Pluggable DWDM: Considerations For Campus and Metro DCI Applications Xiang Zhou and Hong Liu Pla0orm Datacenter Op9cs With input from Vijay, Tad and Vinayak from NetArch group
ECOC 2016 WS 3 Short range optical transmission for emerging 5G, DCI and Metro Networks
Google DC Interconnection Network Intra-DC (Clusters)
500m-1km m
Cluster
Inter-DC (Campus)
2km
DC to Backbone (Metro)
80km
Global Inter-DC (Backbone, LH)
1000s km
Distance
Google Datacenter Traffic Growth
● Datacenter traffic has increased 50x from 2008 to 2014 ● Roughly doubling every year. ○ Faster than Internet growth
Datacenter Campus Network (today) Google’s Perspec9ve ● Interconnec9ng mul9ple DCs with geographic proximity ( ~2km) ● One or mul9ple warehouse computers (Clusters) within each DC ○ Google Jupiter cluster provides 1.3 Petabits bisec9on bandwidth ● Interconnec9on technique ○ Pluggable gray op9cs (typical) ○ Single mode fiber Google Oklahoma data center campus
● Interconnec9ng datacenter/POP and backbone (~40 - 80km) ● Tradi9onal DWDM transport technique ○ Grey Op9cs for client side, colored DWDM op9cs for line side ○ Client side speed typically smaller than lineside speed
DWDM Transport
Backbone router
Challenges Facing Current Solutions q Bandwidth scaling challenge for DCI campus
o Fiber exhaus9on problem
§ Pulling new fiber is expensive § Pulling new fiber also constrained by physical condi9ons
q Cost scaling challenges for DCI Metro (edge access)
o Muxponder based DWDM op9mized for tradi9onal telecom Metro, where § reach up to 300-600km § client speed significantly slower than lineside speed o Not op9mal for DCI metro, where § reach only ~40 - 80km § client speed equal (or close) to the lineside speed
Emerging Technical Trend
Datacenter router/Switch
DWDM directly on Linecard
Backbone router
● DWDM directly on router and switch card (pluggable DWDM) ○ Elimina9ng client interconnec9on and terminal chassis (reduce cost, power) ○ Allowing simpler management and control ○ Poten9al use for both campus and Metro
Pluggable DWDM: Performance Metrics Google’s Perspec9ve q Cost q Power, Size and Density o DC grey optics form factor
q Wavelength Tunability q Link performance (OSNR, dispersion tolerance etc) q Spectral Efficiency (less important than LH)
100G Tunable QSFP28: Enabling Technologies q Serial 100Gb/s direct modulation with 16nm CMOS
o 50-Gbaud PAM4 [1] o 50-Gbaud (equivalent) DMT q Lower power and small footprint tunable lase
PAM4
o With relaxed power,wavelength stability and phase noise requirements q FBG based full C-band tunable DCM (in line systems) [1] over 80km performance demonstrated: D. Sadot et al, OE 2015
DMT (courtesy to Socionext)
Revisit Optical Dispersion Compensation q Multi-span DWDM systems
o Require optical DCM every span o Additional tunable DCM o Cost is high compared to electrical dispersion compensator q Single-span DCI Metro
o Single broadband tunable DCM (FBG based technology)
Broadband TDCM
o Cost and power shared by 40-80 channels § Likely lower than electrical dispersion compensator o Reduce transceiver power § Help meet stringent power density requirement of pluggable optics
Beyond 100G: Coherent or Direct Detection ? q Direct detection
o Require multiple tunable lasers (four for 400G ! ) § Cost and power likely high o Spectral efficiency (SE) difficult to scale § Single sideband modulation could double SE but complex Tx design q Coherent detection
o Single tunable laser for 400G (shared by both the signal and the LO !) § More stringent laser performance o Higher spectral efficiency (2-4 times higher) o Support greater link loss (better OSNR sensitivity) o High DSP power: extra power for pol. and phase recovery
Ultra-Low Power Coherent DSP q 7nm CMOS node
With Baud-rate DSP
q Baud-rate or minimum over-sampling DSP
o
Baud-rate DSP achieves the lowest power consump9on but has limited dispersion compensa9on capability
q Minimal MIMO EQ tap length q Less powerful FEC q 2/3-bit DAC for 16/64-QAM q Remove OTN framer
Residual CD (ps/nm)
Conclusions q Grey op9cs DC campus interconnec9on faces bandwidth scaling challenges q Pluggable (and tunable) DWDM is a promising solu9on for both campus and
Metro without sacrifice of router and switch port density q PAM4/DMT based serial 100G could enable 100G tunable QSFP28 q 7nm ultra-low power coherent DSP likely enables 400G pluggable DWDM
Pluggable DWDM: Considerations For Campus ... - Research at Google
One or mul*ple warehouse computers. (Clusters) within ... Google Oklahoma data center campus ... o High DSP power: extra power for pol. and phase recovery ...
