19.02.2016 Views

IOS XR Platform Hardware Architectures

1oxAq6l

1oxAq6l

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

<strong>IOS</strong> <strong>XR</strong> <strong>Platform</strong> <strong>Hardware</strong><br />

<strong>Architectures</strong><br />

LJ Wobker<br />

Principal Engineer<br />

Lane Wigley<br />

Technical Marketing


Agenda<br />

• Introduction – building a forwarding path<br />

• <strong>Platform</strong> Design & Building Blocks<br />

• <strong>IOS</strong> <strong>XR</strong> <strong>Platform</strong>s<br />

• <strong>Hardware</strong><br />

• Virtualized<br />

Cloud, mobility, and media are driving tremendous traffic growths over provider and web-scale networks. To address this<br />

growth, Cisco offers a diverse portfolio of products running the industry leading <strong>IOS</strong>-<strong>XR</strong>. The <strong>XR</strong> portfolio covers<br />

deployments from web-scale cloud, mobile backhaul, carrier ethernet, data center, and CDN infrastructures. This session<br />

discusses the various hardware architectures for <strong>IOS</strong>-<strong>XR</strong> routing portfolio. The session covers the NCS router portfolio,<br />

the ASR 9000 and <strong>XR</strong> Virtual Router portfolio.


Should I be here?<br />

Today’s topics and not-topics.<br />

Yes!<br />

• <strong>Hardware</strong> architectures<br />

• System-level design<br />

• Data planes<br />

• Packet forwarding<br />

• High-speed, complex,<br />

(expensive!) systems<br />

No!<br />

(or at least not much)<br />

• Routing protocols / design<br />

• Network level designs<br />

• Control plane / OS infrastructure<br />

• Selling a specific product<br />

4


You can’t always get what you want<br />

But if you try sometimes well you just might find you get what you need<br />

• Nothing is free<br />

• Some things are closer to free<br />

• The further from the middle,<br />

the more things “cost”<br />

5


The <strong>IOS</strong> <strong>XR</strong> Router Family<br />

CRS<br />

NCS 5000 NCS 5500 NCS 6000 ASR 9000<br />

6


What’s needed to build a forwarding path?<br />

1. Optical to electrical<br />

2. Transport a signal from optics to NPU<br />

3. Ingress forwarding operations<br />

4. Transport a signal from ingress NPU to egress NPU (fabric)<br />

5. Egress forwarding operations<br />

6. NPU to optics<br />

7. Electrical to optical<br />

2/6<br />

3/5<br />

4<br />

1/7<br />

7


Logical view of forwarding path components<br />

LASERS / RECEIVERS<br />

FAST MEMORY<br />

SILICON<br />

Optics<br />

Tables<br />

SERDES<br />

TCAM MEMORY<br />

TCAM<br />

FAST MEMORY<br />

NPU<br />

TM<br />

Fabric<br />

Interface<br />

ASIC<br />

Optics<br />

Packet<br />

Buffers<br />

SILICON<br />

Fabric<br />

nPower X1<br />

Optics<br />

nPower X1<br />

FIA<br />

Optics<br />

nPower X1<br />

FIA<br />

Optics<br />

nPower X1<br />

FIA<br />

Optics<br />

nPower X1<br />

FIA<br />

SERDES<br />

DRAM<br />

CPU<br />

Ethernet<br />

Switch<br />

