The Photonic Bottleneck

Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on October 10, 2008 at 17:07 from IEEE Xplore. Restrictions apply.

a1208_1.pdf OThI1.pdf

©OSA 1-55752-830-6

The Photonic Bottleneck Kerry Hinton*, Peter M. Farrell+, Rodney S. Tucker* *ARC Special Center for Ultra-Broadband Information Networks, +National ICT Australia, University of Melbourne, Victoria, Australia [email protected]

Abstract: By analyzing the basic physics of all-optical processes, we show that all-optical networks will suffer a “photonic bottleneck” due to the fundamental properties of photons. © 2006 Optical Society of America OCIS codes: (230.4320) Nonlinear optical devices; (250.5300) Photonic integrated circuits

1. Introduction The concept of the “electronic bottleneck” was often discussed in the late 1990’s around the time of the explosive growth of the Internet. At that time, it appeared that the ability of electronic routers to process IP traffic was being out-stripped by the growing demand for Internet capacity. However, it was also apparent that the capacity of WDM optical communications systems was growing even more rapidly, resulting in many researchers and developers viewing all-optical systems as the solution to the electronic bottleneck [1,2]. There is no doubt all-optical phenomena offer much faster dynamics than current electronic devices, but there is much more to obtaining reliable high capacity optical systems than just high speed devices. To be economically viable, the next generation optical transport technologies must provide four times the capacity at no more than 2.5 times the cost of the current generation [3]. If bit rate was the only concern, then 40 Gb/s and higher bit rate systems would have been deployed some years ago. The reality is that the commercial deployment of a new technology is dependent upon its ability to reduce total system costs per customer-paid unit of traffic [4]. This includes CAPEX and OPEX resulting from equipment power consumption and size. Therefore, before all-optical routers can eliminate the “electronic bottleneck”, there must be a feasible path leading to these routers becoming commercially competitive with their evolving electronic counterparts. In this paper we show that the fundamental physics of the photon-photon interaction is a significant impediment against cost-effective deployment of all-optical networks. In fact, the basic properties of photons manifest themselves as a “photonic bottleneck” that will impede the introduction of all-optical networking. 2. The photonic bottleneck in optical communications Optical networks are evolving toward the Generalised Multi-Protocol Label Switching (GMPLS) [5]. GMPLS implements a layered switching stack, depicted on the left of Fig. 1. Also shown in Fig. 1 are the buffering and the signal processing requirements (including switching) as well as the functions provided by each layer and how they are currently implemented. Operations shown in parentheses are not yet commercially available. GMPLS Protocol Stack

Buffering Needs

Signal Processing packet header processing

Large: packet contention, LSP merge, Class of Service

label processing

TDM

Small: Synchronisation, frame processing, TDM

frame processing, FEC, TDM, regeneration

Service to customer Traffic engineering Network management

None

(Wavelength conversion, regeneration)

Transport

Waveband

None

(Wavelength conversion, regeneration)

Transport

Fibre

None

None

Transport

Optical

Wavelength

O/E Interface

Large: packet contention, header processing

LSP

Present day Electronic

Packet

Function

Fig. 1 The GMPLS protocol stack showing the requirements for buffering and signal processing.

The packet and LSP layers at the top of the GMPLS stack require large buffers, primarily to deal with contention [6]. These layers also process the data stream to implement packet and label switching. Although the TDM layer has a much smaller buffer requirement it does employ a significant amount of signal processing to implement TDM, FEC, scrambling, BER monitoring and network functions management functions [7]. Today, the wavelength and lower layers provide transport paths. (In the future, these wavelength and waveband layers may include some signal processing such as wavelength conversion and optical regeneration.) As indicated on the right of Fig. 1, the optical/electronic (O/E) interface currently lies between the WDM and TDM layers. (“WDM” includes both the wavelength and waveband layers.) For all-optical packet switching this interface moves to the very top of the stack. (In “all-optical” network test-beds, it is usually recognised that electronic signal processing is currently more effective. The “all-optical” packet switches in these test-beds electronically process the packet header after splitting it away from the packet payload which is optically buffered but not processed.)

Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on October 10, 2008 at 17:07 from IEEE Xplore. Restrictions apply.

a1208_1.pdf OThI1.pdf

©OSA 1-55752-830-6 It has been shown that all-optical buffering is competitive with electronic (CMOS) buffering only when the buffers are very small [6]. The processing and buffering requirements of the LSP and Packet layers make it clear that the O/E interface cannot be moved into these layers cost-effectively. This is a manifestation of the “photonic bottleneck” which, as will be shown below, also applies to every implementation of all-optical signal processing. Whether or not the packet header is processed electronically, “all-optical” implementations of packet switching must take into account the significant amount of bit-level signal processing which occurs in the lower TDM layer. We now turn our attention to the TDM layer. When the term “electronic bottleneck” emerged, there were proposals to remove or “thin down” the TDM layer (often called “SONET-Lite” [8]). However, it was recognised that many functions of the TDM layer are absolutely essential for good network management, so the proposals moved these TDM functions to other layers rather than eliminate them. Therefore moving the O/E interface above the TDM layer (i.e. all-optical TDM) will require all-optical signal processing to implement these essential network management functions as well as bit-by-bit switching, signal regeneration and wavelength conversion. Although the TDM layer has only small memory requirements, all these functions require processing of the payload. In this case a photonic bottleneck arises because shifting the O/E interface above the TDM layer will be very difficult without significantly increasing CAPEX and OPEX, for reasons discussed below. If we attempt to avoid the photonic bottleneck by removing the TDM layer from all the core-routers (making the core network a WDM routed network) and electronically processing only the LSP/packet headers, then fault location will become difficult. This is because only the header will be regenerated at each node. Payload degradation and payload faults, particularly those which occur inside the routers, will require an extra layer of all-optical fault monitoring thereby increasing CAPEX, OPEX and network complexity: another photonic bottleneck. 3. Fundamental physics electron and photon interactions The three key differences between photons and electrons, which influence the relative size, power and costs of electronic and photonic signal processing are: • The photons used in optical communications are a factor of 106 larger than electrons. The Compton wavelength of photons used in optical communications is ~1 μm. The Compton wavelength of an electron is ~ 2x10-6 μm. The Heisenberg Uncertainty Principle means the devices used to manipulate photons must be at least ~106 times the minimum size of those used to manipulate electrons. Currently the limit on the size of electronic circuitry is ~ 10’s nm, due to lithography [9]. However, there are technologies which will enable smaller electronic devices [10]. There is no similar opportunity to reduce the size of photonic devices. • Photons have zero rest mass and so must propagate. In contrast, electrons have non-zero rest mass and are much easier to confine. Because photons must propagate, they experience losses which must be compensated via a gain mechanism. All gain mechanisms consume pump power and are subject to quantum noise which manifests as amplified spontaneous emission. This cannot be removed as it is a fundamental property of the quantum vacuum [11]. The energy threshold for generating spontaneous photons is relatively low (~ 10-19J), hence the photonic signal must have sufficiently high power to ensure a good signal to noise ratio. • The photon-photon interaction is much weaker than the electron-electron interaction. The physics of photonphoton interactions is described by quantum electro-dynamics (QED). QED represents interactions between photons and electrons using “Feynman Diagrams” which are built up by inter-connecting the basic interaction diagram shown in Fig. 2(a)[11]. In Fig. 2(a), ein represents an incoming electron, γ represents the photon with which the electron interacts and eout is the outgoing electron arising from the interaction. The point where the lines meet is called the “vertex”. For moderate optical intensities, the relative strength is the interaction represented by a Feynman Diagram is, approximately, given by (1/137)(number of vertices) [11].

γ ein

eout

(a) Basic Feynman Diagram

esig γ ecntrl (b) Electronic switching diagram

ecntrl γsig

γsig ecntrl

(c) Photons switched by electrons

ein γsig

γcntrl

eout γsig

(d) χ(2) based photonic switching (PPLN)

γcntrl

γsig

ein γsig

γcntrl

eout

(e) χ(3) based photonic switching (HNLF, SOA)

Fig. 2 Feynman Diagrams for electronic and optical signal processing, including switching, wavelength conversion and regeneration.

The 2-vertex Feynman Diagram shown in Fig. 2(b) relates to electronic switching. The “cntrl” and “sig” subscripts correspond to the control field and signal fields respectively. Electronic switching or control of photons is shown in Fig. 2(c). The 3- and 4-vertex Feynman diagrams for all-optical signal processing, switching, regeneration

Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on October 10, 2008 at 17:07 from IEEE Xplore. Restrictions apply.

a1208_1.pdf OThI1.pdf

©OSA 1-55752-830-6 and wavelength conversion using non-linear susceptibilities χ(2) and χ(3) are shown in Fig. 2(d) and (e) respectively. In these diagrams, the electrons help mediate the interaction. From these diagrams, it is seen that the photonic interaction strength is of order 137 (for χ(2)) and 1372 (for χ(3)) weaker than the electronic interaction in Fig. 2(b). Widely used approaches for all-optical signal processing include [12]: semiconductor optical amplifiers (SOAs), highly non-linear fibres (HNLF) (e.g. chalcogenide fibres) and periodically poled Lithium Niobate (PPLN). Table 1 shows typical experimental results to date along with corresponding values for commercially available electronics. Table 1 Comparison of size and power requirements for all-optical and electronic signal processing.

