Design of Bandwidth Bandwidth Aware Aware and and Design of - - PowerPoint PPT Presentation

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

1

Design of Design of Bandwidth Bandwidth Aware Aware and and Congestion Congestion Avoiding Avoiding Efficient Efficient Routing Routing Algorithms Algorithms for for NoCs NoCs Platforms Platforms

• M. Palesi1, G. Longo1, S. Signorino1, R.

Holsmark2, S. Kumar2, V. Catania1

1DIIT, University of Catania, Italy

{mpalesi, vcatania}@diit.unict.it

2Jőnkőping University, Sweden

{Rickard.Holsmark, Shashi.Kumar}@jth.hj.se

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

2

Outline Outline

Motivation Application specific scenario Bandwidth aware routing algorithm Experiments and Results Architectural implications Conclusions

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

3

Limitations Limitations of

• f Current

Current Routing Routing Algorithms Algorithms

Efforts biased toward performance Side effects like congestion ignored

Estimation and control of congestion is difficult in general Partially tackled by the selection function

Designed only for specific network topologies

Routing Algorithm Design Routing Algorithm Design

Network Topology Network Topology

Buffer

• ccupation

S

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

4

Application Application Specific Specific Scenario Scenario

Information available about

Tasks which communicate and tasks which do never communicate

After task mapping Information about network nodes which communicate

Concurrent/non concurrent communications Communication bandwidth requirement for different pais

Many opportunities

Improving performance (e.g., maximize routing adaptivity) Simplify the estimation/control of congestion Design more effective selection policies

T1 T4 T3 T2 Tn

Network Topology Network Topology Application Specification Application Specification

AS Routing Algorithm Design AS Routing Algorithm Design

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

5

Channel Channel Dependency Dependency Graph Graph

Topology Graph Channel Dependency Graph

l12 l21 l32 l45 l54 l56 l65 l41 l14 l52 l23 l25 l63 l36

P1 P2 P4 P5

l12 l21 l45 l54 l41 l14 l52

P3 P6

l23 l32 l56 l65 l25 l63 l36

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

6

Application Application Specific Specific CDG CDG

l12 l21 l45 l54 l41 l14 l52 l23 l32 l56 l25 l63 l36 l65

• Channel Dependency Graph
• P1 P3

P4 P3 P1 P6

P1 P2 P4 P5

l12 l21 l45 l54 l41 l14 l52

P3 P6

l23 l32 l56 l65 l25 l63 l36 T6 T3 T1 T5 T4 T2 Communication Graph Topology Graph

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

7

Application Application Specific Specific CDG CDG

l12 l21 l45 l54 l56 l65 l41 l14 l52 l23 l32 l25 l63 l36

Application Specific Channel Dependency Graph

P1 P2 P4 P5

l12 l21 l45 l54 l41 l14 l52

P3 P6

l23 l32 l56 l65 l25 l63 l36 T6 T3 T1 T5 T4 T2 Communication Graph Topology Graph

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

8

APSRA Design APSRA Design Methodology Methodology

Application to be mapped Application to be mapped T1 T4 T3 T2 Tn

Communication Graph P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11 P12 P13 Network Topology Mapping Function Mapping Function

APSRA APSRA

Routing Tables

Compression Compression

Compressed Routing Tables

Memory budget Memory budget [Palesi, et al., CODES+ISSS’06] [Palesi, et al., SAMOS’06]

GOAL Maximize adaptivity

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

9

Bandwidth Bandwidth Variation Variation: Multimedia : Multimedia Example Example

Communication bandwidth ranges from 10 to 500 MB/s

Source: Hu and Marculescu, TCAD 24(4), 2005

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

10

Contributions Contributions

Design APSRAs which

are highly adaptive

Translates into high performance, in general

uniformly distribute traffic over the network allow maintenance of load of links under a given bandwidth threshold

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

11

1st 1st Phase Phase

Removing a dependency d

Removing all the paths which use d

As soon as a path is removed

The fraction of bandwidth it transports must be redistributed between the remaining paths

