Aethereal Presentation

1
A MultiPath Routing Strategy with Guaranteed
InOrder Packet Delivery and FaultTolerance for
Networks on Chip S. Murali - Stanford
University, USA D. Atienza, G. De Micheli -
EPFL, Switzerland L. Benini - University of
Bologna, Italy
2
Introduction

• Communication in SoCs rapidly increasing

• Communication architecture is bottleneck
• Scalability, delay, power
• Scaling Interconnect is becoming unreliable
• Networks on Chips (NoCs) needed to tackle all
  these issues.

3
Problem Motivation

• Single path Vs multiple paths for routing

• Multi-path routing advantages
• Evenly spreading traffic
• Fewer traffic bottlenecks, lower NoC
  frequency-gt potentially lower power
• Spatial redundancy support for fault-tolerance
  possible

However, most NoCs dont support multi-path
routing !
4
But, why ?

• Out-of-order packet delivery is a big challenge
• Different path lengths, congestion levels

2
1

• Re-order buffers needed at receivers
• Have large area, power overhead
• Deterministic sizing is infeasible

5
But, why ?

• Packet dropping needed handle buffer-full
  situation

• Macro-networks end-to-end ACK/NACK error control
• Use same protocol to recover from packet dropping
• Most NoCs do not support end-to-end ACK/NACK

Packet re-ordering is a big challenge to overcome
in NoCs!
6
Contributions of this work

• We present
• Multi-path routing that guarantees
  in-orderpacket delivery
• Router architecture to support the scheme
• A methodology to split application traffic
  acrossthe multiple paths
• Integrate support for temporal and spatial
  fault-tolerance with the multi-path design
  methodology

7
Previous Work

• NoC Multi-path routing motivated in Murali DATE
  04
• Several works on NoC mapping consider routing Hu
  DATE 03, Murali DAC 04, Hannson ISSS 05
• Routing schemes well studied in multiprocessor
  networks Dally 04
• Reliable NoC design based on routing Marculescu,
  Piretti ISVLSI 03, Manolache DAC 05
• IBM vulcan multi-processor network support for
  spatial redundancy

8
Multi-path Routing withIn-order Delivery

• Packets are routed on multiple non-intersecting
  paths
• Partially or fully non-intersecting
• In-order in a path, can be out-of-order across
  paths

• Re-ordering needed at path re-convergent nodes
• Change switch architecture to support re-ordering

9
Switch Architecture

• Use a look-up table at each switch
• store packet id of next packet to be received for
  each flow
• Grant access to packet that matches the id
• When a packet is sent, increment the look-up
  table id
• Stall out-of-order packets on switch input
  buffers

10
Switch Architecture (II)

• Packets on a path are in-order
• stalled packets do not block the right packet
• The right packet is on a different path
  thanstalled packets
• When integrated with NoC flow control, leads
  tono packet dropping, unlike the traditional
  schemes!
• Small overhead on the switches(a look-up table
  used)

11
Splitting Traffic across the paths

• Algorithm developed to find such paths for given
  NoC
• Splitting traffic modeled as a Linear Program
  (LP)
• Objective minimize maximum bandwidth requirement
  
• Split traffic flows across non-intersecting paths
• Satisfy latency constraints for different flows

12
Fault-tolerance

• Consider recovery from transient and permanent
  errors
• Transient Errors Replicate critical traffic
• Permanent Errors Allocate enough bandwidth on
  paths to account for permanent link failures
• Replication designed to meet application mean
  time to failure
• Fast and easy to solve the formulation

13
Experimental Results

• Compared dimension-ordered, minimum path and
  proposed multi-path scheme
• Proposed scheme leads to 30.51 reduction in
  power
• Also includes overhead (packet ids, luts) of
  multi-path scheme
• Savings due to reduction in bandwidth
  requirements

14
Experimental Results

• MPEG application mapped onto mesh NoC using an
  existing tool (SUNMAP)
• Multi-path scheme needs lower frequencies to
  achieve same packet latencies as other schemes
• This translates to lower power requirements

MPEG Application
15
Experimental Results

• Bit-Error Rates (BER), amount of critical traffic
  increases gt power consumption increases
• Accounting single path permanent failure gt 2.33
  x increase in power consumption
• Power consumption Vs reliability trade-offs
  possible

MPEG Application
16
Summary

• Multi-path routing desirable bandwidth,
  reliability
• Out-of-order packet delivery big challenge
• Presented multi-path routing with guaranteed
  in-order packet delivery
• Design methodology and router architecture
  support
• Fault-tolerance against transient, permanent
  errors
• Trade-off studies of fault-tolerance and power
  consumption possible

17
Thank You