EnergyAware Mapping for Tilebased NoC Architectures Under Performance Constraints

1
Energy-Aware Mapping for Tile-based NoC
Architectures Under Performance Constraints

• J. Hu and R. Marculescu
• CMU
• ASPDAC03

2
Introduction

• Design flow for the tile-based architecture
• a graph of concurrent tasks
• assignment and scheduling
• Tile mapping (topological placement)

3
Platform Description

• n x n tiles
• XY routing

4
Energy Model
energy consumed when one bit of data is
transported
energy consumed by switch
energy consumed on link
energy consumption of sending one bit of data
from tile ti to tile tj
the number of routers the bit passes
5
Problem Formulation

• Application Characterization Graph (APCG)
  gG(C,A)
• Vertex ci IP
• Arc ai,j communication from ci to cj
• v(ai,j) communication bits from ci to cj
• b(ai,j) minimum bandwidth (bits/sec)
• Architecture Characterization Graph (ARCG)
  gG(T,P)
• Vertex ti tile
• Arc pi,j routing path
• e(pi,j) average energy consumption
• L(pi,j) set of links that make up the path pi,j

6
Problem Formulation
7
Significance of the Problem

• Saves 50 energy in 3x3 tiles
• Savings increase as the system size scales up.
• Mapping problem is a constrained quadratic
  assignment problem gt NP-hard
• Propose an efficient branch-and-bound algorithm

8
The Algorithm of Energy-Aware Mapping
9
The Algorithm of Energy-Aware Mapping

• Def 3 The cost of a node is the energy consumed
  by the communication among those IPs that have
  already been mapped.
• Def 4 Let M be the set of vertices in the APCG
  that have already been mapped. A node is called a
  legal node if and only if, for any link lk, it
  satisfies the following condition

10
The Algorithm of Energy-Aware Mapping

• Def 5 The Upper Bound Cost (UBC) of a node is
  defined as a value that is no less than the
  minimum cost of its legal, descendant leaf nodes.
• Def 6 The Lower Bound Cost (LBC) of a node is
  defined to be the lowest cost that its descendant
  leaf nodes can possibly achieve.

11
The Branch-and-Bound Algorithm

• Branch
• unexpanded node is selected from the tree
• The next unmapped IP is enumeratively assigned to
  the set of remaining unoccupied tiles
• The corresponding new child nodes are generated
• Bound
• A node can be trimmed away w/o further expansion
  if either its cost or its LBC is higher than the
  lowest UBC that has been found during the
  searching

12
The Branch-and-Bound Algorithm

• UBC calculation
• The next unmapped IP ck with the highest
  communication demand
• LBC calculation

13
Speed-up Techniques

• IP ordering
• so that IPs with higher communication demand will
  mapped earlier
• Priority queue
• The lower the cost of the node, the higher the
  priority the node has for branching
• Decrease the minimum UBC
• Symmetry Exploitation
• For the system with 16 tiles, we only need to
  investigate those nodes which map the first IP to
  the tiles denoted by (0,0),(0,1) and (1,1)

14
Pseudo code
When the length of the PQ reaches a threshold
value, strict criteria are applied.
15
Experimental results
16
Video/Audio Application
17
Conclusion

• The mapping problem for regular tile-based
  architectures
• Efficient algorithm
• Minimize the total communication energy
• Future work
• Combine the IP selection and the task
  partitioning/scheduling into this framework