Control<br />

Ethernet<br />

8


Liquid cooling<br />

Acoustics<br />

NEBS<br />

Heat Sinks<br />

Die size<br />

On-chip<br />

Serial<br />

DDR3<br />

LLDRAM<br />

Materials<br />

Signal Integrity<br />

Junction temp<br />

Cooling<br />

Airflow<br />

Chassis<br />

Backplane<br />

Filters<br />

Cable Management<br />

Fabric<br />

Connectors<br />

Routing Slot pitch<br />

SERDES<br />

Busbar<br />

Power<br />

PCB<br />

Interposer<br />

Optics<br />

Process<br />

Silicon<br />

CPAK<br />

Fast Convergence<br />

Run to completion<br />

N:N<br />

QSFP28<br />

Pipeline<br />

PPS<br />

MPO<br />

LR4<br />

Features<br />

Stats<br />

Programmability<br />

Router<br />

<strong>Hardware</strong><br />

Building<br />

Blocks<br />

IPoDWDM<br />

Ops/sec<br />

Capacity<br />

Bandwidth<br />

FIB size<br />

Memory<br />

Buffering<br />

ACL scale<br />

TCAM<br />

GDDR5<br />

HMC<br />

SRAM<br />

HBM


Challenge: Scale routers faster than components<br />

Note: exponential scale<br />

4096x<br />

1024x<br />

256x<br />

64x<br />

16x<br />

4x<br />

1x<br />

Moore’s Law<br />

?<br />

?<br />

?<br />

?<br />

1998 2000 2002 2004 2006 2008 2010 2012 2014 2016


Technology Trends – POS/Ethernet Standards<br />

Fastest Interface Bandwidth<br />

4096x<br />

1024x<br />

256x<br />

64x<br />

16x<br />

Interface Speeds<br />

10G<br />

40G<br />

100G<br />

Moore’s Law<br />

400G<br />

4x<br />

1x<br />

OC-48<br />

GE<br />

OC-12<br />

1998 2000 2002 2004 2006 2008 2010 2012 2014 2016


Technology Trends – POS/Ethernet Standards<br />

Fastest Interface & Router Bandwidth<br />

4096x<br />

1024x<br />

256x<br />

64x<br />

16x<br />

640G<br />

Interface Speeds<br />

40G<br />

10G<br />

Buffered Router Bandwidth<br />

8T<br />

2.4T<br />

100G<br />

Moore’s Law<br />

24T<br />

16T<br />

58T<br />

400G<br />

4x<br />

1x<br />

OC-48<br />

GE 28G<br />

8G OC-12<br />

1998 2000 2002 2004 2006 2008 2010 2012 2014 2016


Technology Trends - Optics<br />

Bandwidth/Volume<br />

4096x<br />

Driven by physics, not Moore’s Law, costs shifting<br />

SerDes dependency (NPU interface)<br />

Cooling challenge – 30C lower max temp vs. ASICs<br />

1024x<br />

256x<br />

64x<br />

16x<br />

4x<br />

1x<br />

GBIC<br />

10G XFP<br />

10G SFP+<br />

100G CFP<br />

100G CPAK<br />

40G QSFP+<br />

100G QSFP28<br />

1998 2000 2002 2004 2006 2008 2010 2012 2014 2016


Silicon & Modular Router PPS – Per NPU/LC<br />

*Mpps for Standalone and Modular/Buffered<br />

4096x<br />

*represents a combination of memory, Silicon & SerDes<br />

1024x<br />

256x<br />

64x<br />

16x<br />

4x<br />

1x<br />

CPU<br />

4<br />

12000<br />

16<br />

75<br />

CRS-1<br />

14x4<br />

ASR 9000<br />

CRS-3<br />

125<br />

45x4<br />

ASR 9000<br />

3200<br />

1200<br />

800 700x6<br />

280x5<br />

150x4<br />

NCS 5500<br />

NCS 6000<br />

ASR 9000<br />

SoC – no buffer<br />

Modular – buffered<br />

1998 2000 2002 2004 2006 2008 2010 2012 2014 2016


