A Dynamic Link Allocation Router

23/4/19Advanced Processor Group The School of Computer Science

A Dynamic Link Allocation Router

Wei Song, Doug Edwards

Advanced Processor Group

The University of Manchester


Overview

• Network-on-a-Reconfigurable-ChipNetwork-on-a-Reconfigurable-Chip

• The Dynamic Link Allocation Flow control method

• The Dynamic Link Allocation Router (DyLAR)

• Conclusion


The NoRC Platform

Network Interface

Router

Processor

• NoRC: network on a reconfigurable chip

• Running multimedia applications

• Connection oriented • Stochastic routing algorithm• GALS: fully asynchronous rout

ers linked by CHAIN


Connection Oriented Routing

• Flit Definitions

Master Slave

Tim

e

Request to reserve a path

OK ACK

Data transmissionsData transmissionsData transmissionsData transmissionsData transmissionsData transmissions

False ACK

Data transmissions (end)


False Ack

data request content flit type flit header

data flit type flit header

Request Flit

Other Flits


The High Retry Rate

0 100 200 300 4000

2000

4000

6000

8000

10000

12000

14000

16000

Ave

rage

Fra

me

Late

ncy

(ns)

Frame Injection Rate (kfps)

Simulation results of a 6x6 NoC with 12 functions in network.

0 100 200 300 400

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Loss

Rat

e

Frame Injection Rate (kfps)

Flit Level Loss Rate Frame Level Retry rate

7%

Virtual Channels are required to reduce the retry rate.Virtual Channels are required to reduce the retry rate.


Overview

• Network-on-a-Reconfigurable-Chip

• The Dynamic Link Allocation Flow contThe Dynamic Link Allocation Flow control methodrol method


• Conclusion


Major Design Targets

• Implement some kind of virtual channels

• Increase the bandwidth of CHAIN links

• Reduce the area and power of the router


Increase the bandwidthC

C

C

C

C

C

C

C

C

C

C

C

C

C

C

C

C C

C

C C

C

DI00

DI01

DI10

DI11

DI20

DI21

DI30

DI31

DO00

DO01

DO10

DO11

DO20

DO21

DO30

DO31

ACKI ACKO

Dual-rail 0

Dual-rail 1

Dual-rail 2

Dual-rail 3

Asynchronous Links work better with the lower wire count.Asynchronous Links work better with the lower wire count.


Increase the bandwidth

A B C


Increase the bandwidth

A B C

A B C

Spatial division multiplex (SDM) is a good choice for Spatial division multiplex (SDM) is a good choice for asynchronous NoCs.asynchronous NoCs.


Problems of SDM

SDM has the low bandwidth efficiency.SDM has the low bandwidth efficiency.

Master Slave

Tim

e


OK ACK

Data transmissionsData transmissionsData transmissionsData transmissionsData transmissionsData transmissions

False ACK

Data transmissions (end)


False Ack

N1 N2 N3

Spare sub-link


Problems of SDM

N1 N2 N3

N1 N2 N3


Dynamic Link Allocation

• Divide the sub-link allocation apart from the path reservation

• Allocate idle sub-link to active communications that reserved this link

• All communications fairly compete for the bandwidth


Overview



• The Dynamic Link Allocation Router (DThe Dynamic Link Allocation Router (DyLAR)yLAR)

• Conclusion


Dynamic Link Allocation Router (DyLAR)

Tran Control

Tran Control

Tran Control

Arbiter

Output Buffer

Output Buffer

Output Buffer

Data Switch

Request Switch

Input Buffer

Input Buffer

Input Buffer

DyLAR Router


Path Reservation Stage

Tran Control

Tran Control

Tran Control

Arbiter

1

2

3

4


Data Transmission Stage

Tran Control

Tran Control

Tran Control

Arbiter


Head-of-line (HOL) Problem

00 01 02 03

10 11 12 13

20 21 22 23

30 31 32 33

00 01 02 03

10 11 12 13

20 21 22 23

30 31 32 33


Backpressure

tran_

ctl

tran_

ctl

data switch

request switch

arbiter

tran_

ctl

tran_

ctl

data switch

request switch

arbiter


Backpressure

tran_ctl

tran_ctl

data switch

request switch

arbiter

tran_ctl

tran_ctl

data switch

request switch

arbiter


Overview




• ConclusionConclusion


Conclusion

• Contribution– A new flow control method– Implement the first asynchronous spatial division rout

er• Advantages

– Smaller latency under zero load– Larger overall throughput under heavy load– Smaller retry rate (smaller power consumption)

• Problems– An extra request switch in each router– Extra control logic– Increase the latency to pass a router


Thank You!

Questions?


Detailed Procedures

– Request– Path reserved– Sending data– Release path


Request Procedure

tran_ctl

tran_ctl

data switch

request switch

arbiter


Request Procedure

tran_ctl

tran_ctl

data switch

request switch

arbiter


Request Procedure

tran_ctl

tran_ctl

data switch

request switch

arbiter


Request Procedure

tran_ctl

tran_ctl

data switch

request switch

arbiter


Request Procedure

tran_ctl

tran_ctl

data switch

request switch

arbiter


Request Procedure

tran_ctl

tran_ctl

data switch

request switch

arbiter


OK Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter


OK Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter


OK Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter


OK Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter


OK Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter


OK Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter


Data Flits

tran_ctl

tran_ctl

data switch

request switch

arbiter


Data Flits

tran_ctl

tran_ctl

data switch

request switch

arbiter


Data Flits

tran_ctl

tran_ctl

data switch

request switch

arbiter


Data Flits

tran_ctl

tran_ctl

data switch

request switch

arbiter


Data Flits

tran_ctl

tran_ctl

data switch

request switch

arbiter


Data Flits

tran_ctl

tran_ctl

data switch

request switch

arbiter


Data Flits

tran_ctl

tran_ctl

data switch

request switch

arbiter


False Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter


False Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter


False Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter


False Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter


False Ack

tran_ctl

tran_ctl

data switch

request switch

arbiter

Documents

A Dynamic Link Allocation Router