optical trends in the data center
TRANSCRIPT
Optical Trends in the Data Center
Doug ColemanManager, Technology & Standards
Distinguished AssociateCorning Cable Systems
Data Center Environment
• Higher speeds• Higher density• Higher reliability• Lower capex • Lower opex• Green
VCSELs Drives MMF Value Proposition
• 850 nm VCSEL– Highly Efficient Manufacturing
Process (>100,000 / wafer)– Ease of Packaging into
Transceiver TOSA• Lowest Transceiver Price
– 10G Serial x Serial• 2:1 ($SM/$MM)
– 40G Serial x Parallel• 5:1 ($SM/$MM)
– 100G Serial x Parallel• 30:1 ($SM/$MM)
Vertical Cavity Surface Emitting Laser
Data Center Environment
Source: Corning Cable Systems
Data Center Multimode Cable Channel Distribution
Trunk LengthProduct Manufactured 2009-2011
50% 1 Trunk, 40% 2nd Trunk, 10% 3rd Trunk
0
200
400
600
800
1000
1200
1400
1600
1800
2000
10 20 30 40 50 60 70 80 90 100
110
120
130
140
150
160
170
180
190
200
210
220
230
240
250
More
Length (m)
Freq
uenc
y
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Frequency Cumulative %
Average = 54.2 m
100m: 88%
Source: Corning Cable Systems
Standard Specified Distances
850 nm Ethernet Distance (m)
1G 10G 40G 100G
OM3 1100 300 100 100
OM4 1100 400 / 550* 150 150
850 nm Fibre Channel Distance (m)
4G 8G 16G
OM3 380 150 100
OM4 480 190 125
*Engineered Length
Migration to OM4 …..
Finisar 32G presentation slide at the T11.2 Fibre Channel meeting (06/2010)
TIA-942A recommends OM4
Fibre Channel (32G) and Ethernet (100G) utilized OM4 to define distance objectives…..
Transmission and Cable Standards recommend OM4
Data Center Trends: Electronics and Connectivity
Optical Electronics and Connectivity Focus
Low Cost Low Power High Density
Increase data rates, transceiver size reduction, server consolidation/virtualization, multi-core processors, line card density, embedded optics, cloud computing, increased interconnect density
Evolution of VSR Computing Optical Interconnects
Source: Finisar 2011
Market Trends Toward 100GE
• Embedded optics adoption in data center – Reduction of signal trace length at higher data rates requires close proximity of
optics with ASIC– Enabling 5x – 10x denser I/O than edge mounted pluggable optics
Enabling Dense I/O for Data Centers
Source: Avago
MMF Transceivers Trend
• 850 nm MM VCSEL – Pluggable Transceivers• Continue dominating the DC demand from 10GE
to 40GE and to 100GE due to power, density and cost
• 850 nm MM VCSEL – Embedded Optics • Getting more adoption in DC and HPC especially
when moving to 25G per lane
1310 nm Silicon Photonics
• Optical interconnect boasts peak transfer rates of 1.6Tb/s
• 32 to 64-fiber connector
• 1310 nm MMF to 300 m with performance of 25Gb/s
• High BW SiP links server CPUs to storage units within the rack
• Allows servers to be replaced easily and independently.
Source: Intel
The Need for Speed: 10/40/100G
Server Virtualization Drives Higher Data Rates
• Multiple applications running in parallel on one server. (eg., 20 to100 apps/server)
• 10 to 50 servers consolidated • Increases utilization efficiency to =< 90%• Multi-core processors (4,8,16..50)• PCIe2: 8 lanes @ 5G 8b/10b• PCIe3: 16 lanes @ 8G 128/130b• Increased Memory • Less connectivity and reduced electronic
ports – Drive utilization of high BW optical
connectivity to mitigate system bottlenecks and support required increased I/O speeds.