Technology Trends – Memory<br />

Commodity & Custom Bandwidth<br />

4096x<br />

1024x<br />

256x<br />

64x<br />

16x<br />

4x<br />

1x<br />

0.8<br />

SDR<br />

1.6<br />

DDR<br />

Scaling FIB above ~256K IPv4<br />

usec vs. msec buffering<br />

Operations / second critical for FIB<br />

DDR2<br />

5<br />

DDR3<br />

10<br />

ASR 9000<br />

2 nd & 3 rd gen<br />

NCS 6000<br />

55<br />

GDDR5<br />

HMC<br />

28<br />

NCS 6000<br />

160<br />

100<br />

25<br />

HBM<br />

DDR4<br />

325<br />

1998 2000 2002 2004 2006 2008 2010 2012 2014 2016


Technology Trends – SerDes<br />

High-performance Electrical Link (Speed in GHz)<br />

4096x<br />

1024x<br />

256x<br />

64x<br />

Optics to NPU<br />

NPU to fabric<br />

NPU to TCAM<br />

NPU to serial memory<br />

16x<br />

4x<br />

1x<br />

1.25<br />

1.25<br />

2.5<br />

5<br />

11.5<br />

15<br />

25<br />

1998 2000 2002 2004 2006 2008 2010 2012 2014 2016


Technology Trends Compared<br />

4096x<br />

1024x<br />

256x<br />

64x<br />

Interfaces<br />

PPS<br />

16x<br />

4x<br />

Optics<br />

Memory<br />

SerDes<br />

1x<br />

1998 2000 2002 2004 2006 2008 2010 2012 2014 2016


Modular Buffered Router Bandwidth<br />

Systems growing much faster than components – exponential scale<br />

4096x<br />

1024x<br />

8T<br />

24T<br />

16T<br />

58T<br />

256x<br />

2.4T<br />

64x<br />

640G<br />

Interfaces<br />

PPS<br />

16x<br />

4x<br />

1x<br />

8G<br />

28G<br />

Optics<br />

Memory<br />

SerDes<br />

1998 2000 2002 2004 2006 2008 2010 2012 2014 2016<br />

*similar trends for conductive materials, connectors, fans, power supplies, CPUs, …


<strong>Platform</strong> Design


Why not one platform?<br />

Requirements have a butterfly effect<br />

• An individual requirement may dramatically impact architecture<br />

• Every component “ahead of the curve” adds complexity<br />

• Key Drivers<br />

• Time / Cost<br />

• Buffering<br />

• FIB scale<br />

• System Scale<br />

• Features/Flexibility<br />

20


Time – Comparing 1T LCs in 2013 vs. 2016<br />

• NCS 6000 1T – 200G NPU w/ full FIB, deep buffers, programmable<br />

• Significant investment in nPower X1 silicon development<br />

• Custom memory for FIB and buffering – extremely high bandwidth and ops/sec<br />

• $271M acquisition of Lightwire for small, low-power 100G optics<br />

• ASR 9000 1.2T – similar speed, flexibility, and power – 3 years later<br />

• Commercial NPU, customized for Cisco<br />

• Commodity memories<br />

• Cisco fabric, backward compatible with earlier generations<br />

• QSFP28 optics<br />

• Full range of chassis sizes<br />

21


Buffering<br />

Ports<br />

• On-chip limited to microseconds<br />

• But… Off-chip requires<br />

• ASIC I/O pins – fewer interfaces<br />

• Board space and power<br />

• Commodity bandwidth is limited<br />

• Many devices may be needed<br />

• Better suited for buffering than FIB (1 read/write per packet optimized for 128B read/write)<br />

