Packet Classification Peng He (Institute of Computing Technology, - - PowerPoint PPT Presentation

▶
packet classification
SMART_READER_LITE
LIVE PREVIEW

Packet Classification Peng He (Institute of Computing Technology, - - PowerPoint PPT Presentation

Jul 23, 2023 122 likes •448 views

Meta-algorithms for Software based Packet Classification Peng He (Institute of Computing Technology, China) Gaogang Xie(Institute of Computing Technology, China) Kav Salamatian (University of Savoie, France) Laurent Mathy (University of

slide-1
SLIDE 1

Meta-algorithms for Software based Packet Classification

Peng He (Institute of Computing Technology, China) Gaogang Xie(Institute of Computing Technology, China) Kavé Salamatian (University of Savoie, France) Laurent Mathy (University of Liège, Belgium)

slide-2
SLIDE 2
  • Motivation
  • Key Observations
  • Two Meta Methods

Memory Consumption Model Characterizing range distribution uniformity

  • The AutoPC framework and SmartSplit algorithm
  • Experiment Results.

Agenda

2 ICNP 2014

slide-3
SLIDE 3
  • Motivation
  • Key Observations
  • Two Meta Methods

Memory Consumption Model Characterizing range distribution uniformity

  • The AutoPC framework and SmartSplit algorithm
  • Experiment Results.

Agenda

3 ICNP 2014

slide-4
SLIDE 4

Packet Classification

4 ICNP 2014

Packet Classification: find the highest priority rule that matches a packet Classifier: a set of rules Packet classification is key for

  • Security
  • Traffic monitoring and analysis
  • QoS

Source IP Destination IP Source Port Destination Port Protocol Action 120.0.0.0/24 198.12.130.0/2 0:65535 11:17 0xFF/0xFF Accept 138.42.83.1/0 174.3.18.0/8 50:10000 0:65535 0x06/0xFF Deny

Packet classification prevalent in modern routers

slide-5
SLIDE 5
  • TCAM based Solutions

fast, deterministic performance. power-hungry, expensive.

  • RAM based (algorithmic) Solutions

Tradeoff between memory size and memory accesses. Theoretical Bounds

 O(log N) speed and O(Nk) space  O(logk-1N) speed and O(N) space

All existing algorithms are heuristic algorithms, exploiting special ruleset structures.

Two Solutions

5 ICNP 2014

slide-6
SLIDE 6

Performance Unpredictability

6 ICNP 2014

HyperSplit EffiCuts Ruleset 1 Ruleset 2 same algorithm, different ruleset, same algorithm, different ruleset, Choose the right algorithm for different ruleset!

slide-7
SLIDE 7
  • Study this performance unpredictability

 Two ruleset features

  • Develop methods to predict performance

 Two meta methods

  • choose the right algorithm and develop more efficient one

 The AutoPC framework and SmartSplit Algorithms

  • Workaround

Compare all the alternative algorithms, choose one with better performance. Need more than 24 hours to build a tree for some rule-

  • sets. (INFOCOM 09’)

Our contributions

7 LANMAN 2013

slide-8
SLIDE 8
  • Motivation
  • Key Observations
  • Two Meta Methods

Memory Consumption Model Characterizing range distribution uniformity

  • The AutoPC framework and SmartSplit algorithm
  • Experiment Results.

Agenda

8 ICNP 2014

slide-9
SLIDE 9

Geometric View of rulesets

9 ICNP 2014

Each rule can be viewed as a hyperrectangle Algorithms performs “cuts” or “splits”

  • n the space to reduce the search

space.

slide-10
SLIDE 10

An Example (HiCuts and HyperSplit)

10 ICNP 2014

Equal sized cutting A lot of rule duplications, large memory size Unequal sized splitting Less rule duplications, smaller memory size

2 dimension rule-set Each field has 4 bits 00* 01* 10*11* [0,13] [14,15]

slide-11
SLIDE 11

Non-uniformly distributed ranges

11 ICNP 2014

Need 8 equal-sized cuts to separate R1 and R2, however R3, R4 and R5 get duplicated. 111* and 110* just concentrate in a small part of the full range [0,15]

range distribution on Field 1 is not uniform, resulting in more memory accesses or more rule duplications.

slide-12
SLIDE 12

Sparse rules and orthogonal structure rules

12 ICNP 2014

No matter how to cut, rules get duplicated. Memory size increases. Rules can be easily sperated

slide-13
SLIDE 13

EffiCuts

13 ICNP 2014

Untangle the orthogonal structures Smallest memory size No rules get duplicated

slide-14
SLIDE 14

Orthogonal Structures can be invalid.

14 ICNP 2014

2 memory accesses 1 memory access Field 1 is enough to separate rules. Orthogonal structures can be ignored.

slide-15
SLIDE 15

Discussions

15 ICNP 2014

  • “Orthogonal structures” should be considered, and rules

should be eventually splitted in order to untangle these structure and avoid memory explosion.

  • When splitting a ruleset, if a dimension appears that contain
  • nly small ranges, it should be used to separate the rules

with a single tree.

  • Equal-sized cutting becomes more efficient when ruleset

ranges are uniform, if not splitting with non- equal sized intervals should be considered.

slide-16
SLIDE 16
  • Motivation
  • Key Observations
  • Two Meta Methods

Memory Consumption Model Characterizing range distribution uniformity

  • The AutoPC framework and SmartSplit algorithm
  • Experiment Results.

