Dynamic Pipelining: Making IPLookup Truly Scalable

1
Dynamic Pipelining Making IP-Lookup Truly
Scalable

• Jahangir Hasan and T.N. Vijaykumar
• School of Electrical and Computer Engineering,
  Purdue University
• ACM SIGCOMM05, August 21-26, 2005, Philadelphia,
  Pennsylvania, USA

2
Five challenges of scalability

• Routing table size
• Lookup throughput
• Implementation cost
• Power dissipation
• Routing-table update cost

3
Pipelined IP-Lookup

• Hardware-Level Pipelining (HLP)
• Data-Structure-Level Pipelining (DLP)

4
Hardware-Level Pipelining (HLP)

• A Pipelined Memory Architecture for High
  Throughput Network Processors 15
• Timthy Sherwood, George Varghese, and Brad
  Calder, UCSD
• 30th annual International Symposium on Computer
  Architecture (ISCA03), IEEE

5
(No Transcript)
6
(No Transcript)
7
(No Transcript)
8
(No Transcript)
9
Data-Structure-Level Pipelining (DLP)

• Fast Incremental Updates for Pipelined Forwarding
  Engines 1
• Anindya Basu and Girija Narlikar, Bell
  Laboratories
• IEEE INFOCOM03 and IEEE/ACM Transactions on
  Networking, Vol. 13, No.3, June 2005

10
Controlled Prefix Expansion (1)

• Fast Address Lookups Using Controlled Prefix
  Expansion 17
• V. Srinivasan and G. Varghese, Washington
  University in St. Louis
• ACM Transactions on Computer Systems, Vol. 17,
  No. 1, February 1999, pages 1-40

11
Controlled Prefix Expansion (2)

• Optimal Expanded Trie with Fixed Strides
• Optimal Expanded Trie Using Varying Strides

12
Leaf-pushed Fixed Strieds

• Only tries with fixed strides are considered.
  Although variable-stride tries could be more
  memory efficient, they are difficult to maintain
  during incremental updates. Since each node can
  have a different stride, there is no
  straightforward way to determine the strides for
  trie nodes that are created when new prefixes are
  added.
• We focus on leaf-pushed tries. Although updates
  to leaf-pushed tries can result in more pipeline
  writes, leaf pushing allows for a higher pipeline
  throughput and more efficient use of the pipeline
  memory. All our optimizations, however, are also
  applicable to nonleaf pushed tries.

13
(No Transcript)
14
(No Transcript)
15
(No Transcript)
16
Characteristics of the BGP routing table
17
Characteristics of the BGP routing table
18
Characteristics of the BGP routing table
19
Controlled Prefix Expansion for Optimal Expanded
Trie with Fixed Strides

• Let Tj,r be the optimal memory requirement (in
  terms of the number of the trie nodes) for
  covering bit positions 0 through j using r trie
  levels (assuming that the leftmost bit position
  is 0).
• Let nodes(i) be the number of nodes in the 1-bit
  trie at level i.

20
(No Transcript)
21
(No Transcript)
22
(No Transcript)
23
A New Algorithm for Pipelined Architectures
24
A New Algorithm for Pipelined Architectures
25
(No Transcript)
26
(No Transcript)
27
Reducing Write Bubbles

• A. Separating Out Updates to short Routes
• B. Node Pullups
• C. Eliminating Excess Writes
• D. Caching Deleted Subtrees

28
A. Separating Out Updates to short Routes

• Storing all short routes, of up to 8 bits length,
  in a separate table with 28 256 entries.

29
(No Transcript)
30
B. Node Pullups
31
(No Transcript)
32
C. Eliminating Excess Writes add
33
C. Eliminating Excess Writes withdraw
34
D. Caching Deleted Subtrees