals 0, the edge is a residual edge, and
could be found in any partitioned hash tables.
8
![Page 9: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/9.jpg)
Node-Character-Based Edge-Grouping
9
![Page 10: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/10.jpg)
EVALUATION RESULTS
10
![Page 11: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/11.jpg)
EVALUATION RESULTS
11
![Page 12: Authors: Yang Xu, Zhaobo Liu, Zhuoyuan Zhang, H. Jonathan Chao Conf. : ANCS 2009](https://reader036.vdocuments.mx/reader036/viewer/2022062321/56814000550346895dab351d/html5/thumbnails/12.jpg)
EVALUATION RESULTS
12
Suppose the on-chip SRAM access frequency is 400 MHz, the smallest size of IP packet is 64 bytes. The proposed pipeline can achieve a throughput between 19.5Gbps and 91Gbps with different types of classifiers.