CAMP: Fast and Efficient IP Lookup Architecture

1
CAMP Fast and Efficient IP Lookup Architecture

• Sailesh Kumar, Michela Becchi,
• Patrick Crowley, Jonathan Turner
• Washington University in St. Louis

2
Context

• Trie based IP lookup
• Circular pipeline architectures

3
Context

• Trie based IP lookup
• Circular pipeline architectures

IP address 111010
Prefix dataset
0 P1
000 P2
0010 P3
0011 P4
011 P5
10 P6
11 P7
110 P8
4
Context

• Trie based IP lookup
• Circular pipeline architectures

IP address 111010
Trie
Prefix dataset
P1
0 P1
000 P2
0010 P3
0011 P4
011 P5
10 P6
11 P7
110 P8
P7
P6
P5
P2
P8
P3
P4
5
Context

• Trie based IP lookup
• Circular pipeline architectures

IP address 111010
Trie
Prefix dataset
0 P1
000 P2
0010 P3
0011 P4
011 P5
10 P6
11 P7
110 P8
Stage 1
Stage 2
Stage 3
Stage 4
6
Context

• Trie based IP lookup
• Circular pipeline architectures

IP address 111010
Trie
Circular pipeline
Prefix dataset
0 P1
000 P2
0010 P3
0011 P4
011 P5
10 P6
11 P7
110 P8
Stage 1
Stage 2
Stage 3
Stage 4
7
CAMP Circular Adaptive and Monotonic Pipeline

• Problems
• Optimize global memory requirement
• Avoid bottleneck stages
• Make the per stage utilization uniform
• Idea
• Exploit a Circular pipeline
• Each stage can be a potential entry-exit point
• Possible wrap-around
• Split the trie into sub-trees and map each of
  them independently to the pipeline

8
CAMP (contd)

• Implications
• PROS
• Flexibility decoupling of maximum prefix length
  from pipeline depth
• Upgradeability memory bank updates involve only
  partial remapping
• CONS
• A stage can be simultaneously an entry point and
  a transition stage for two distinct requests
• Conflicts origination
• Scheduling mechanism required
• Possible efficiency degradation

9
Trie splitting
Direct index table

• Define initial stride x
• Use a direct index table with 2x entries for
  first x levels
• Expand short prefixes to length x
• Map the sub-trees

Subtree 1
Subtree 3
Subtree 2
E.g. initial stride x2
10
Dealing with conflicts

• Idea use a request queue in front of each stage
• Intuition without request queues,
• a request may wait till n cycles before entering
  the pipeline
• a waiting request causes all subsequent requests
  to wait as well, even if not competing for the
  same stages
• Issue ordering
• Limited to requests with different entry stages
  (addressed to different destinations)
• An optional output reorder buffer can be used

11
Pipeline Efficiency

• Metrics
• Pipeline utilization fraction of time the
  pipeline is busy provided that there is a
  continuous backlog of requests
• Lookups per Cycle (LPC) average request
  dispatching rate
• Linear pipeline
• LPC1
• Pipeline utilization generally low
• Not uniform stage utilization
• CAMP pipeline
• High pipeline utilization
• Uniform stage utilization
• LPC close to 1
• Complete pipeline traversal for each request
• pipeline stages trie levels
• LPC gt 1
• Most requests dont make complete circles around
  pipeline
• pipeline stages gt trie levels

12
Pipeline efficiency all stages traversed

• Setup
• 24 stages, all traversed by each packet
• Packet bursts sequences of packets to same entry
  point
• Results
• Long bursts result in high utilization and LPC
• For all burst size, enough queuing (32)
  guarantees 0.8 LPC

13
Pipeline efficiency LPC gt 1

• Setup
• 32 stages, rightmost 24 bits, tree-bit map of
  stride 3
• Average prefix length 24
• Results
• LPC between 3 and 5
• Long bursts result in lower utilization and LPC

