Root Filetype
Root Filetype
Root Filetype
Datacenter Networks
Neelakandan Manihatty Bojan ([email protected]), Jingyun Zhang ([email protected])
AbstractFuture Datacenters requirements demands the deployment of optics within the Datacenter. A better understanding
of optical technologies will enable better design choices. The aim
of the paper is enlighten the reader with the developments in
optics and its technology (for short range communication ie., <5
km) and how they can be exploited to meet the demands of future
Data center networks (DCN). We start with a general overiew
of current DCN architecture, its bottleneck (at chip, board and
rack level) and the motivation to move towards optics. Then we
shall discuss on the recent developments (up-to January 2015) in
optical components and their technologies, comparing the various
design choices (through parameters) and their applicability in a
datacenter center scale networks. In the later section of the paper
we discuss about how the developments in optical components
and their technology are changing the design space of optical
subsystem used in DCN. Finally we give some insights on how the
above developments can be exploited to increase the throughput,
lower the latency and power consumption in the Datacenters
(addressing in terms of the compute, interconnect and storage
networks).
I.
INTRODUCTION
II.
Bisection bandwidth
Network diameter
Path diversity
Ideal throughput
Scalability
8.5 percent of data center traffic will be generated between data centers (data replication and software/system updates).
Fig. 1.
Fig. 2.
Adding to the above issues is the unpredictable traffic profiles of the data centers. Datacenter have a slightly more varied
traffic patterns compared to the supercomputer traffic that have
very high locality. There has been a lot of investigation about
the type of traffic in a data center [7],[8],[9]. The trace analyis
indicates that data center exhibit ON/OFF behavior and are
heavily tailed [10].
New applications of the future will push the Datacenters
to its limits. This raises the following questions:
Are the current Datacenter design and network architecture future proof?
Future data center needs to sustain the increasing bandwidth requirements, reduce the power consumption and decrease the overall latency. The above requirements require us
to do a critical evaluation of the limits of current data center
infrastructure and explore new technologies and architectures
to meet the future data center requirements.
III.
Electronics
7
7
7
7
7
3
3
3
Optics/Photonics
3
3
3
3
3
7
7
7
Need for massive efforts towards integration and assembly of discrete optical components
Cost
Absence of optical memories forces us to explore novel scheduling techniques for optical packet
switched networks [14].
O PTICAL C OMPONENTS
TABLE II.
Techniques
Benefits
B. Modulator
A modulator is a device that is used to change the properties of the carrier signal that is used for the transmission
of the data. An optical modulator is a device that is used
for manipulating the light beam from source. Based on the
Direct modulation:
Direct modulation involves changing the intensity of
the light beam by modulating the current that is
driving the light source. This technique is limited
by the bandwidth chirping effect when applying
and removing current to laser diodes having narrow
linewidth. This technique is ideal when the data rates
are in the low gigabit range (<3 GHz) and transmission distances are less than 100 km.
External modulation:
In this case modulation is performed outside the laser
cavity. A separate device called light modulator is used
to modulate the intensity or phase of the light source.
The light source is turned on continuously and the
light modulator acts like a shutter controlled by the
information being transmitted. This type of application
is suited for high bandwidth application but this comes
at the cost of expensive and complex circuitry to
handle high frequency RF modulation signal.
External modulation uses a Mach-Zehnder interferometer (MZI) waveguide structure fabricated on Lithium
Niobate. Lithium Niobate is used because it has low
optical loss and high electro-optic coefficient. Electrooptic effect is the one in which the refractive index of
the material changes in response to an applied electric
field. The The waveguide region is slightly doped with
impurities to increase the refractive index for guiding
the light. Light entering the input of the modulator is
split between the two paths. One of the paths (bottom
one) is unmodulated while the upper path has two
electrodes across it. So when a voltage is applied due
to the electro-optic effect, the upper path experiences
a higher delay (resulting in a phase change) compared
to the lower path. Hence based on the application of
the voltage across the electrodes, the light will have
high output (due to constructive interference when no
voltage is applied) or low output (due to destructive
interference when voltage is applied). Phase modulation can be performed using a very similar effect using
Polarization Maintaining Fiber (PMF).
Electro-absorption modulator (EAM) is a device that
is used to control the intensity of the laser beam by
electric voltage. EAM exploit the Franz-Keldysh effect
wherein, the applied electric field changes the ab-
Still in research,
need to make multi-core fibers more affordable and couplers cheap
Challenges
Complex modulation formats results in
receiver complexity and higher power consumption,
need linear electronic components,
should also be made cost and power effective,
tighter loss budget.
Number of channels are limited (160)
WDM lasers are not commodity yet (might not be cost efficient)
TABLE III.
