Constellation Shaping: Can It be Useful for Datacenter Reach Communication ? Xiang Zhou and Hong Liu Google Pla2orms Datacenter Op;cs [email protected] ECOC’17 Workshop 4

Datacenter (DC) Optics ●  Used for networking fabric above servers ●  Very large volume ●  Pluggable grey optics (PSM4, CWDM4)

High Efficiency Servers (www.google.com/about/datacenters)

●  Key performance metrics ○  ○  ○  ○  ○ 

Bandwidth and cost Power and density Serviceability Cabling efficiency Latency

10G => 40G => 100G=>200G/400G Optics

DC Optics Technology Evolution Mod: 100G PAM4 Lane: 8 Power:<15W

●  50Gb/s and 100Gb/s PAM4 for 400Gb/s switch I/O interface

Per lane speed (Gb/s)

100

Mod: 50G PAM4 Lane: 8 Power:<12W Mod: 25G NRZ Lane: 4 Power:<3.5W

50

25 10

800G OSFP

Mod: 10G NRZ Lane: 1 Power: <1W

107.8x22.58x9mm

Mod: 10G NRZ Lane: 4 Power:< 2.5W

400G OSFP 100G QSFP28

2007

Year

40G QSFP+

10G SFP+

2010

2014

2017

2020

Optical receiver sensitivity (dBm)

Can PAM-M Scale to 200Gb/s Throughput ? Thermal noise and bandwidth limited

>5dB

●  Shannon Mutual Informa;on theory used for achievable sensi;vity calcula;on ●  DSP assump;ons ○  3-tap Tx FFE (ENOB=5.5) ○  17-tap Rx FFE (ENOB=5.5) ●  Op;mal modula;on format depends on component BW o  With 40Hz BW, >5dBm sensi;vity improvement from PAM4 to PAM5

Individual component bandwidth (GHz) ●  Identical analog BW assumed for all Individual components (DAC, ADC, Driver, Modulator, PD and TIA) ●  MZM with ER=5.5dB, PD R=0.8, TIA=16pA/sqrt(Hz)

●  Without other level-dependent impairments, PAM8 performs best when component BW<38GHz !

Value of FlexPAM (with fine granularity SE) PAM8

PAM7 PAM6 PAM5 PAM4 PAM2

SE: spectral efficiency

●  PAM-M with direct detec;on (DD) scales poorly toward level-dependent op;cal impairments: mul;-path interference (MPI) and laser RIN ●  For similar noise tolerance, should select the modula;on with the smallest constella;on size ●  FlexPAM with fine SE granularity allows single chip useful for different BW condi?ons/use cases ○  PAM4 for BW>45GHz, LR ○  PAM5 for BW=[40 45]GHz, LR/SR ○  PAM6 for BW=[35 40]GHz, SR

Probabilistic Constellation Shaping (PCS) ? DC Optics

Long Haul ●  ●  ●  ●  ● 

With optical amplifier Average power constrained system Coherent detection Performance limited by OSNR To achieve the same SE, PCS with increased constellation size increases Euclidean distance thus noise tolerance

2b/s coherent PAM4 d1 2b/s coherent PCS-PAM8 d2

●  ●  ●  ●  ● 

No optical amplifier Peak power constrained system Direct detection (DD) Performance limited by thermal noise To achieve the same SE, PCS with increased constellation size reduces noise tolerance

2b/s DD-PAM4 d1 2b/s DD-PCS-PAM8 d2

Value of Geometric Constellation Shaping (GCS) Uniform constellation

Non-uniform constellation Bigger upper eye opening

Tx electrical signal

Tx optical signal

MPI=23dB, ER=7.5dB

MPI=23dB, ER=7.5dB

●  GCS useful for alleviating level dependent impairments such as MPI and RIN

FlexPAM DSP: One Example PAM-M selec;on

Tx DSP 200Gb/s binary bits

2b to 3b distribu?on matching

Bit pabern preserved FEC encoder

Bit to PAM symbol mapping

Geometric shaping and pre-EQ

PAM-M selec;on

Rx DSP Rx EQ and decision

PAM symbol to Bit mapping

FEC decoder

3b to 2b distribu;on matching

200Gb/s binary bits

●  Irregular PAM such as PAM5 can be treated as a special case of PCS-PAM8 with equal probability for level 1 to 5 while zero probability for level 6 to 8

