hybrid on-chip data networks

Hybrid On-chip Data Networks

Gilbert HendryKeren Bergman

Lightwave Research Lab

Columbia University

2

Chip-Scale Interconnection Networks

Intel Polaris IBM Cell AMD Opteron

• Chip multi-processors create need for high performance interconnects

• Performance bottleneck of on-chip networks and I/O• Power dissipation constraints of the chip package

• > 50% of total power comes from interconnects*

* N. Magen et al., “Interconnect-power dissipation in a microprocessor,” SLIP 2004.

3

Motivation

• CMPs of the future = 3D stacking• Lots of data on chip• Photonics offers

key advantages

4

Why Photonics?

TX RX

ELECTRONICS: Buffer, receive and re-transmit at

every router.

Each bus lane routed independently. (P ∝ NLANES)

Off-chip BW is pin-limited and power hungry.

Photonics changes the rules for Bandwidth, Energy, and Distance.

OPTICS: Modulate/receive high bandwidth

data stream once per communication event.

Broadband switch routes entire multi-wavelength stream.

Off-chip BW = On-chip BW for nearly same power.

RX

TX

RX RX

TX

RX

TX RXTX TX TXTXTX TX

RX

5

Hybrid Network Premise

Optical processing difficult and limited

Source, destination routing inefficient

Use electronics for routing, optics for switching and transmission

Hybrid Circuit-Switching

6

Hybrid Circuit-Switched Networks

Step 1: Path SETUP request

Electronic SETUP Msg

Source coreDestination Core

7


Step 2: Path ACK

Electronic ACK Msg

8


Step 3: Transmit Data

Photonic Switch Use Information

9


Meanwhile: Path Contention

Path BLOCKED Msg(Backoff)

10


Step 4: Path TEARDOWN

Electronic TEARDOWN Msg

Source coreDestination Core

11


• Energy-efficient end-to-end transmission

• High bandwidth through WDM

• Electronic network still available for small control messages*

• Network-level support for secure regions

• Path setup latency• Path setup contention

(no fairness)

Pros: Cons:

* [G. Hendry et al. Analysis of Photonic Networks for a Chip Multiprocessor Using Scientific Applications. In NOCS, 2009]

Programming and Communication

13

Shared Memory

Implicit Communication

Explicit Communication

scaling

“… [ OpenMPon large systems] often performs worse than message passing due to a combination of false sharing, coherence traffic, contention, and system issues that arise from the difference in scheduling and network interface moderation”

~ ExascaleReport

14

Partitioned Global Address Space



[G. Hendry et al. Circuit-Switched Memory Access in Photonic Interconnection Networks for HPEC. In Supercomputing, Nov. 2010]

Access Method

Local Read Optical Receive

Local Write Optical send

Remote Read Electronic request, optical receive

Remote Write Optical send

Shared R/W ?

15

Message Passing



* [G. Hendry et al. Analysis of Photonic Networks for a Chip Multiprocessor Using Scientific Applications. In NOCS, 2009]

• Complex, dynamic access patterns

• Relatively larger blocks of data• Scientific computing

16

Streaming



1

2

3

4

Input DataOutput Data

Persistent optical circuits

• Embedded / specialized systems (Graphics, Image + Signal Proc.)• Execution mode of general-purpose systems (Cell Processor)

Electronic Plane

18

Electronic Router

Arbiter

…

Control Router

Data Switch

Buffer Crossbar

Buffer Cntrl

Data Path

XbarCntrlRequest Bus

Flow Control

XbarAllocation

Data Switch Allocation

Routing Logic

Credits In

XbarCntrl

Ring Cntrl

Ring Cntrl

• Low frequency operation (~ 1GHz)• 1 VC (typically)• Small buffers (64-28)• Narrow Channels (8-32)

19

Network Gateway

Core

Core

Core

Core

Tx/Rx

Net

wor

k IF

Bidirectional Waveguide

Bidirectional Electronic Channel

Control RouterElectronic Crossbar

5-port photonic switch

To/From Control plane

To/From Data plane

Se

rializ

atio

n

Driv

ers

Des

eria

lizat

ion

Re

ceiv

ers

[P. Kumar et al. Exploring concentration and channel slicing in on-chip network router. In NOCS, 2009]