Server Virtualization Drives Higher Data Rates
Source: Dell’Oro. 07/ 2012
Data Centers Are Going through Data Rate Increases
• Core switch
• 100 GbE non‐blocking
• Metro/Campus 100G connectivity LR/ER
• WAN connectivity through 100G DWDM
• Servers to TOR switch
• 40 x 10 GbE
Copper
or Fiber
• TOR switch
• 400G Rack Capacity
• 2‐4 X 40/100 GbE uplinks
• QSFP/CFP/CFP2 (MMF/SMF)
TOMORROW
• Core switch
• 10 GbE non‐blocking
• Metro/Campus 10G connectivity LR/ER
• WAN connectivity through 40G DWDM
• Servers to TOR switch
• 40 x 1 GbE
• Copper CAT5, 100 m
• TOR switch
• 40 G Rack Capacity
• 2‐4 X10 GbE uplinks
• SFP+ MMF or SMF
TODAY
1G
Nx40/100G
10G
40/100G10G
Nx10G
10G/40G/100G Ethernet Switch Line Card Density
Source: 100GbE Electrical Backplane/Cu Cable CFI IEEE 802 Plenary, Dallas, TX, Nov 2010
96 fibers OM3/OM4528 fibers OM3/OM48 fibers SMF (CWDM)768 fibers OM3/OM4
Legacy Data Center Collapsed Architecture: EDGE, Aggregation and Core Switch Consolidation into MDA
Data Center Flat Architecture: Top of Rack EDGE Switch
• 4 QSFP+ 40G ports • 48 SFP+ ports – 10G or 1G• 48 server connections
Source: Blade Networks Technologies
Source: FCI
Top of Rack EDGE Switch/Server Interconnect: 10GBASE-T
• Significant switch power requirements• 10G copper 4 to 5 watts
per port (40 nm)• Need <1w/10m (LOM)• Major silicon chip (28 nm)
development required to reduce power (2014-2015)
– Yield and $$$ issues• 10G optical switches 1 to 4
watts per port– Typical SFP+ 0.5 watts
• 10G copper Latency – 2us/PHY
10GBASE‐T Card
Source: Fulcrum Microsystems
Top of Rack EDGE Switch/Server Interconnect: 10G SFP+ Active Optical Cables
• Direct-Attached SFP+ Transceiver– Distribution to edge interconnect– Edge to server interconnect
• Performance Attributes:– Low cost
• Optical components selected to designed distance
• Reduced manufacturing and testing costs
– Small diameter, low weight & flexible– Distance capability (=>10 m)– No cleaning connector concerns– No connector insertion loss concerns– No mismatch transceiver concerns
SFP+ AOC
Source: Cisco
Data Center Flat Architecture: Top of Rack EDGE Switch
• 16 QSFP+ 40G ports • Each QSFP+ port can handle
four 1/10G server connections • 64 x 1/10G server connections
Source: Cisco MPO to LC Harness
Source: INF‐8438i
Twinax Harness
Data Center Flat Architecture
Today’s Data Center
ToR or EoR Switches
Source: Cisco Data Center Infrastructure 2.5 Design Guide, Cisco Validated Design I, December 6, 2007.
Tomorrow’s Data Center
Source: Data Center Basics and the Role of Optical Fiber, Discerning Analytics, January 30, 2012.
