HardwareBased IP Routing Using Partitioned Lookup Table

1
Hardware-Based IP Routing Using Partitioned
Lookup Table

• Mohammad J. Akhbarizadeh
• Student Member, IEEE,
• Mehrdad Nourani,
• Senior Member, IEEE

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 13, NO.
4, AUGUST 2005
2
Outline

• THE IFPLUT ALGORITHM
• THE IFPLUT ARCHITECTURE
• EXPERIMENTAL RESULTS
• PS IFPLUTgt IP Forwarding based on Partitioned
  Lookup Table

3
THE IFPLUT ALGORITHM Notations, and
Definitions

• P P1,P2,PG is the set of prefixes G
  collected by a backbone router.
• There are N output ports, let Q 1,2,,N
  denote the set of these port indices.
• FT (Forwarding Table) is an organized ei (pi ,
  qi) set of pairs, where pi ? P and qi ? Q.
• IPNi IP address
• leni Length of the prefix

4
THE IFPLUT ALGORITHM Notations, and
Definitions

• Definition 1 Two prefixes are disjoint if none
  of them is a prefix of the other.
• Definition 2 A prefix pi ? P is called an
  enclosure , if there is at least one prefix pj ?
  P (j?i and leni lt lenj) such that pi is a prefix
  for pj.
• Definition 3 If both prefixes pi and pj point to
  the same egress q and pi is an enclosure of pj
  then we call an identical enclosure of pj.
• Definition 5 If there is no enclosure in , then
  is a disjoint set.

5
THE IFPLUT ALGORITHM
SAMPLE SET OF FORWARDING TABLE ENTRIES (IPv4)
6
THE IFPLUT ALGORITHM

• For a forwarding table FT,FT is defined as
  follows.
• If (pi, q), (pj, q) ? FT and pi is an identical
  enclosure of pj then FT is generated by
  replacing (pi, q) with d entries (pi1, q) through
  (pid, q), such that
• (A) pi1, pi2, .. pid and pj are disjoint
• (B) pi1 U pi2 U pid U pj pi
• Here is the distance between and (i.e., the
  difference between their length).
• Consequently, FT contains no identical
  enclosures.

7
THE IFPLUT ALGORITHM

• Partitioning Rule We partition set FT into N
  subsets PFT1,PFTN such that for partition we
  have
• Having partitioning rule applied to set FT,
  every PFTk is a disjoint set.
• Given an IP address to be looked up, searching a
  PFT will result in zero or one match.

8
THE IFPLUT ALGORITHM
9
THE IFPLUT ALGORITHM

• The percentage of identical enclosures cannot be
  too high because a prefix must 1) be an
  enclosure and 2) point to the same egress as its
  child.
• This is based on our empirical analysis of the
  routing tables of three major network access
  points, AADS, MAE-WEST, and PacBell.

10
THE IFPLUT ALGORITHM
11
THE IFPLUT ALGORITHM
12
THE IFPLUT ALGORITHM
13
THE IFPLUT ARCHITECTURE Basic IFPLUT
forwarding architecture
Basic IFPLUT forwarding architecture.
14
THE IFPLUT ARCHITECTURE

• If a particular egresss PFT becomes much larger
  than the other PFTs we say the imbalance
  distribution of prefixes has occurred. This may
  become a serious problem as it may lead to
  under-utilization or even malfunctioning of
  IFPLUT.
• We decided to break the one-to-one correspondence
  between PLUs and partitions and have a big pool
  of search units and let a management unit
  dynamically assign them to partitions whenever
  the need exists.

15
THE IFPLUT ARCHITECTURE

• In this design, there is a pool of distinct
  memory units. Each partition (corresponding to an
  egress port) will be assigned zero or more memory
  units. To keep track of which unit is assigned to
  which partition, a register is accompanying each
  unit.
• This register will store the identification (Id)
  of the associated partition, which can simply be
  the corresponding egress port index.