External Concentration

The Photonic Plane

21

Wavelength Division Multiplexing

λ

waveguide

22

Silicon Photonic Waveguide Technology

[Vlasov and McNab, Optics Express 12 (8) 1622 (2004)]

C23(1559 nm)

C28

(1555 nm)C46

(1541 nm)C51

(1537 nm)

before injection into waveguide

after 5-cm waveguide and EDFA

[B. G. Lee et al., Photon. Technol. Lett. 20 (10) 767 (2008)]

1.28 Tb/s Data Transmission Experiment(occupies small slice of available WG BW)

100 ps

Silicon photonic waveguides provide low-power optical interconnects in CMOS-compatible platform.

Low-loss (1.7 dB/cm), high-bandwidth (> 200 nm) silicon photonic waveguides can be fabricated in commercial CMOS process.

23

Ring Resonator Operation

λ

λ

modulator/filter

Broadband spatial switch

24

Silicon Photonic Modulator and Detector Technology

[M Watts, Group Four Photonics (2008)]

[M Lipson, Optics Express (2007)]

85 fJ/bit demonstrated at 10 Gb/s Scalable to < 25 fJ/bit

18 Gb/s demonstrated

[S Koester, J. Lightw. Technol. (2007)]

Ge-on-Si Detectors: 40-GHz bandwidths 1 A/W responsivities

Receivers (detectors w/ CMOS amplifiers): 1.1 pJ/bit demonstrated at 10 Gb/s Scalable to < 50 fJ/bit

(CW) LASER

modulator detector

25

Higher Order Switch Designs

[A. Biberman, IEEE Phot. Tech. Letters (2010)]

26

On-Chip Topology Exploration

• Photonic Torus • Nonblocking Photonic Torus

[A. Shacham et al., Trans. on Comput., 2008] [M. Petracca et al. IEEE Micro, 2008]

27

On-Chip Topology Exploration

• TorusNX • Square Root

[J. Chan et al. JLT, May 2010]

28

Photonic Plane Characteristics

• Insertion Loss• Noise• Power

29

Insertion Loss and Optical Power Budget

Nonlinear Effects

WDM FactorO

pti

cal

P

ow

er B

ud

get

Worst-case Insertion Loss

Detector Sensitivity

30

Insertion Loss vs. Bandwidth

Network Size

Num

ber

of λ

Topologies

31

Simulation Results

4×46×6

8×810×10

12×12

14×14

16×16

18×18

0

10

20

30

40

50

Inse

rtio

n Lo

ss (

dB)

Topology Size (nodes)

Torus Topology

20.625.6

31.237.0

42.848.6

54.560.3

4×46×6

8×810×10

12×12

14×14

16×16

0

10

20

30

40

50

Inse

rtio

n Lo

ss (

dB)


Non-BlockingTorus Topology

18.7 25.331.5

38.044.1

50.656.8

18×18

63.2

4×46×6

8×810×10

12×12

14×14

16×16

18×18

0

10

20

30

40

50

Inse

rtio

n Lo

ss (

dB)


TorusNX Topology

15.819.5

23.227.1

31.034.9

38.842.7

4×48×8

16×16

0

10

20

30

40

50

Inse

rtio

n Lo

ss (

dB) Square Root Topology

12.221.5

30.6

Propagation Crossing Dropping Into a Ring

0 2 4 6 8Topology Size (nodes)

32

Simulation Results

0 100 200 3001

10

100

Num

ber

ofW

avel

engt

h C

hann

els

Number of Access Points

Torus Topology

100

Non-Blocking Torus Topology

10 20 301

10

Num

ber

ofW

avel

engt

h C

hann

els


20 dB30 dB40 dB

Improved

20 dB30 dB40 dB

Original

TorusNX Topology

0 100 200 3001

10

100

Num

ber

ofW

avel

engt

h C

hann

els


Square Root Topology

0 100 200 3001

10

100

Num

ber

ofW

avel

engt

h C

hann

els


20 dB30 dB40 dB

Improved

20 dB30 dB40 dB

Original

Original is based on the IL results from previous slide, Improved is based on a hypothetical improvement in crossing loss from 0.15 dB to 0.05 dB.