• Stalled – High-end graphics and networking moving to custom<br />

• Custom memories<br />

• Fewer devices are needed to reach bandwidth, fewer pins<br />

• Development and per-unit costs – Extremely expensive<br />

Ports<br />

Ports<br />

NPU<br />

Ports<br />

Ports<br />

NPU<br />

FIB<br />

Ports<br />

Ports<br />

FIB<br />

Buffers<br />

22


FIB Scale<br />

FIB<br />

• On-chip FIB limited to ~256K entries today<br />

Ports<br />

NPU<br />

Ports<br />

• Large FIB tables require external memory:<br />

Buffers<br />

• ASIC I/O pins redirected from interfaces to memory<br />

• 4-8 reads for every packet – high operations per second<br />

• Board space (density) and power<br />

23


System Scale<br />

Four common approaches<br />

FIB<br />

• Fixed – single NPU / Forwarding ASIC<br />

• All resources dedicated to network interfaces<br />

Ports<br />

NPU<br />

Fabric<br />

• May or may not have external memories<br />

• Fixed – multiple NPUs<br />

Buffers<br />

• Connected via mesh or fabric chips<br />

• Modular – Expand with line cards<br />

• Resources for fabric and usually external memories<br />

• Multi-chassis<br />

• Adds fiber connections to fabric cards (more power & board space)<br />

• Increases software complexity<br />

24


Features and Flexibility<br />

Network Interface<br />

• Packet Processing Engine (PPE)<br />

• C programmable<br />

• Run to completion<br />

• Anything is possible, no pure optimization<br />

• Wide range of pipeline programmability<br />

• NCS 5000 < NCS 5500 < ASR 9000<br />

CRS PPES<br />

• Packet rate is strongly correlated to<br />

cost, power, and flexibility<br />

Ingress<br />

Port<br />

Link<br />

Term Parser Layer<br />

Tunnel<br />

VLAN<br />

Service<br />

Trans<br />

Term<br />

Fwd<br />

PMF<br />

Egress<br />

FEC<br />

Res<br />

FEC<br />

Res<br />

NCS 5500 PIPELINE<br />

25


<strong>IOS</strong> <strong>XR</strong> <strong>Platform</strong>s


The <strong>IOS</strong> <strong>XR</strong> Router Family<br />

CRS<br />

NCS 5000 NCS 5500 NCS 6000 ASR 9000<br />

27


NCS 6000 – First Dense 100G<br />

• Highest router bandwidth capacity<br />

• 128 Tbps first generation multi-chassis (2013)<br />

• Only Tbps card on the market in 2013-2015<br />

• With buffers and full FIB<br />

• Dramatic improvement in power efficiency<br />

• $6000/100G/month power savings over CRS-3 64x100<br />

• Merchant fabric<br />

• Highly programmable forwarding<br />

• Custom memories for FIB and buffering<br />

• Custom CPAK optics<br />

28


NCS 6000 Line Card Architecture<br />

• Slice architecture<br />

• Optics, NPU & FIA per slice<br />

• Run to Completion NPU w/ PPEs<br />

• 1 Generation (so far)<br />

• 1T line cards w/ 200G NPUs (40 nm)<br />

• 400G NPUs in lab<br />

• TCAM for ACL/QoS scale<br />

Optics<br />

Optics<br />

Optics<br />

Optics<br />

Optics<br />

Optics<br />

T<br />

C<br />

A<br />

M<br />

Tables<br />

PPEs<br />

Packet<br />

Buffers<br />

nPower X1<br />

nPower X1<br />

nPower X1<br />

nPower X1<br />

nPower X1<br />

T<br />

M<br />

Fabric<br />

Interface<br />

ASIC<br />

FIA<br />

FIA<br />

FIA<br />

FIA<br />

• Deep buffers (50+ msec)<br />

DRAM<br />

CPU<br />

10X 100GE<br />

29


ASR 9000 Series<br />

• Full range of chassis and interfaces<br />

• 40G, 200G, and 1T generations<br />

• Highly programmable forwarding<br />

• Partnered for a customized NPU – cisco software<br />

• Highest scale, feature, and QoS capabilities<br />

• Cisco fabric<br />

• Commodity memories for FIB and buffering<br />

• Mostly commodity optics<br />

30


ASR 9000 System Design<br />

• Backplane or midplane<br />

• Flexible switch fabric options<br />

• RP and fabric may be integrated into RSP<br />

• Variable # of fabrics for increased capacity and redundancy<br />

• First stage of fabric on line card<br />

• Multiple airflow designs<br />

• Front-to-back – 9922, 9912, 9010<br />

• Side-to-back – 9006<br />

• Side-to-side – 9001, 9004<br />

• Modular options for lower-speed and legacy interfaces<br />

31


ASR 9000 Line Card Architecture<br />

• Slice architecture<br />

• Optics, NPU & FIA per slice<br />

• Flexible pipeline NPU<br />

• 1 st stage of fabric on line card<br />

• 3 Generations (so far)<br />

• 40G line cards w/ 15G NPUs (90 nm)<br />

• 200-360G line cards w/ 60G NPUs (55 nm)<br />

• 800G-1.2T line cards with 240G NPUs (28 nm)<br />

• TCAM for ACL/QoS scale<br />

CPAK<br />

CPAK<br />

CPAK<br />

CPAK<br />

CPAK<br />

CPAK<br />

CPAK<br />

CPAK<br />

FIB TCAM<br />

NP-5c<br />

NP-5c<br />

NP-5c<br />

NP-5c<br />

Packet Buffers<br />

FIA<br />

FIA<br />

FIA<br />

FIA<br />

DRAM<br />

Fabric<br />

CPU<br />

• Huge FIB (5M+) & buffers (200 msec)<br />

8X 100GE<br />

32


ASR 9000 Modular Line Card Architecture<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