Conclusions !  Conventional probabilistic shaping not useful for peak power limited short reach optical systems !  Geometric constellation shaping can be utilized to mitigate level dependent optical impairments such MPI and RIN !  FlexPAM with Hine granularity SE might be needed to scale per lane rate beyond 100Gb/s "  Enable single chip for different bandwidth condi;ons/use cases

Constellation Shaping: Can It be Useful for ... - Research at Google

Communication ? Xiang Zhou and Hong Liu ... Long Haul. ○ With optical amplifier ... Euclidean distance thus noise tolerance. Probabilistic Constellation ...

1MB Sizes 2 Downloads 256 Views

Recommend Documents

HOW CAN IT BE
... clear that "the message has never been received, nor proclaimed, nor given free ... beginning rays of light in the "most precious message" God sent to His ... reject the message and provisions of Righteousness by Faith in and following 1888.

FIX IT WHERE IT FAILS: PRONUNCIATION ... - Research at Google
a grapheme-to-phoneme (G2P) engine to generate pronunciations for words not in the list. ... their corrections from voice search logs. We learn previously un-.

Book how hard can it be? pdf free download
Book how hard can it be? pdf free download

Rogue Femtocell Owners: How Mallory Can ... - Research at Google
in a significant way, because it is entirely passive and not foiled by .... 3. Fig. 3. Close-up of packet sizes that are produced during a phone call. Moving to Figure ...

Data storage and IT services for research at JYU About IT ... - GitHub
Nov 21, 2016 - Video services (ACP, Moniviestin), Confluence Wiki, Plone sites, etc. Storage ... access. Possible solution to implement international federated logins: ... Some Azure cloud services (e.g. notebooks) available via the campus.

Can learning kernels help performance? - Research at Google
Canonical hyperplane: for support vectors, ... Support vectors:points along the margin and outliers. ..... DC-Programming algorithm (Argyriou et al., 2005).

What is the Kalman Filter and How can it be used for Data Fusion?
(encoders and visual) to solve this issue. So for my math project, I wanted to explore using the Kalman Filter for attitude tracking using IMU and odometry data.

Can student test scores provide useful measures of school principals ...
Sep 1, 2016 - nor does mention of trade names, commercial products, or organizations .... Figure 1. Percentage of any difference in single-year ratings across ...

Can student test scores provide useful measures of school principals ...
Sep 1, 2016 - education program. The study ...... Bachelor's ...... galore and Claire Postman conducted excellent programming and research assistance, and.

Apparatus for shaping cookie dough
Aug 1, 1988 - relative to said dough in the longitudinal direction of said elongate mass of dough to separate a portion of dough from said elongate mass of ...

Mathematics at - Research at Google
Index. 1. How Google started. 2. PageRank. 3. Gallery of Mathematics. 4. Questions ... http://www.google.es/intl/es/about/corporate/company/history.html. ○.

Simultaneous Approximations for Adversarial ... - Research at Google
When nodes arrive in an adversarial order, the best competitive ratio ... Email:[email protected]. .... model for combining stochastic and online solutions for.

Asynchronous Stochastic Optimization for ... - Research at Google
Deep Neural Networks: Towards Big Data. Erik McDermott, Georg Heigold, Pedro Moreno, Andrew Senior & Michiel Bacchiani. Google Inc. Mountain View ...

SPECTRAL DISTORTION MODEL FOR ... - Research at Google
[27] T. Sainath, O. Vinyals, A. Senior, and H. Sak, “Convolutional,. Long Short-Term Memory, Fully Connected Deep Neural Net- works,” in IEEE Int. Conf. Acoust., Speech, Signal Processing,. Apr. 2015, pp. 4580–4584. [28] E. Breitenberger, “An

Asynchronous Stochastic Optimization for ... - Research at Google
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.

UNSUPERVISED CONTEXT LEARNING FOR ... - Research at Google
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-.

Limited memory can be beneficial for the evolution ... - Semantic Scholar
Feb 1, 2012 - through ''image scoring'' (Nowak and Sigmund, 1998), where individuals monitor ...... Zimmermann, M.G., Eguıluz, V.M., San Miguel, M., 2004.