ing rapidly. While applications to date have focused on small screens for portable electronics, such as mobile phones and MP3 players, manufacturers are beginning to scale up the technology. Just six months ago, the largest OLED screen measured 21 inches; with Samsungs latest offering, this has jumped to 40 inches. The 40 inch OLED display has a pixel widescreen for- The SID show took place in Boston, mat and is driven by an amor- US, on 2427 May.

NEWS BUSINESS

OUTSOURCING
USUkraine venture targets design
By employing a team of experienced engineers in Ukraine, a small US consultancy has found a way to provide optical design services at an unprecedented price. The big difference is that we only charge around $40 (731) per hour for our design services, while many firms here in the US charge $150 or more, John Ellis, president of Massachussets-based Optics for Hire, told OLE. Weve done everything from the design of endoscopes and spectrometers, to underwater cameras and automotive lighting. Since it was founded two-and-ahalf years ago, Optics for Hire has won contracts from 30 customers, including well known names such as Laser Components and GE Engines. Following the success of its SAFET Y mechanical and laser systems. We have 18 scientists based in Lviv, Ukraine, who have many years of experience in optical, mechanical and electrical design, and we are now recruiting for more, said Ellis. As well as producing designs on a computer using packages like Zemax, we can also perform prototype manufacturing using local machines. The firm was founded after the demise of a company called Constellation 3D a designer of optical disk drives at which Ellis was vicepresident of marketing. Constellation was employing Ukrainian designers and, after it went bust, Ellis came up with the idea of creating an optical design consultancy that made the most of their skills, and Optics for Hire was born. son to blink or turn away. He added that the use of laser pointers that are more powerful than 5 mW could damage a persons vision. Robertson has previously published research on red laser pointers. I tested different powers up to 5 mW and could not create recognizable damage in the human eye, he said. We know that the retina is infinitely more sensitive to shorter wavelengths. The green lasers appear much brighter because of the shorter wavelength, and can cause more damage.

Motorola

Motorola believes its NED is ready to take on existing flat-panel display technologies.
the quality of the resulting display has, to date, been disappointing. On a back plate, only 3 mm behind each sub-pixel, we place a small structure that contains about 1000 CNTs arranged such that a properly applied voltage excites each CNT to bombard the colour phosphors with electrons, explained the spokesperson.
With a thickness of just 3.3 mm, the prototype is a 5 inch diagonal section of a 42 inch high-definition television, and has a refresh rate of 60 Hz. Motorola estimates that a 42 inch NED running typical video would consume 75 W. In comparison, the firm says a similar LCD would consume around 180 W
OPTICAL & ILLUMINATION DESIGN SOFTWARE
Optical design thats fast, easy, and accurate.
CODE V is the most powerful software available for optical design, analysis, and fabrication support. No matter what type of system youre designing, you can be confident that CODE V will help transform your requirements into results. And it's easy to share CODE V's optimization and analytic capabilities with other scientific and engineering applications, enabling a seamless, end-to-end solution to your design challenges. LightTools delivers the exceptional design and analysis features you need in illumination design software, combined with an easy-to-use interface. Libraries of sources, materials, and optical properties simplify many steps in the design process. Its fast, accurate ray tracing makes it easy to visualize how your system works. From concept to completion, LightTools allows you to iterate and perfect your design quickly and accurately.
New releases now available
Want to know more? Contact our nearest office:
U.S.A. Optical Research Associates Ph: +info@opticalres.com www.opticalres.com Germany Optics & Energy Concepts Ph: +49 (0) 30 info@oec.net www.oec.net France OPTOPRIM Ph: +33 (0) 80 info@optoprim.com www.optoprim.com U.K./Ireland Kidger Optics Associates Ph: +44 (0) info@kidger.com www.kidger.com

IN BRIEF

OPTICAL STORAGE Toshiba of Japan has unveiled a triple-layer HD DVD-ROM (readonly) disk with a capacity of 45 GB, enough to record 12 h of highdefinition content. The company has also released a double-sided, dual-layer hybrid ROM disk comprised of a dual-layer 30 GB HD DVD-ROM side and a dual-layer 8.5 GB DVD-ROM side. ACQUISITIONS Fabrinet is to acquire JDS Uniphases operations in Fuzhou, China. The manufacturing-services provider will take on all assets and operations related to the Fuzhou plant, including research and development, and the manufacture of products such as crystals and precision optics. The acquisition is expected to be completed by the end of June 2005. CONTRACT S e2v Technologies, UK, has received a $5.7 m (74.5 m) order for customized CCD sensors from US-based Dentrix Dental Systems. Delivery will commence in June 2005. Dentrix will incorporate the CCDs into its digital dental X-ray imaging system. CONTRACT S Thales Laser Diodes (TLD) has been awarded a 72 m contract by the European Space Agency. TLD will supply laser diode stacks for use in the agencys ADM-AEOLUS programme, a satellite that will study the circulation of wind energy on Earth and phenomena such as El Nio.