PHY<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

PHY<br />

2/4/8X 10GE<br />

10G SFP+<br />

10G SFP+<br />

PHY<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

10G SFP+<br />

PHY<br />

20X 10G<br />

Bay<br />

0<br />

FIB TCAM<br />

NP-5c<br />

FIB TCAM<br />

Packet Buffers<br />

FIA<br />

Packet Buffers<br />

Fabric<br />

1G SFP<br />

1G SFP<br />

1G SFP<br />

1G SFP<br />

1G SFP<br />

1G SFP<br />

1G SFP<br />

1G SFP<br />

1G SFP<br />

PHY<br />

40G<br />

QSFP+<br />

40G<br />

QSFP+<br />

PHY<br />

40G<br />

QSFP+<br />

PHY<br />

100G<br />

CPAK<br />

PHY<br />

100G<br />

CPAK<br />

100G<br />

CPAK<br />

PHY<br />

Bay<br />

1<br />

NP-5c<br />

FIA<br />

DRAM<br />

CPU<br />

1G SFP<br />

20X 1G<br />

1/2X 40G<br />

1/2X100G<br />

MODULAR 3 RD GENERATION LC<br />

33


NCS 5500 – 3 rd generation 100G<br />

• Dense 100GE with deep buffers<br />

• 8 & 16-slot modular, 1 & 2 RU fixed<br />

• Up to 57.6 Tbps in ½ rack in 2016<br />

• 2.4T - 3.6T line cards<br />

• Options for route and ACL scale via TCAM<br />

• Highly integrated Silicon<br />

• Single ASIC for forwarding & fabric interface<br />

• Dramatic power reduction to 0.24 W/Gbps<br />

• QSFP28 optics<br />

• GDDR5 commodity buffers<br />

34


NCS 5500 System Design<br />

AIR INLET<br />

• New chassis design for <strong>XR</strong><br />

• Orthogonal direct connect<br />

• Horizontal line cards and vertical fabric<br />

• Direct connection between line cards and fabric cards<br />

• No midplane<br />

• Distributed air intake between cards<br />

• Fans cover fabric cards<br />

FAN<br />

REMOVED<br />

• Provides cool air equally to all optics<br />

• Optics require ~30C cooler operation than silicon<br />

• Avoid preheating air from module to module<br />

REAR VIEW<br />

35


NCS 5500 Line Card Architecture<br />

• Slice architecture<br />

• Optics & Pipelined Forwarding ASIC per slice – integrated FIA<br />

• Optional TCAM for FIB/ACL scale<br />

• 1st Generation<br />

• 3.6T line cards FA at 600G<br />

• 2.9T line cards FA at 720G + TCAM<br />

• Deep VoQ buffers (50+ msec)<br />

QSFP28<br />

QSFP28<br />

QSFP28<br />

QSFP28<br />

QSFP28<br />

QSFP28<br />

Optics x 6<br />

Optics x 6<br />

Optics x 6<br />

Optics x 6<br />

Optics x 6<br />

DRAM<br />

36X 100GE<br />

Forwarding<br />

ASIC<br />

FA<br />

FA<br />

FA<br />

FA<br />

FA<br />

CPU<br />

Buffers<br />

36


Buffers<br />

TCAM<br />

Buffers<br />

TCAM<br />

QSFP28<br />

QSFP28<br />

QSFP28<br />

QSFP28<br />

QSFP28<br />

QSFP28<br />

SFP+<br />

SFP+<br />

SFP+<br />

SFP+<br />

SFP+<br />

SFP+<br />

QSFP28<br />

QSFP28<br />

QSFP28<br />

QSFP28<br />

CPU<br />

DRAM<br />

CPU<br />

DRAM<br />