Parameters
Modulation
Datarate per channel
Operation in Optical bands
Typical Distance
Threshold Current
Bias Current
Side Mode Supression Ratio
Output Power
Cost
Wavelength Stability with temperature variance
VCSEL [25]
Direct Modulation
40 Gbps (SM)/ 56 Gbps (MM)
850 nm
upto 100m
0.99 mA
12 mA
tbd
7 mW
reduced (can build VCSEL arrays)
Upto 40G for temperature upto 85 C
1.8 mW
tbd
SiP [27]
Electro Absorbtion Modulation
25 Gbps-30 Gbps
1320 nm - 1337 nm
Upto 2 Km
tbd (lowered Ith [28])
20mA - 40 mA
>40 dB
30 mW [[29]]
become lower as technology matures
better wavelength stability [29]
TABLE IV.
Operation Window (nm)
Bandwidth (GHz)
Noise Performance
Quantum Efficiency
price
Si
400-1000
10-20
low
low
low
Ge
900-1600
51
high
low
low
InGaAsP
1000-1300
22
low
high
high
InGaAs
900-1700
30
low
high
high
SiGe
1300-1500
10-25
low
moderate
NA
SiGraphene
1310-1650
402
low
moderate
NA
Fig. 4.
TABLE V.
CWDM
<8 active wavelengths per fiber
Defined by wavelengths
Short-range communications
Uses wide-range frequencies
Wavelengths spread far apart
Wavelength drift is possible
Breaks the spectrum into big chunks
Light signal is not amplified
DWDM
>8 active wavelengths per fiber
Defined by frequencies
Long-haul transmissions
Narrow frequencies
Tightly packed wavelengths
Needs precision lasers to avoid drifting
Dices the spectrum into small pieces
Signal amplification maybe used
Fig. 5.
a) physical structure of Micro Ring Resonator b) Transmission
Spectrum of Micro Ring Resonator[68]
H. Memories
that differentiates optical waveguide medium. The telecommunication networks, long and ultra long-haul networks were
heavily dependent on the SMF for most of their data and voice
transmission because of its high bandwidth capabilities. MMF
were primarily used for solutions wherein bandwidth can be
traded off for lower cost. The data center networks have a
different set of requirements that can exploit the benefits of
both SMF and MMF. Choosing a right waveguide medium
requires a careful understanding on the performance requirements of the deployed infrastructure, upfront capital investment
and long term scaling requirements. Exploring novel optical
waveguides is currently a hot research area. Polymer based
waveguides have received significant interest amongst research
groups trying to over the limits of electronic PCB. Authors
in [75] give an interesting approach to use Silicon photonics
based optical link instead of VCSEL based multimode fibers.
Industries have shown a lot of interest in Active Optical
Cables (AOC) for short range high data rate transmissions.
It is called active because all the opto-electronic components
required for transmission and reception are assembled inside
the package of the connectors. Current 10G AOCs are more
prominent but there are research efforts towards developing a
Terabit/second AOC by exploiting WDM, PAM4 modulation,
spacial multiplexing using multi core fibers.
Fig. 8.
V.
TABLE VI.
Parameters
Driving light source
Distance (for 100 GbE)
Spectral efficiency
Bandwidth support
Assembly Tolerance
Cost
Types
Application areas
pros and cons
Polymer
VCSELs
Very short distance
low
Upto 30 Gb/s
stringent
costly
Optical PCB
SMF
FP laser, SiP laser (1310 nm)
upto 1.3 km
high
>50 Gb/s
relaxed
economical
Long distance DCI
Unlimited bandwidth enables better
network design due to reduced constraints
Plasmonics for datacom, as they reduce the size of optochips. Plasmonics can confine light to much smaller dimensions that optics cannot do as it is diffraction limited. Initial results were obtained in EU-Platone project. The project
involved using Plasmonics with Si photonics. The following
were the observations:
A. Si based
Silicon photonics [77] is the ideal candidate as it is CMOS
compatible. But there is still no Si based Laser.
B. III-V on SOI
This is a very interesting fabrication technology that encompasses two complementary technologies
VI.
VII.
VIII.
DATACENTER PERSPECTIVE
Interconnection network: This refers to the interconnect infrastructure mainly provided by the
switches(TOR switch, aggregation swtiches and core
switches) and their cabling.
Fig. 10.