Size Power

SOA[13] mm’s mW

HNLF[14] 10’s cm – 10’s m W’s – 10’s W

PPLN[15] cm’s 10’s mW – 100’s mW

Electronic 35nm[16] nW’s – μW’s[17]

Table 1 accords with the argument made above that all-optical switching technologies are orders of magnitude larger in both size and power requirements compared with current electronic technology. The interaction strength increases with the control signal intensity, I, interaction region length, L, and nonlinear coefficient, χ(2), χ(3) [15,18]. As shown above, χ(2) and χ(3) are constrained by the strength of the photonphoton interaction. To enhance the interaction we must increase I and/or L. Increasing the optical intensity, I, by increasing the power will add to system power costs, approach the optical damage threshold as well as impinge on laser safety standards, which are a non-negotiable requirement of optical communications systems [19]. (Reducing the cross-sectional area is limited by the need to confine the optical power in the waveguide [20].) Therefore increasing I will increase system costs. Increasing L necessitates attending to device losses and any phase matching requirements, and so will add to the equipment footprint and the requirement for environmental control. Increasing L using resonant structures will increase losses and reduce the switching speed. Therefore, increasing L also increases system costs. In contrast, electronic signal processing is known to be effective with much smaller footprint and power consumption [9]. Future optical networks will use a combination of photonic and electronic technologies. The best way to manage photonic and electronic bottlenecks will be to use each technology in the role for which it is best suited. 4. Conclusions We have argued that the fundamental properties of photons manifest themselves as a photonic bottleneck. The implementation of all-optical buffering, bit-by-bit switching and/or signal processing, wavelength conversion or regeneration will unavoidably require more space, material and/or power than the electronic counterpart, thereby increasing CAPEX and OPEX. We believe the photonic bottleneck will be a serious impediment to the development of truly all-optical networks. 5. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20]

W. Ha, et.al, “Demonstration of photonic fast packet switching at 700Mbit/s data rate,” Electron. Letts. Vol.27, p789, 1991 M. Dhodhi, et.al, “Bottlenecks in next generation DWDM-based optical networks,” Computer Communications, Vol. 24, p.1726, 2001 S. Reddy, L. Taira, M. Resso, “Signal Integrity Concerns when Modulating Laser Transmitters at Gigabit Rates,” DesignCon 2002, (available at http://cp.literature.agilent.com/litweb/pdf/5988-6590EN.pdf) For examples see: www.cisco.com/go/ciscoatwork, www.nortel.com/products/01/optera/metro/msp/collateral/nn-100520-0502.pdf G. Bernstein, B. Rajagopalan, D. Saha, “Optical Network Control; Architecture, Protocols & Standards,” Addison Wesley, 2004 R. Tucker, “Petabit-per-Second Routers: Optical vs Electronic Implementation,” OFC 2006, Paper OFJ3, 2006 J-P. Vasseur, et.al, “Network Recovery: Protection and Restoration of Optical, SONET-SDH, IP and MPLS,” Morgan Kaufmann, 2004 N. Ghani, et.al, “On IP-WDM Integration,” IEEE Commun. Mag. Vol.38, p.72, 2000 “International Technology Roadmap for Semiconductors”, 2005 Edition, International Roadmap Committee, http://public.itrs.net/ “Technology Roadmap for Nanoelectronics,” R. Eompano (Editor), Information Society Technologies, 2001 D. Marcuse, “Principles of Quantum Electronics,” Academic Press, 1980 P. Sales, P. Beneit, C. Politi, “Physial layer limiations in high-speed electro/optical nodes. The European approach,” Proc. 2nd International Conf. on Broadband Networks, p.328, 2005 T. Houbavlis, et.al., “All-Optical Signal Processing and Applications Within the Esprit Project DO_ALL,” JLT. Vol.23, p.781, 2005 S. Wantabe, “Optical signal processing using nonlinear fibers,” J. Opt. Fiber Commun. Rep., Vol.3, No.1, p1, 2006 C. Langrock, et.al, “All-Optical Signal Processing Using χ(2) Nonlinearities in Guided-Wave Devices,” JLT. Vol.24, p.2579, 2006 P. Bai, et.al., “A 65nm Logic Technology Featuring 35nm Gate Lengths, Enhanced Channel Strain, 8Cu Interconnect Layers, Low-k ILD and 57μm2 SRAM Cell,” IEEE IEDM 2004 Tech. Digest, p.657, 2004 G. Moore, “No Exponential is Forever: But “Forever” Can Be Delayed!” Proc. IEEE ISSCC., Paper 1.1, 2003 B. Saleh, M. Teich, “Fundamentals of Photonics,” Wiley Interscience, 1991 ITU-T Recommendation G.664, “Optical safety procedures and requirements for optical transport systems,” 03/2003 M. Foster, K. Moll, A. Gaeta, “Optimal waveguide dimensions for nonlinear interactions,” Optics Express, Vol.12, No.13, p.2880, 2004