NCS 5501/5502 Architecture<br />

Forwarding<br />

ASIC<br />

• NCS 5501<br />

• 1RU – Single 800 Gbps Forwarding ASIC<br />

• 4x 100G + 40x 10G<br />

• NCS 5502<br />

• 2RU – 8 600Gbps Forwarding ASICs<br />

• Integrated switch fabric<br />

• 48x 100G<br />

• Optional TCAM for scale<br />

• Deep VoQ buffers (50+ msec)<br />

40x10G<br />

NCS 5501 800G<br />

Switch<br />

Forwarding<br />

ASIC<br />

QSFP x 6 FA<br />

Fabric<br />

QSFP x 6 FA<br />

QSFP x 6 FA<br />

QSFP x 6 FA<br />

4x100G<br />

NCS 5502 4.8T<br />

Switch<br />

QSFP x 6 FA<br />

QSFP x 6 FA<br />

QSFP x 6 FA<br />

48x100G


NCS 5000 <strong>Platform</strong>s<br />

Extending <strong>IOS</strong> <strong>XR</strong> for satellite, ToR, and beyond<br />

• Maximizing Silicon capabilities<br />

• No external memories – small FIB, usec buffers<br />

• All bandwidth to ports, no fabric<br />

• Low cost and power<br />

• Full <strong>IOS</strong> <strong>XR</strong> Routing<br />

NCS 5001<br />

NCS 5002<br />

Forwarding ASIC<br />

32x100<br />

FIB<br />

Buffers<br />

NCS 5011<br />

38


CRS-X<br />

• Scaling CRS into 100G<br />

• 4x 100G<br />

• 40x 100G<br />

• 2x 100G + 5x 40G<br />

• 3x 100G + 1x 100G IPoDWDM<br />

• Up to 51.2T via multi-chassis<br />

• Fully compatible with CRS-1 & CRS-3<br />

39


Virtualized <strong>IOS</strong> <strong>XR</strong><br />

• <strong>IOS</strong> <strong>XR</strong> on x86 hardware with Linux VM & containers<br />

• Router creation in seconds – rapid service deployment<br />

• Runs NCS 6000 software (recompiled) in emulated NPU<br />

• Hosted & small PE with L2 & L3 VPNs<br />

• Ideal route reflector<br />

• 64 bit<br />

• 10+ M routes<br />

40


Virtualized <strong>IOS</strong> <strong>XR</strong> – Performance & QoS<br />

• Targeted for 5-50 Gbps forwarding<br />

• 20+ Gbps with features per socket<br />

• 3-layer Hierarchical QoS<br />

41


Call to Action<br />

• Visit the World of Solutions for<br />

• Walk in Labs – NCS 6000 Zero Packet Loss demo, NCS 5000, NCS 5500<br />

• Technical Solution Clinics<br />

• Lunch and Learn Topics<br />

• DevNet zone related sessions


Complete Your Online Session Evaluation<br />

• Please complete your online session<br />

evaluations after each session.<br />

Complete 4 session evaluations<br />

& the Overall Conference Evaluation<br />

(available from Thursday)<br />

to receive your Cisco Live T-shirt.<br />

• All surveys can be completed via<br />

the Cisco Live Mobile App or the<br />

Communication Stations


You can’t always get what you want<br />

But if you try sometimes well you just might find you get what you need<br />

• Summary & Questions<br />

44


Thank you

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!