1. Toward Advocacy-Free Evaluation of Packet Classification Algorithms
Author:
Haoyu Song and Jonathan S. Turner
Publisher:
IEEE TRANSACTIONS ON COMPUTERS 2011
Presenter:
Hsin-Mao Chen
2011/05/04

2. Outline Introduction High-Level Review Challenges Approach
Example Illustrated

3. Introduction
Each approach has its own advantages and disadvantages in terms of throughput, cost, power dissipation, ease of implementation, scalability.
Enforce some consistent and fundamental criteria for the algorithm evaluation, so their performance and potentials are directly comparable.

4. Outline Introduction High-Level Review Challenges Approach
Example Illustrated

5. High-Level Review Algorithm solution TCAM solution
Map the packet classification problem to the point location problem in a multidimensional space.
TCAM solution
Store ternary bit strings and perform parallel searches on all of its entries simultaneously.

6. Algorithmic solutions
One Theme – Space and Time Tradeoff
Cross-producting [3] and RFC [4].
Two Strategies – Cutting and Projecting

7. Algorithmic solutions
Three Techniques – Splitting, Intersecting, and Grouping
Splitting: Woo’s modular packet classification [5], HiCuts [6], and HyperCuts [7].
Intersecting: BV [8] and ABV [9].
Grouping: Tuple Space Search(TSS) [16], and 2D compressed Tuple Space Search [17].

8. TCAM Solution Low density and high cost High power consumption
20 times more than SRAMs
Hundreds of times more than DRAMs
High power consumption
Poor arbitrary range support
An original filter can become as many as 900 expanded filters in the worst case [20].
Poor multiple-match support

9. Outline Introduction High-Level Review Challenges Approach
Example Illustrated

10. Challenges Incommensurable Evaluation Results
Not base on the same filter sets.
Not share a common implementation model.
Irreproducible Implementations and Evaluation Results
Parameter settings and the filter sets used.
Incomplete Evaluation and Unconvincing Results
Tradeoffs, heuristics, and optimizations.

11. Outline Introduction High-Level Review Challenges Approach
Example Illustrated

12. Approach
Throughput, storage, incremental update support, preprocessing time, scalability to the size of filter sets, adaptability to the structure of filter sets, implementation cost, and power dissipation.
Evaluation results should be normalized in a directly comparable way.

13. Approach
On-chip resource: search engines, packet dispatcher, and the memory interface.
Off-chip memory: all data structures.

14. Approach Memory consumption: cost
Average number of bytes consumed per filter.
One algorithm may involve multiple data structures and tables, but the evaluation fails to cover all of them.
Only the overall memory consumption is reported, but the size of the filter set applied is unknown.
Memory bandwidth consumption: throughput
The number of bytes per memory access.
The number of dependent memory access per packet lookup.

15. Approach

16. Outline Introduction High-Level Review Challenges Approach
Example Illustrated

17. Example Illustrated HiCuts Bucket size: 16 Space measure factor: 2
Dimension choose option: smallest of sum of # filters at each child of node.

18. Example Illustrated

19. Example Illustrated Preprocessing time
2.93 GHz Intel Q6800 CPU and 3 GB RAM