Fig. 11. Diagram showing the penetration of optics closer to the ASIC (will
be updated)
C OMPUTE NETWORK
A. High bandwidth
There is fundamental limitation in the bandwidth scaling
of electonic based systems due to Chip-IO bottleneck (due
to poor scaling of the pins in the processor package against
Moores law), electronic PCB limitations at high frequencies
and rack level bottlenecks ( power and form factor requirements resulting in lower port densities). Optics has tried to
explore newer venues to overcome chip level, board level and
rack level bottlenecks. Some of them are mentioned below:
Optical interposers:
Chip level bandwidth limitations can be overcome
using optical interposers. Optical interposer is the convergence of photonic and electronic system through
Silicon photonic technology. These have been primarily used to achieve high bandwidth inter chip communication but they can also be used to achieve high
bandwidth communication between interposer and onboard memory. Authors in [82] demonstrated a high
density (3.5 Tbps/sqarecm) inter-chip optical interconnect based on Silicon photonics. Further improvement
in transmission density (6.6 Tbps/squarecm) have been
achieved for the above system using better optical
components and a 1x4 optical splitter [83].
TABLE VII.
Component
Network Switch
Network Interface Card
OS Network Stack
Speed of light (in Fiber)
Delay
10-30 s
2.5-32 s
15 s
5ns/m
Round-Trip
100-300 s
10-128 s
60 s
0.6-1.2 s
B. Low latency
C. Energy efficiency
Our context of the end host latency considers the time taken
from the arrival of the data (as optical signal) to the receiver
interface (in a NIC card) to the time it has been received and
processed by the application in the host operating system.
Hence the end host latency is a combintaion of delays of
the transceiver circuitry, NIC processing, buffering, operating
systems (containing the protocol stacks) and the applications
running on it.
The advancements in optics has lead to better and faster
receivers with high bandwidth reception capabilites leading to
low latency high bandwidth optical transceivers. If the above
A. High bandwidth
Increasing bandwidth in point to point networks involves
imporoving the fiber infrastructure or using advance multi-level
modulation schemes. But interconnect networks do not have
dedicated point to point links connecting two nodes instead
they are connected through switches (for scalability and cost
reasons). Hence there are two important factors that needs to
be considered for increasing the bandwidth of the interconnect
networks: the first is scaling up the bandwidth of the cabling
infrastructure to push higher data rates, and the second is the
need for faster processing of the data at the switches.
In Context, Ya et al proposes a possible solution to overcome this problem by proposing a physical layer burst mode
protocol and by clock gating at MAC layer. By exploiting
clock gating, high power consuming MCML components can
be disabled during periods of idle activity.
X.
I NTERCONNECT NETWORK
B. Low latency
Latency and jitter are dependent on the network topology
and traffic conditions. Understanding the latency profile of the
DC network is very challenging, partly due to the variety of
applications running in the datacenter.But knowledge of network latency can be very useful to do latency optimization and
latency engineering in datacenters. This section will discuss
about the various components that constitute the latency in the
interconnect network and how the deployment of optics will
affect the latency.
Latency is introduced in the network is due to buffering,
propagation, switching, queuing and processing delays. The
sources of latency in interconnect networks are:
network interface delays: This mainly includes serialization delays like signal modulation and data
Fig. 13.
For a intra-data center environment, assuming short distance optical cables are used, the most dominant cause of
network latency are the router and switching delays, queuing
delays and the serialization delays.
C. Energy efficiency
Fig. 14.
XI.
S TORAGE NETWORK
To handle the increasing data rates in electronic interconnects, the copper traces needs to be densely packed
and also increase the number of traces. The maximum
bandwidth that can be achieved by densely packing
copper traces is limited by the effects of electromagnetic interference between adjacent copper traces,
resistive loss (due to skin effect) and dielectric loss of
the medium. This in turn increases the cost and lowers
scalability of the storage network based on electronic
interconnnects.
higher density
B. Low latency
Optics provides certain latency advantages but this can be
greatly improved in the storage networks with the use of right
storage medium. Flash based memory is one of the promising
solution that can provide the lowest latencies when integrated
with optics. IDC study [100] [101] provides detailed insights
into the rapidly growing market for enterprise storage systems
that leverage flash storage media.
All-flash data centers can provide a lot of advantages [102]
Simplified operations
C. Energy efficiency
Replacing active edge board transceivers will passive components, there by reducing the density.
XII.
OTHER APPLICATIONS
R EFERENCES
[1]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
X. Wang, Z. Cheng, K. Xu, H. K. Tsang, and J.-B. Xu, Highresponsivity graphene/silicon-heterostructure waveguide photodetectors, Nature Photonics, 2013.
L. Components, Germanium Detectors and Position Sensors, ISO
9001.
O. Optoelectronics, High Speed Silicon Photodiodes.
M. Geis, S. Spector, M. Grein, R. Schulein, J. Yoon, D. Lennon,
S. Deneault, F. Gan, F. Kaertner, and T. Lyszczarz, Cmos-compatible
all-si high-speed waveguide photodiodes with high responsivity in
near-infrared communication band, Photonics Technology Letters,
IEEE, vol. 19, no. 3, pp. 152154, Feb 2007.