Characteristics. 1. Cervical and upper thoracic fusion, typically of three or more levels. 2 ..... The clinical practice of airway management in patients with cervical.
... the apps below to open or edit this item. Infosys Campus Drive at CRIT.pdf. Infosys Campus Drive at CRIT.pdf. Open. Extract. Open with. Sign In. Main menu.
When nodes arrive in an adversarial order, the best competitive ratio ... Email:[email protected]. .... model for combining stochastic and online solutions for.
[27] T. Sainath, O. Vinyals, A. Senior, and H. Sak, âConvolutional,. Long Short-Term Memory, Fully Connected Deep Neural Net- works,â in IEEE Int. Conf. Acoust., Speech, Signal Processing,. Apr. 2015, pp. 4580â4584. [28] E. Breitenberger, âAn
for sequence training, although in a rather limited and controlled way [12]. Overall ... 2014 IEEE International Conference on Acoustic, Speech and Signal Processing (ICASSP) ..... Advances in Speech Recognition: Mobile Environments, Call.
grams. If an n-gram doesn't appear very often in the training ... for training effective biasing models using far less data than ..... We also described how to auto-.
Aug 27, 2008 - Before modeling CCF, we first model community-user co- occurrences (C-U) ...... [1] Alexa internet. http://www.alexa.com/. [2] D. M. Blei and M. I. ...
been proposed as a novel alternative to SA (Kadowaki ... lowest energy in m states as the final solution. .... for Ï = argminÏ loss(X, Ï), the energy function is de-.
Dec 4, 2017 - information can be included. The distances between clusters correspond to the audience overlap between the video sources. For example, cluster 104a is separated by a distance 108a from cluster 104c. The distance represents the extent to
26-30 September 2010, Makuhari, Chiba, Japan. INTERSPEECH ... phone calls are famously inexpensive, but this is not true in most developing countries.).
the Fourier frequency domain. Results show improved performance for noise reduction in an easily pipelined system. Index Termsâ Noise Measurement, Video ...
set of types of links having different capacities and costs that can be used to .... all Steiner vertices in M have degree at least 3 and are contained in the small-.
language recognition system. We train the ... lar approach to language recognition has been the MAP-SVM method [1] [2] ... turned into a linear classifier computing score dl(u) for utter- ance u in ... the error rate on a development set. The first .
Feb 14, 2005 - engines and information retrieval systems in general, there is a real need to test ... IR studies and Web use investigations is a task-based study, i.e., when a ... education, age groups (18 â 29, 21%; 30 â 39, 38%, 40. â 49, 25%
Feb 23, 2016 - 10) Optimized Queuing Management [IOPS] ... center, high availability in the presence of host failures also requires storing data on multiple ... disks to provide durability, they can at best be only part of the solution and should ...
clinicians and educators employ it for automated assessment .... We call this new phone sequence ..... Arlington, VA: Center for Applied Linguistics, 1969.
Proportion of costs for energy will continue to grow, since. Moore's law keeps ... Challenge: Are there alternative designs that would .... semi-structured sources.
circumstances is examined, leading to a number of practical results related to the informativeness of the bounds in ... and one from Nevo and Rosen (2012b)âloosen IV assumptions in different ways, and are relevant to different types of ...... As su