Designing Scalable Routers with a New Switching Architecture

1
Designing Scalable Routers with a New Switching
Architecture

• Department of Computer Science, Tsinghua
  University, Beijing, P.R. China
• Zuhui Yue
• Youjian Zhao
• Jianping Wu
• Xiaoping Zhang

Proceedings of the Joint International Conference
on Autonomic and Autonomous Systems and
International Conference on Networking and
Services (ICAS/ICNS 2005)
2
Outline

• The basic CR (Cellular Router) architecture
• Design of the line card
• Improvement of the basic architecture
• Minimal routing algorithms in the CR architecture
• Performance evaluation

3
The basic CR architecture -A single rack with a
layer of line cards
Basic Element Numbers(BENs)
Basic Element
4
The basic CR architecture -A single rack with a
layer of line cards
5
The basic CR architecture -A single rack with
multilayer line cards
In each rack, a line card can be identified with
(l, st ) . l represents the layer where the
line card is located st represents the BEN of
the line card
6
The basic CR architecture -Multi-rack system
Cobweb architecture
A line card can be identified as (CN, l, st) CN
Cobweb Number (CN).
7
The basic CR architecture -Multi-rack system

• Step1 define the central point as Polar Point
  (PP).
• Step2 there are six axes which start from the
  PP. They are defined as 0-axis, -p/3-axis,
  -2p/3-axis, p-axis, 2p/3-axis and p/3-axis
  respectively.
• Step3 first place one rack on the PP. Then we
  place the second rack on the p/3-axis. After we
  have placed six racks around the PP clockwise
• Step4 define the PP as Circle 0 and the six
  racks around it as Circle 1, etc. We call this
  architecture as Cobweb architecture.

8
The basic CR architecture -Multi-rack system
9
Design of the line card
The line card structure in CR architecture
10
Design of the line card

• CR architecture distributes the switching task
  into each line card.
• The inner switching fabric can be normal
  crossbar.
• We can introduce separated queues, named Virtual
  Channel Queue (VCQ), for each output.

11
Improvement of the basic architecture Fault
tolerance considerations
Node failure in the BE
12
Improvement of the basic architecture Mirror
Points
Modification of the BE
13
Improvement of the basic architecture Mirror
Points
Mirror Points in a system with two circles of
racks
14
Minimal routing algorithms in the CR architecture

• Each link is unidirectional, so there are two
  links between any adjacent nodes. We define the
  length of each link as 1.
• We further assume that the architecture uses
  storeand-forward flow control with each node
  having buffers of infinite length.

15
Minimal routing algorithms in the CR architecture

• s source node.
• d destination node.

16
Minimal routing algorithms in the CR
architecture- Selection of next-hop node

• Every node can be identified as (?, ?)
• ? the length of OP
• ? is the value of ? POX.

17
Minimal routing algorithms in the CR
architecture- Selection of next-hop node
Method based on the PP
18
Performance evaluation

• We define the length of each packet as 1.
• The destination of each arrival packet is
  randomly chosen.
• The queues on each node are organized as VCQs and
  the length of each queue is 100.
• When there are two optional next-hop nodes we
  always select the right neighbor.

19
Performance evaluation
Throughput of RNF algorithm