Trends toward collapsed architecture that accommodates more east/west vs. north/south traffic Low latency and low oversubscription required in new systems
Servers***
Data Center prioritized list* 1st priority** (close) 2nd priority*** (distant) 3rd priority
Data Center Trend: Flat Architecture
Source: Cisco
Data Center Trend: Network Monitoring
• Network layer data must first be extracted in order to apply the analysis tools– SPAN (mirroring) ports (active)– Port Tap(passive for optical)
• What is monitoring looking for?– Security threats – Performance issues – Optimization (I/O bottlenecks)
SAN
Core Switch
Distribution Switch
Access Switch
Server Device
SAN Director Switch
Storage Device
WAN/ISP
Data Center Trend: Network Monitoring
Local Area Network (LAN)
IP Network
Ethernet monitoring is growing from core port monitoring all the way down into access layer
Fibre Channel monitoring is typically between switch and storage array
Data Center Trend: Optical Tap
• Network Security and Performance
• OPEX savings through improved operations
• Better tracking and adherence to SLAsenables move to cloudinfrastructures
Configuration Options
Configuration BIntegrated MPO/LC
Configuration CIntegrated MPO/MPO
Configuration ANon-Integrated LC/LC
Data Center Trend: Optical Tap Example
Monitor Device
Data Rate Link Insertion Loss (coupler)
Fiber Type
Link Distance
8G Fibre Channel 70% LIVE 2.2dB (std) OM4/OM3 5/‐m
8G Fibre Channel 70% LIVE 1.8dB (Corning) OM4/OM3 85/75m
Device 1
Device 2
Traditional Data Center Fabrics
• Ethernet – LAN (1/10/40/100G)– OM3/OM4 Fiber– Non-deterministic
• Fibre Channel– SAN (4/8/16G)– OM3/OM4 Fiber– Deterministic
Source: Info‐Advantage
Fibre Channel speedMAP
• “FC” used throughout all applications for Fibre Channel infrastructure and devices, including edge and ISL interconnects. Each speed maintains backward compatibility at least two previous generations (I.e., 8GFC backward compatible to 4GFC and 2GFC)
• Line Rate: All “…GFC” speeds listed above are single-lane serial stream I/O’s. All “…GFCp” speeds listed above are multi-lane I/Os‡ Dates: Future dates estimated
ProductNaming
Throughput(MBps)
Line Rate(GBAUD)
T11 SpecTechnically
Completed (Year)‡
MarketAvailability(Year)‡
FC
1GFC 200 1.0625 1996
2GFC 400 2.125 2000
4GFC 800 4.25 2003
8GFC 1600 8.5 2006
16GFC 3200 14.025 2009
1997
2001
2005
2008
2011
32GFC 6400 28.05 2013
64GFC 12800 TBD 2016
128GFC 25600 TBD 2019
2015
Market Demand
Market Demand
256GFC 51200 TBD 2022
512GFC 102400 TBD 2025
Market Demand
Market Demand
128GFCp 25600 4X28.05 2014 2015
1TFC 204800 TBD 2028 Market Demand
Fibre Channel4/8/16G Variants: OM2, OM3, OM4
FC-0 400-M5-SN-I 800-M5-SN-S 1600-M5-SN-SData Rate (MB/s) 400 800 1600
Operating Range (m) 0.5-150 0.5-50 0.5-35
Loss Budget (dB) 2.06 1.68 1.63
Multimode Cable Plant for OM2 Limiting Variants
FC-0 400-M5E-SN-I 800-M5E-SN-I 1600-M5E-SN-I
Data Rate (MB/s) 400 800 1600
Operating Range (m) 0.5-380 0.5-150 0.5-100
Loss Budget (dB) 2.88 2.04 1.86
Multimode Cable Plant for OM3 Limiting Variants
FC-0 400-M5F-SN-I 800-M5F-SN-I 1600-M5F-SN-I
Data Rate (MB/s) 400 800 1600
Operating Range (m) 0.5-400 0.5-190 0.5-125
Loss Budget (dB) 2.95 2.19 1.95
Multimode Cable Plant for OM4 Limiting Variants
Fibre ChannelFC-PI6 32G
• Fibre Channel March 2010– 32G activity started– Expected completion: 2013– Commercial products: 2014
• Approved Objectives– OM3/OM4 70 m to100 m– SMF 10 km– SFP+ form factor
SFP+
70m OM3 100m OM4
Eye diagram shows distortions caused by
jitter and ISI
Source: Avago
FC-PI6 Project - 32G Optical Objectives
• Backward compatibility to 8GFC and 16GFC• Same external connectors as present connector