16
THE IFPLUT ARCHITECTURE Improved IFPLUT
forwarding architecture
Improved IFPLUT forwarding architecture.
17
THE IFPLUT ARCHITECTURE Concept of the
search unit
Concept of the search unit.
18
THE IFPLUT ARCHITECTURE
selector
Modified selector block.
19
THE IFPLUT ARCHITECTURE
Comparator
Comparator slice
20
THE IFPLUT ARCHITECTURE
Macro selector

• The critical path of the selector goes through
  only one gt block and it is very fast but,
  because of a large amount of interconnect needed
  between the sub-blocks of the selector, this unit
  is not scalable beyond a limit.

21
THE IFPLUT ARCHITECTURE
Update

• To keep track of the identical enclosures we add
  a hashing table to the IFPLUT architecture.
• Each entry of this table contains two fields a
  prefix and its Id. The prefix part stores the
  identical enclosure prefix and the Id part links
  it to the expanded prefixes that represent it in
  the IFPLUT.
• One proper choice for the Id value would be the
  prefix length. We refer to this table as the
  Expansion Table or simply ExpTab.

22
THE IFPLUT ARCHITECTURE
Update

• The Id field is also added to each entry in the
  IFPLUT. In PFTk, prefix pi comes with idi . The
  value of this Id is zero for all original
  prefixes. For an expanded prefix, the Id field
  copies the value of the longest identical
  enclosure in the ExpTab that parents it.

23
THE IFPLUT ARCHITECTURE
Update

• Add Say (p1, qk) is going to be added to our
  IFPLUT. The new route prefix will be written into
  PFTk. Two special situations should be taken care
  of

24
THE IFPLUT ARCHITECTURE
Update
25
THE IFPLUT ARCHITECTURE
Update

• Remove To remove prefix p1 ? PFTk from IFPLUT
  the system has to take care of two special
  situations as well

26
THE IFPLUT ARCHITECTURE
Update
27
THE IFPLUT ARCHITECTURE
Metrics and Tradeoffs

• N is the number of ports
• M is the number of SUs
• S is the size of each SU memory
• G is the expected size of the FT
• For given M and G values, the size of each SU
  memory (S) must at least be.

28
THE IFPLUT ARCHITECTURE
Metrics and Tradeoffs

• The worst case memory usage comes about when all
  partitions except one have only one entry each.
• In this case each of those N-1 partitions will
  get a single SU so that SUs will be assigned to
  N-1 partitions.
• In this scenario the rest of the FT entries
  (i.e.,G - (N-1) entries) belong to the last
  partition. So, the IFPLUT is left with G (N-1)
  entries to be fit in the remaining M - (n-1)
  SUs.
• S( M - (N-1) ) ? (G (N-1)) ? lower bound

29
THE IFPLUT ARCHITECTURE
Metrics and Tradeoffs

• To be ready for the worst case, we need to have
  the minimum value of S as
• S max G/M , (G-N1)/(M-N1)
• As G M the condition that is always true.So, S
  can be always determined as

30
THE IFPLUT ARCHITECTURE
Metrics and Tradeoffs

• Actual memory size is in total
• which is always greater than or equal to G.
• The difference between these two values (G - G)
  is the amount of memory waste that must be
  tolerated to be safe at the worst case scenario.
  So, the memory waste W is given as

31
THE IFPLUT ARCHITECTURE
Metrics and Tradeoffs

• M always has a value between N and G,
• i.e. N?M ?G. From the above equation, memory
  waste is maximum when M has its minimum value and
  it is minimum when is maximized. Hence
• Wmin 0, When M G
• Wmax (G-N)(N-1) , When MN

32
EXPERIMENTAL RESULTS

• For the purpose of experimentation, a system with
  an FT capacity of G 20000 entries was designed
  and evaluated.

IMPLEMENTATION OF COMPARATOR BLOCKS (COST
EXPRESSED AS THE NUMBER of TWO-INPUT NAND GATES)
33
EXPERIMENTAL RESULTS

• M show (number of search units, which is also
  the number of the corresponding macro selectors
  inputs)
• K (size of search units in the first stage)
• Q (size of search units in the second stage)
• N 32 Ports

34
EXPERIMENTAL RESULTS
Overhead cost of different IFPLUT designs.