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

12

1st 1st Phase Phase

Removing a dependency d

Removing all the paths which use d

As soon as a path is removed

The fraction of bandwidth it transports must be redistributed between the remaining paths d d due to Path 3

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

13

1st 1st Phase Phase

Removing a dependency d

Removing all the paths which use d

As soon as a path is removed

The fraction of bandwidth it transports must be redistributed between the remaining paths

Remove d

Path 3 must be removed 25 MB/s must be redistributed between the remaining paths 66 MB/s 33 MB/s 33 MB/s 66 MB/s 33 MB/s 66 MB/s 33 MB/s 33 MB/s 33 MB/s

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

14

1st 1st Phase Phase

∑

∈

− × × =

C c

d c PT c P c P d c PT c B d ) ) , ( ) ( ( ) ( ) , ( ) ( ) cost(

2 2

Removing a dependency d

Removing all the paths which use d

As soon as a path is removed

The fraction of bandwidth it transports must be redistributed between the remaining paths

Strategy

Removing the dependency d which minimizes the overhead of bandwidth that should be allocated to the remaining paths that do not use d

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

15

2nd 2nd Phase Phase

After Phase 1 we have a routing function which is

Deadlock free Provides more adaptivity to communications characterized by higher communication bandwidth

But…

It is possible that the agregate bandwidth (AB) on some links exceeds the capacity of that link

“Some” routing paths passing on that link, must be removed to reduce the AB on that link down to the link capacity

∀ link l → AB(l) ≤ Cap(l)

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

16

Routing Routing and and Selection Selection

Body H

Inputs

Adaptive Routing Function Adaptive Routing Function Selection Function Selection Function

Routing Algorithm Neighbours information Locally stored information Outputs

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

17

Load Load Balancing Balancing Selection Selection Function Function

Dst … … … … … … … … … … nd E S 0,75 0,25 … … … … … … … … … … Admissible Out Selection Prob Overhead

The probability to select output channel l is proportional to the number of admissible paths starting from l and that can be used to reach the destination

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

18

Experimental Experimental Setup Setup

8x8 mesh based NoC Buffer size 4-flits Simulation time 100,000 cycles Warmup time 20,000 cycles Traffic injection distribution

Poisson (for synthetic traffic scenarios) Self-similar (for MMS traffic)

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

19

Stdev Stdev Reduction Reduction

Percentage reduction of standard deviation of the aggregated bandwidth in network links

0% 5% 10% 15% 20% 25% 30%

U n i f

• r

m B i t

• r

e v e r s a l B u t t e r f l y S h u f f l e T r a n s p

• s

e 1 T r a n s p

• s

e 2 H

• t
• s

p

• t

_ C H

• t
• s

p

• t

_ T R M M S

Stdev reduction (%)

APSRA-BW APSRA-BWL

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

20

Aggregate Aggregate bandwidth bandwidth

Aggregate bandwidth per link for a 9x9 mesh-based NoC under uniform traffic

APSRA APSRA-BWL

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

21

Delay Delay Reduction Reduction

Average delay reduction obtained when APSRA-BW and APSRA-BWL are used taking APSRA as a baseline

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

22

Average Average Delay Delay Variation Variation

Average delay variation under uniform traffic for different ranges of communication bandwidth

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

23

Links Links Utilization Utilization

Links utilization under uniform traffic for APSRA and APSRA-BWL

→ → → →

APSRA-BWL APSRA

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

24

Router Router Architecture Architecture

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

25

Area Area Overhead Overhead

5000 10000 15000 20000 25000 30000 35000 40000 45000 Arbiter XBar FIFO WHRT Ctrl Routing Function Selection Function

Area (um^2)

RND LB

+5% overall area overhead

NoCS 2008 Newcastle University, UK – 7th-11th April 2008

26

Conclusions Conclusions

Bandwidth aware routing algorithm

Highly adaptive Reduces the variation of load in the network links Ensures that the link bandwidth is not violated

Evaluate the idea for irregular mesh topology