Route Lookup

1
Route Lookup

• CSE 525 Advanced Networking
• Computer Science and Engineering Department
• Winter 2004

2
Papers covered

• Adaptive Data Structures for IP Lookups by
  Ioannis Ioannidis, Ananth Grama, Mikhail Atallah
• XOR-based schemes for fast parallel IP lookups by
  Giancarlo Bongiovanni, Paolo Penna

3
Motivation

• Addresses the Need of efficient route look up,
  which is motivated by
• Exponential growth of internet more entries in
  routers
• Increase in the transmission speed faster lookup

4
Contribution 1

• Addresses the problem of generalizing LC-tries
  and Extensible hashing to provide guarantees for
  memory sub-optimal structures
• This paper present a greedy algorithm for this
  problem

5
Related Work

• Hardware and cache based solution
• Look up acceleration using memory placement and
  pipelining
• Protocol based solutions
• All demands modification and raise complexity
• Does not avoid prefix matching problem

6
Approach

• Combining Trie and Extensible hashing
• e.g. If Trie is taking 6MB of memory out of 8MB
  router memory, how it should be modified to make
  best use of 2MB of excess memory?
• Added access statistics on node to determine the
  frequency

7

• Path compression remove internal node with 1
  child and put skip value
• Level compression replace the i highest complete
  levels of the binary trie with a single node of
  degree 2i

8

• Each node is one word long
• First 5 bit represent branching factor
• Next 7 bit represent Skip Value
• Reaming 20 bit represent pointer to left most
  node
• Assumption Store children of a node in a
  consecutive location for space efficient solution

9
(No Transcript)
10
(No Transcript)
11
(No Transcript)
12
Challenge

• The challenge is to identify parts of the trie
  that can be expanded into hash tables without
  exceeding available memory while yielding most
  benefit in terms of memory accesses.
• This problem is a variation of Knapsack problem
  with a complexity of having non-static items

13
Experiment Result

• Did experiment on Mae-West routing table size
  little over 5 105
• Applied hashing at level 16 and 24
• For a increase of 10 memory they achieve an
  increase of 25 in performance

14
Experiment Result
15
Future Work

• Assumption changes in routing table doesnt
  demand re-calculation of entire structure but
  the rate which rate which routing tables get
  updated, this may be most important open problem
  in this
• Improvement of the scheme concerning time
  complexity and approximation ratio

16
Contribution 2

• Proposes a novel scheme for fast IP look up in
  the routers, which is handing the data _at_ gigabit
  speed
• The Algorithm exploits parallelism in routers,
  that reduces size of routing table and suggest
  parallel lookup

17
Related work

• Optimized forwarding table to fit in to L2 cache
• Multiprocessor routers by assigning different
  packets to different processor
• Hardware based solution
• L2 should grow as fast as routing table
• Hardware based solution not scalable and too
  expensive

18
Approach

• Algorithm uses more than one processors in order
  to reduce the size of the routing table
• Take a routing table T split into two T1 and T2
  then a packet is processed in parallel by two
  processors having T1 and T2 in their L2 cache.
  The final result is obtained by combining the two
  result

19
Approach
20
Approach

• This paper talks about splitting a table in such
  a way that combination of two partial result done
  in a simplest way the XOR of two sequence
• Lookup (a, T) Lookup (a, T1) ? Lookup (a, T2)

21
Split phase
22
Algorithm

• Split Phase Divide table in two or more table
• Compact phase T1 and T2 may contains redundant
  information
• Update It doesnt not yield an overhead in the
  process of updating
• Label change
• Entry insertion/deletion

23
(No Transcript)
24
Experiment Result

• Did experiment on real routing table of 5
  different routers
• Tup contains all non-leaf node ltlt T
• Resolves the issue of memory size of the
  structure for the IP lookups
• T/Tup with no exceptions
• Simplifies the BMP problem and yield the data
  structure of smaller size

25
Conclusion and Future Work

• Possibility to optimize other parameters
• Any advantage of splitting table into more then
  two tables
• Problem is still Open, Does any of the existing
  solutions get simpler or more efficient because
  of this?

26

• Questions?