14
Nodes-to-stages mapping

• Objectives
• Uniform distribution of nodes to stages
• Minimize the size of the biggest stage
• Correct operation of the circular pipeline
• Avoid multiple loops around pipeline
• Simplified update operation
• Avoid skipping levels

15
Nodes-to-stages mapping (contd)

• Problem Formulation (constrained graph coloring)
• Given
• A list of sub-trees
• A list of colors represented by numbers
• Color nodes so that
• Every color is nearly equally used
• A monotonic ordering relationship without gaps
  among colors is respected when traversing
  sub-trees from root to leaves
• Algorithm (min-max coloring heuristic)
• Color sub-trees in decreasing order of size
• At each steps
• Try all possible colors on root (the rest of the
  sub-tree is colored consequentially)
• Pick the local optimum

16
Min-max coloring heuristic - example
T4
T3
T2
T1
Present coloring If 1 on new root If 2 on new root If 3 on new root If 4 on new root
Color 1 1
Color 2 2
Color 3 4
Color 4 5
17
Min-max coloring heuristic - example
T4
T3
T2
T1
Present coloring If 1 on new root If 2 on new root If 3 on new root If 4 on new root
Color 1 1 2 5 3 2
Color 2 2 3 3 6 4
Color 3 4 6 5 5 8
Color 4 5 9 7 6 6
18
Min-max coloring heuristic - example
T4
T3
T2
T1
Present coloring If 1 on new root If 2 on new root If 3 on new root If 4 on new root
Color 1 1 2 5 3 2
Color 2 2 3 3 6 4
Color 3 4 6 5 5 8
Color 4 5 9 7 6 6
19
Min-max coloring heuristic - example
T4
T3
T2
T1
Present coloring If 1 on new root If 2 on new root If 3 on new root If 4 on new root
Color 1 3 4 5 4 5
Color 2 6 8 7 8 7
Color 3 5 6 7 6 7
Color 4 6 8 7 8 7
20
Min-max coloring heuristic - example
T4
T3
T2
T1
Present coloring If 1 on new root If 2 on new root If 3 on new root If 4 on new root
Color 1 3 4 5 4 5
Color 2 6 8 7 8 7
Color 3 5 6 7 6 7
Color 4 6 8 7 8 7
21
Min-max coloring heuristic - example
T4
T3
T2
T1
Present coloring If 1 on new root If 2 on new root If 3 on new root If 4 on new root
Color 1 5
Color 2 7
Color 3 7
Color 4 7
22
Evaluation settings

• Trends in BGP tables
• Increasing number of prefixes
• Most of prefixes are lt26 bit (24 bit) long
• Route updates can concentrate in short period of
  time however, they rarely change the shape of
  the trie
• 50 BGP tables containing from 50K to 135K
  prefixes

23
Memory requirements
CAMP
Level based mapping
Height based mapping

• Balanced distribution across stages
• Reduced total memory requirements
• Memory overhead 2.4 w/ initial stride 8, 0.02
  w/ initial stride 12, 0.01 w/ initial stride 16

24
Updates

• Techniques for handling updates
• Single updates inserted as bubbles in the
  pipeline
• Rebalancing computed offline and involving only a
  subset of tries
• Scenario
• migration between different BGP tables
• imbalance leads to 4 increase in occupancy of
  larger stage

25
Summary

• Analysis of a circular pipeline architecture for
  trie based IP lookup
• Goals
• Minimize memory requirement
• Maximize pipeline utilization
• Handle updates efficiently
• Design
• Decoupling of stages from maximum prefix length
• LPC analysis
• Nodes to stages mapping heuristic
• Evaluation
• On real BGP tables
• Good memory utilization and ability to keep
  40Gbps line rate through small memory banks

26

• Thank you!

27
Addressing the worst case

• Observations
• We addressed practical datasets
• Worst case tries may have long and skinny
  sections difficult to split
• Idea adaptive CAMP
• Split trie into parent and child subtries
• Map the parent sub-trie into pipeline
• Use more pipeline stages to mitigate effect of
  multiple loops around pipeline