Agenda

16 ICNP 2014

slide-17
SLIDE 17

Memory Consumption Model

17 ICNP 2014 R5 R6 R7 R8 R1 R2 R3 R4

F1 F2

Key Observations: Cutting F1 will usually cause (small, large) rules duplicate. Cutting F2, (large,small) rules will get duplicated.

ls rules X cuts sl rules Y cuts

𝑚𝑡 × 𝑍 + 𝑡𝑚 × 𝑌 rule duplications 90% memory is for rule duplications. R1, R2, R3, R4 are (small, large) rules while R5, R6, R7 and R8 are (large small) rules.

slide-18
SLIDE 18

Memory Consumption Model

18 ICNP 2014

We need

𝑚𝑡 𝑐𝑗𝑜𝑢ℎ/2 cuts on F1, 𝑡𝑚 𝑐𝑗𝑜𝑢ℎ/2 cuts on F2

Assume: in each divided space, there are binth rules. Half of rules are (large, small) rules and the other half are (small, large) rules. In total, we have

𝑚𝑡 𝑐𝑗𝑜𝑢ℎ/2 × 𝑡𝑚 + 𝑡𝑚 𝑐𝑗𝑜𝑢ℎ/2 × 𝑚𝑡

rule duplications

slide-19
SLIDE 19
  • How about (small, small) rules and (large, large)

rules?

See paper for details.

  • How to minimize the estimation error?

See paper for details

Memory Consumption Model

19 ICNP 2014

𝑁𝑡𝑡 = 𝑡𝑡 × 𝑄𝑈𝑆 𝑁𝑚𝑡 = 𝑚𝑡 × 𝑡𝑚 + 𝑡𝑡 × 𝛽 𝑐𝑗𝑜𝑢ℎ/2 × 𝑄𝑈𝑆 𝑁𝑡𝑚 = 𝑡𝑚 ×

𝑚𝑡+𝑡𝑡 ×(1−𝛽) 𝑐𝑗𝑜𝑢ℎ/2

× 𝑄𝑈𝑆 𝑁𝑚𝑚 = 𝑚𝑚 × 𝑡𝑚 + 𝑡𝑡 × 𝛽 𝑐𝑗𝑜𝑢ℎ/2 × 𝑚𝑡 + 𝑡𝑡 × (1 − 𝛽) 𝑐𝑗𝑜𝑢ℎ/2 × 𝑄𝑈𝑆

slide-20
SLIDE 20

Characterizing range distribution uniformity

20 ICNP 2014

Characterizing range distribution by the shape

  • f interval tree!

Build a interval tree for the ranges on each field.

slide-21
SLIDE 21

The shape of Interval trees

21 ICNP 2014

Quasi-Balanced subtrees 𝐶𝑠𝑏𝑢𝑗𝑝 =

# 𝑂𝑝𝑒𝑓𝑡 𝑗𝑜 𝑢ℎ𝑓 𝑙𝑢ℎ 𝑚𝑓𝑤𝑓𝑚 #𝑂𝑝𝑒𝑓𝑡 𝑗𝑜 𝑢ℎ𝑓 𝑙−1 𝑢ℎ 𝑚𝑓𝑤𝑓𝑚 ≥ 1.5

Balance tree distance D Max Balance tree distance Dmax Dmax = 3 for this tree

slide-22
SLIDE 22

The threshold of range distribution uniformity

22 ICNP 2014

When 𝑬𝒏𝒃𝒚 ≤

𝟐 𝟑 𝒎𝒑𝒉 #𝒔𝒗𝒎𝒇𝒕 𝒄𝒋𝒐𝒖𝒊

One should use equal-sized cut based algorithm

See paper for details.

slide-23
SLIDE 23

SmartSplit Algorithm

23 ICNP 2014

ss sl ls ll all

Split the ruleset according to large/small ranges on IP fields Merge rulesets (ss, sl) or (ss, ls) Use different algorithms on three rulesets.

rs1 rs2 rs3 al1 al2 al3

slide-24
SLIDE 24

The AutoPC framework

24 ICNP 2014

Automatically perform the tradeoff for a given ruleset.

slide-25
SLIDE 25
  • We conduct extensive experiments for memory

model and compare the performance of AutoPC and SmartSplit.

Experiments Results

25 ICNP 2014

  • The performance of

different algorithms is related the logarithm of the memory

slide-26
SLIDE 26

Experiments Results

26 ICNP 2014

  • For 60 rulesets of

various size, the real memory size .vs. the memory estimation

slide-27
SLIDE 27

Memory estimation is fast

27 ICNP 2014

1000x faster then really building a decision tree

slide-28
SLIDE 28

Experiments Results

28 ICNP 2014

  • Compare SmartSplit

and EffiCuts

10x small memory 2~4x fast

slide-29
SLIDE 29

Real Evaluation

29 ICNP 2014

  • On Xeon machines, the

SmartSplit is in average 2 times faster.

  • AutoPC can choose

right algorithm, which can be 3.8 / 18.9 faster than using EffiCuts or HyperSplit alone.

slide-30
SLIDE 30
  • Orthogonal structure and non-uniformly distributed ranges

have impact on algorithm performance

  • Memory size can be roughly estimated by simply counting
  • rthogonal structure rules (memory consumption model)
  • Through exploiting uniformity of range distributions, we can

improve the performance (SmartSplit)

  • Through carefully choose right algorithm, we can

automatically achieve better tradeoff between memory size and memory accesses (AutoPC)

Conclusion

30 ICNP 2014

slide-31
SLIDE 31

Thank you!

hepeng@ict.ac.cn