– LC and SFP+• Cable length of 100 m on OM4 cables
– Duplex fiber– Serial transmission– FEC
• Power goal at the port is less power per port than comparable 40GE port in 2014/15 timeframe
• Auto-Negotiation down to 8GFC and 16GFC• Products ship in 2014
IEEE Ethernet 802.3 40/100G
• IEEE 802.3 • 40 and 100 Gbps • At least 100 m on OM3 multimode fiber• At least 150 m on OM4 multimode fiber• At least 10 km on single-mode fiber• At least 40 km on single-mode fiber (100G only)• At least 2 km on single-mode fiber (40G only)• At least 7 m on copper cable assembly
• Key project dates• 802.3ba 40/100G standard completed June 2010• 802.3bg 40G standard completed March 2011
40G Ethernet Parallel Optics: OM3/OM4
12F MPO Connector Interface
QSFP Transceiver Source: Avago
Ethernet 40G and 100G: OM3/OM4
OM3 and OM4 distances contingent upon 1.5 and 1.0 total connector loss, respectively
40G Optical Transceiver: OM3/OM4
QSFP Transceiver technology• Standard 12F MPO
Connector• =< 1.0 watts per port• Now used for 40/64G
InfiniBandSource: Zarlink
40G eSR4 QSFP+
• 40G eSR4 Parallel Optics Extended Reach QSFP+ Transceiver• OM3/OM4:
• 300 m / 400 m (Industry Connectivity)• 330 m / 550 m (CCS Connectivity)
• CCS modeling shows 12% data center lengths >100 m• Internal testing demonstrated 1250 m with random
transceiver and OM4 fiber.• Commercially available now
Potential 40G Solutions
100G Ethernet Parallel Optics: OM3/OM4
24F MPO Connector Interface
Source: USConec
100G Optical Transceiver: OM3/OM4
• CXP Transceiver technology• Standard 24-fiber MPO
Connector• =< 3 watts per port
Source: Molex24F MPO Pinless Connector
100G Polarity: Two-Row Transceiver Connectivity
Note: Only one scheme shown, but examples of several methods will be shown within the standard.
No industry traction expected for 802.3ba 100GBASE-SR10
40/100G Optical Transceiver: SMF
CFP Transceiver technology• Standard duplex LC
Connectors• =< 20 watts per port• 3 to 4 ports per card• Large footprint
• Equivalent to two 10G XENPAKs
Source: CFP MSA
Duplex LC Connector
40/100G: Twinax Copper
• Traditionally, used for short-length InfiniBand connectivity and must be factory-connectorized
• No guidance included in the Ethernet 802.3ba standard for CAT UTP/STP copper cable
Source: FCI
QSFP Direct-Attached Twinax Cable
IEEE 40GBASE-T
• 40GBASE-T Study Group Approved July 2012
• 40GBASE-T 802.3bq Task Group Approved May 2013• Chip and Copper Cable Manufacturers Driving Interest
• Support is likely to be slow• Projected standard completion: 2015
• Baseline objectives: • Four-pair, balanced twisted-pair copper cabling • Up to two connectors• Up to at least 30 m
• Cable: Cat8 Shielded• 2000 MHz
IEEE 802.3bm 40 and 100G Fiber Optic Cables Task Group
Task Group will develop guidance in accordance with approved distance objectives
– Define a 40 Gb/s PHY for operation over at least 40 km of SMF– Duplex fiber, single wavelength expected
– Define a 100 Gb/s PHY for operation up to at least 500 m of SMF– Options under consideration include: 4x25G parallel optics (8F), duplex fiber wave
division multiplexing (WDM), and duplex fiber Pulse Amplitude Modulation (PAM) (probably will not happen – no consensus)
– Define a 100 Gb/s PHY for operation up to at least 100 m of OM4− 4x25G parallel optics (8F) expected
– Define a 100 Gb/s PHY for operation up to at least 20 m of OM4– 4x25G parallel optics (8F) expected (probably will be removed – not needed based
on minimal economical value)
• Expected Standard completion date: May 2015
Market Trends Toward 100GE 100GE (4x25G) pluggable form factor is still evolving!