M. J. Connelly, Semiconductor optical amplifiers. Springer, 2002.
R. Bonk, Linear and Nonlinear Semiconductor Optical Amplifier for
Next-Generation Optical Networks, ser. Karlsruhe Series in Photonics
and Communications / Karlsruhe Institute of Technology.
KIT
Scientific Publishing, April 2013, vol. 8.
B. Mersali, G. Gelly, A. Accard, J.-L. Lafragette, P. Doussiere,
M. Lambert, and B. Fernier, 1.55 mu m high-gain polarisationinsensitive semiconductor travelling wave amplifier with low driving
current, Electronics Letters, vol. 26, no. 2, pp. 124125, Jan 1990.
K. Magari, M. Okamoto, H. Yasaka, K. Sato, Y. Noguchi, and
O. Mikami, Polarization insensitive traveling wave type amplifier
using strained multiple quantum well structure, Photonics Technology
Letters, IEEE, vol. 2, no. 8, pp. 556558, Aug 1990.
T. Akiyama, M. Sugawara, and Y. Arakawa, Quantum-dot semiconductor optical amplifiers, Proceedings of the IEEE, vol. 95, no. 9, pp.
17571766, Sept 2007.
H. Wang, E. T. Aw, M. Xia, M. G. Thompson, R. V. Penty, and I. H.
White, Temperature independent optical amplification in uncooled
quantum dot optical amplifiers, in Optical Fiber Communication
Conference. Optical Society of America, 2008, p. OTuC2.
R. Brenot, M. Manzanedo, J.-G. Provost, O. Legouezigou, F. Pommereau, F. Poingt, L. Legouezigou, E. Derouin, O. Drisse, B. Rousseau
et al., Chirp reduction in quantum dot-like semiconductor optical
amplifiers, ECOC 2007, 2007.
S. Pathak, M. Vanslembrouck, P. Dumon, D. Van Thourhout, and
W. Bogaerts, Optimized silicon awg with flattened spectral response
using an mmi aperture, Lightwave Technology, Journal of, vol. 31,
no. 1, pp. 8793, Jan 2013.
A. Liu, L. Liao, Y. Chetrit, J. Basak, H. Nguyen, D. Rubin, and
M. Paniccia, Wavelength division multiplexing based photonic integrated circuits on silicon-on-insulator platform, Selected Topics in
Quantum Electronics, IEEE Journal of, vol. 16, no. 1, pp. 2332, Jan
2010.
S. Park, K.-J. Kim, I.-G. Kim, and G. Kim, Si micro-ring mux/demux
wdm filters, Opt. Express, vol. 19, no. 14, pp. 13 53113 539, Jul
2011. [Online]. Available: http://www.opticsexpress.org/abstract.cfm?
URI=oe-19-14-13531
S. Janz, A. Balakrishnan, S. Charbonneau, P. Cheben, M. Cloutier,
A. Delage, K. Dossou, L. Erickson, M. Gao, P. Krug, B. Lamontagne,
M. Packirisamy, M. Pearson, and D.-X. Xu, Planar waveguide echelle
gratings in silica-on-silicon, Photonics Technology Letters, IEEE,
vol. 16, no. 2, pp. 503505, Feb 2004.
Ciena, Dwdm vs cwdm, 2010, http://www.ciena.com/technology/
dwdm-vs-cwdm/.
H. Liu, C. Lam, and C. Johnson, Scaling optical interconnects
in datacenter networks opportunities and challenges for wdm, in
High Performance Interconnects (HOTI), 2010 IEEE 18th Annual
Symposium on, Aug 2010, pp. 113116.
D. Qian, M.-F. Huang, E. Ip, Y.-K. Huang, Y. Shao, J. Hu, and
T. Wang, High capacity/spectral efficiency 101.7-tb/s wdm transmission using pdm-128qam-ofdm over 165-km ssmf within c- and
l-bands, Lightwave Technology, Journal of, vol. 30, no. 10, pp. 1540
1548, May 2012.
A. Sano, T. Kobayashi, S. Yamanaka, A. Matsuura, H. Kawakami,
Y. Miyamoto, K. Ishihara, and H. Masuda, 102.3-tb/s (224 x 548gb/s) c- and extended l-band all-raman transmission over 240 km using
pdm-64qam single carrier fdm with digital pilot tone, in Optical Fiber
Communication Conference and Exposition (OFC/NFOEC), 2012 and
the National Fiber Optic Engineers Conference, March 2012, pp. 13.
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[74]
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
[88]
[89]
[90]
[91]
[92]
[93]
[94]
[95]
[96]
[97]
[98]
[99]
[100]
[101]
[102]