Optical power budget

Optical power budget

33



34

Noise and Crosstalk

Laser Noise

Inter-Message Crosstalk

Intra-Message Crosstalk

Modulation Noise

Crosstalk

Filter

Coherent noise

Incoherent noise

35

Effects of Noise

Network Size

Op

tical

SN

R

Number of λ Network Load

36

Simulation Results

0

10

20

30

40

50

Opt

ical

SN

R(d

B)

100 101 102 103 104 105 106 107

Message Size (bit)

TorusNon-blocking TorusTorusNXSquare Root

The line at OSNR=16.9 dB is where a bit-error-rate of 10-12 can be achieved, assuming an ideal binary receiver circuit and orthogonal signaling.

Results •Results are plotted for network size of 8×8 at saturation, at the detectors.• Maximum OSNR = ~45 dB (due to laser noise)• Minimum OSNR < 17 dB (due to message-to-message crosstalk)• Variations between networks due to varying likelihood of two message intersecting on network topology.

System Performance• SNR measures the likelihood of error-free transmission.• Lower SNR designs will require additional retransmission, resulting in lower throughput performance.

37



38

Power Usage

0V1V

n-regionp-regionElectronic Control

0V

1V

OhmicHeater

Thermal Control

λ

Tran

sm

issio

n

Injected Wavelengths

Off-resonance profile

On-resonance profile

• Laser Power• Active Power

• Modulating• Detecting• Broadband

• Static Power• Thermal tuning

• Tx\Rx Power• Drivers• TIAs

39

Energy Per Bit

10-13

10-12

10-11

10-10

10-9

10-8E

nerg

ype

rB

it(J

/bit)

10-7

100 101 102 103 104 105 106 107

Message Size (bit)

TorusNon-blocking TorusTorusNXSquare Root

40

Power Breakdown

Router Logic43%

Router Buffer44%

Electronic Wire3%

Detector3%

Modulator4%

PSE2%

Thermal1%

Router Logic45%

Router Buffer44%

Electronic Wire2%

Detector2%

Modulator4%

PSE2%

Thermal1%

• Results based on randomly generated traffic with message sizes of 100 kbit, with network in saturation.• Data was collected on 64 nodes topologies constrained to a total surface area of 2 cm × 2 cm.

Torus Topology Nonblocking Torus Topology

• 7 wavelengths @ 10 Gbps/each• Power Dissipation = 1.59 W


41

Power Breakdown

Router Logic37%

Router Buffer31%

Electronic Wire1%

Detector10%

Modulator17%

PSE1%

Thermal3%

Router Logic34%

Router Buffer31%

Electronic Wire7%

Detector8%

Modulator14%

PSE2%

Thermal4%

Square Root Topology TorusNX Topology



Performance

43

Performance

• Uniform random traffic• 256 cores, 64-node network

44

Scientific Applications0.00001

0.0001

0.001

0.01

Cactus GTC MADbench PARATEC

Exe

cutio

n T

ime

(s) E-Mesh P-Mesh

0.00001

0.0001

0.001

0.01

0.1

Cactus GTC MADbench PARATEC

Ene

rgy

(J)

E-Mesh P-Mesh

Other Interesting Issues

46

Memory Access

Processor Core

Network Router

Memory Access Point

[G. Hendry et al. Circuit-Switched Memory Access in Photonic Interconnection Networks for HPEC. In Supercomputing, Nov. 2010]

47

Other Arbitration Means - TDM

[G. Hendry et al. Silicon Nanophotonic Network-On-Chip Using TDM Arbitration. In HOTI, Aug. 2010]

48

Wavelength Granularity

• Original Re-design

λ λ

Scalable number of WDM channels

49

Conclusion

• Some applications / programming models definitely well-suited to a circuit-switched photonic network

• Interesting tradeoffs and design space• Photonic physical layout / design• System-level benefits from device improvement• Network-level improvements

hybrid on-chip data networks

Documents