Future Multimode 100G: Relative Connectivity Cost/Circuit
PMD Max Distance (m)
Fiber Count
Relative Connectivity Cost /Circuit at 100m
100GBASE-SR10-OM3 100 20 1.53100GBASE-SR10-OM4 150 20 1.78100GBASE-SR4-OM3 100 8 1100GBASE-SR4-OM4 150 8 1.13100GBASE-NR4-SM no WDM
2000 8 0.81
100GBASE-NR4-SM with WDM
2000 2 0.24
100GBASE-LR4-SM CFP2 10000 2 0.24
100GBASE-LR4-SM CFP 10000 2 0.24
Future Multimode 100G: Relative Transceiver Cost/Circuit
PMD Max Distance (m)
Relative Module Cost
Range
100GBASE-SR10 150 1 1
100GBASE-SR4 100 1 0.8, 1.2, 1.6
100GBASE-NR4 no WDM 2000 5 1, 3 - 6
100GBASE-NR4 with WDM 2000 6 5-8
100GBASE-LR4-SM CFP2 10000 25 5-25
100GBASE-LR4-SM CFP 10000 32 15-32
Future Multimode 100G: Relative Link Cost/Circuit
0
5
10
15
20
25
0 500 1000
100GBASESR10-OM3100GBASESR10-OM4100GBASESR4-OM3100GBASESR4-OM4100GBASENR4 - NoWDM100GBASENR4 -with WDM100GBASELR4CFP2100GBASELR4 CFP
IEEE High-Speed Ethernet Industry Connections
• Interim Meeting September 2012– Scope focused on building
consensus related to the next speed of Ethernet for wire line applications
– Straw polls showed strong support for 400G as next generation Ethernet
– Study group approved March 2013
• Most likely first designs will be 16x25G (32 MMF)
Source: USConec
Generations of 400GbEFormative Stages of 400GbE Still Evolving
Generation 1st Gen 2nd Gen
Optical Module CDFP – Copper and MMF
4X CFP4 ‐ SMF CFP2 – MMF and SMF
Electrical Interface (Gb/s)
CDAUI–16 16 lanes of retimed 25G
CDAUI–16 16 lanes of retimed 25G
CDAUI–8 8 lanes of 50G
Availability 2016 2016 2020?
C = 100CD = 400D = 500
Source: Brocade
When Is Terabit Coming?
Key:
EthernetEthernet ElectricalInterfacesHollow Symbols = predictionsStretched Symbols = Time Tolerance
1T
100G
10G
1G
400G
40G
nx10.3125G
2010 2015 2020 2025
Standard Completed
nx25.8G
400GbE16X25G
100GbE10X10G
40GbE4X10G
Dat
a R
ate
and
Line
Rat
e (b
/s)
100GbE4X25G
nX50G
400GbE8X50G
400GbE4X100G
100GbE1X100G
TbE?10X100G
nX100G
1.6TbE?16X100G
Source: Brocade
40/100G Data Center Architecture: Top of Rack EDGE Switch
40/100G: 2 x 24-fiber OM3/OM4 Uplink
Duplex LC 10G ports
Duplex LC 10G ports + 12‐fiber MPO 40G / 100G ports12‐fiber MPO 40G ports and 100G ports
Optical Connectivity 40/100G OM3/OM4
10G LC modules independently changed out for 40G and/or 100G MPO panels
Contact Info
• Doug Coleman
• E-mail: [email protected]
• Phone: 828-901-5580
• Fax: 828-901-5488
• Address: 800 17th Street NW Hickory, NC 28601