AMTrie: An OC192 Parallel Multidimensional Packet Classification Algorithm for Network Processor

1
AM-Trie An OC-192 Parallel Multidimensional
Packet Classification Algorithm for Network
Processor

• Bo Zheng, Chuang Lin and Xuehai Peng
• IMSCCS06

2
Outline

• Introduction
• Background
• AM-Trie Classification Algorithm
• Optimizing the space complexity
• Performance evaluation

3
Introduction

• NPU combines high-speed of ASIC and the
  flexibility of software together
• AM-Trie( Asymmetrical Multi-bit Trie )
• Height is much lower than normal trie
• Eliminate trace back
• A parallel algorithm which fits NPU
• Good scalability

4
Background

• NPU has the following features
• Parallel processing
• Multiple hardware thread
• Optimized ALU
• Most ALU operation can be completely in 1 cycle
• Coprocessor for networking
• Optimized memory management and optimized DMA
  unit
• Ex sram queue instructions are provide speed up
  queuing operation

5
AM-Trie Classification Algorithm

• The first idea of AM-Trie
• express arbitrary length of prefix using the
  redundancy.
• All possible prefix is , 1, 0, , 11111( k 1),
  totally 2(k1) 1 prefixes.
• Uses k 1 bits to express k-bit prefix
• Defines Serial number

6
AM-Trie creation

• Each node of AM-Trie have two pointer
• pNext ( point to child ) pRule ( point to
  rule)
• Firstly, divide the policy into predifined parts
  h1, h2,hl.
• Then, calculate the serial number of each section
  use the redundant expression
• Calculate the offset value of each level
  according to the serial number

7
AM-Trie creation
8
AM-Trie creation

• Ex

9
AM-Trie creation

• Why do we call it AM-Trie?
• Only right part of the trie may have children
  nodes
• The trie looks asymmetrical

10
AM-Trie Search, Compression andUpdate

• In AM-Trie search, if a leaf node doesnt contain
  a rule, we should go on searching its prefix
  node.
• Ex 0110gt

11
AM-Trie Search, Compression andUpdate

• After push and compression

12
Multidimensional Packet Classification Using
AM-Trie

• In order to perform multidimensional
  classification using AM-Trie, we profit from the
  idea of Aggregated Bit Vector (ABV)
• Operation in different dimension are independent,
  they can be parallel executed
• NPU is fit for the action

13
Multidimensional Packet Classification Using
AM-Trie

• Notation
• N rules,W bits, h height, d dimension
• (1) the initial complexitygt
• (2) the search complexitygt O( h d )
• (3) the space complexitygt O(N2)
• In worst case, the require storage is less than

14
Optimizing the Space Complexity of AM-Trie
algorithm

• Fixed stride AM-Trie

15
Fixed stride AM-Trie
16
Optimizing the Space Complexity of AM-Trie
algorithm

• Variable stride AM-Trie
• The complexity of finding out optimal division is
• Not fit in real application

17
Variable stride AM-Trie
18
Performance Evaluation

• Resource allocation
• IXP2400
• Packet RX, Packet TX, Buffer Management and
  Scheduling module use 1 ME respectively, and 4
  MEs carry on parallel AM-Trie classification as
  well as IP header processing
• IXP2800
• 2 for RX, 2 for TX, Buffer Management, Statistic
  and Scheduling module use 1 ME respectively. 8
  MEs for AM-Trie classification

19
Performance Evaluation
20
Performance Evaluation