Start-up develops FEDs

Evening Telegraph, Dundee ( D C Thomson & Co Ltd)
QFT will exploit Ravi Silvas (left) and Mervyn Roses (middle) work. Also shown: Roy Clarke (seated) and chairman Peter Dervyer.
Quantum Filament Technologies (QFT), a one-month-old spin-off from the universities of Dundee and Surrey, UK, believes it has hit upon a technology that will allow the cost-effective production of high-quality flat screens known as field-emission displays (FEDs). FEDs are essentially flat CRTs that are a few millimetres thick. In the past, several firms have tried to commercialize them, but attempts have been abandoned owing to manufacturing difficulties. QFTs solution lies with producing an array of micrometre-sized tips on the surface of an amorphous silicon backplane. These tips then direct electrons towards each pixel on a phosphor-coated screen. QFT has come up with a realizable solution to produce FEDs, Roy Clarke, the companys managing director, told OLE. The method is
based on standard materials and processing methods. It creates a true flat CRT with all the associated benefits [of a CRT], such as brightness, speed and viewing angle, but with the size and geometry of an LCD or a plasma display. QFTs approach uses a laser to produce a unique internal structure within the amorphous silicon. The internal processing creates structures that produce the highly desirable field emission properties, explained Clarke. lot of the field A enhancement takes place within the internal structure of the silicon and not in the geometry of the tips. On the back of this idea, QFT has just secured first-round funding in excess of 71 m. This will enable a prototype development programme to be run in collaboration with the universities of Dundee and Surrey, as well as a university in Korea.
QFTs current prototypes can only be demonstrated in a vacuum. However, this new programme will allow the company to produce stand-alone devices that show off all the features of its field-emission technology. Clarke says the technology is scalable and could be used in any display. In the near term we know that niche markets, such as the medical and avionics industries, are seeking high-performance displays to replace conventional instrumentation in life-critical environments, he revealed. But the technology is equally suited to mass-market sectors, such as wall-hanging TVs. QFTs ultimate goal is to see its idea used across all display sectors and to license the technology to major manufacturers. According to Clarke, QFT is already talking to several Asian display companies.

CRTs are about to be replaced by their flatter cousins on a grand scale.

Oliver Graydon

NEWS ANALYSIS
Can plasma TVs fight off LCDs?
Plasma display-panels have dominated the commercial market for large-screen TVs for five years. As alternatives such as LCDs reach larger sizes, Robert Thomas asks if the days of plasma displays are over.
Despite issues with picture-quality, high prices and, in some cases, high running costs, the market for plasma display-panel (PDP) TVs has grown steadily over the last five years. In some ways, PDP TV is to consumer electronics what the sport utility vehicle is to the automotive industry: popular, big, heavy, power-hungry and, some would say, ostentatious. ments Department of Environment, Food and Rural Affairs, notes that PDP TVs consume significantly more energy than CRT TVs: around 450 W in comparison to 90 W. However, proponents of plasma displays claim that such data do not account for the fact that PDPs simply have a larger screensize than typical CRTs. That said, there is no doubt that PDP TVs certainly do consume more power per unit screen-size than competing technologies. A recent test by Consumer Reports concluded that a 30 inch LCD TV consumed just 134 W, whereas a 42 inch PDP TV consumed 347 W. Although these figures are not as bad as critics of PDP technology suggest, it remains that large PDP TVs are relatively power hungry.

Market dynamics

Many analysts believe PDPs have had their day and that the technology was only a stop-gap solution; the question is when, not if, other display technologies will overtake PDP The most obvious contender is. the liquid-crystal display (LCD). LCD TVs already dominate the market for small- and mediumsized flat-panel TVs. In recent years, manufacturers have overcome production difficulties and can now make screens in excess of 32 inch. Large (37 inch) LCD TVs are already competing with 42 inch PDP TVs. Furthermore, a joint venture between Sony and Samsung, S-LCD, is planning to open a plant to produce 40 inch LCD screens. Price is unlikely to be an issue in the battle of the flat-panel technologies, given that the prices of both PDP and LCD TVs are falling. In 2003, SRI Consulting Business Intelligence predicted that sales of 30 inch and larger LCDs would surpass sales of PDPs by 2008. However, the transition seems to be happening much quicker, especially in Japan. A recent survey by GfK Marketing Services Japan claimed that sales of large-screen LCD TVs overtook sales of PDP TVs in February of this year. But key PDP manufacturer Matsushita Electric Industrial Co was quick to respond, claiming that PDPs still account for 90% of the global market for large-screen TVs.

Implications

If the bad press relating to plasmascreen TVs continues to increase throughout 2005, manufacturers may have difficulty convincing consumers that they have improved the performance of their products enough. Once the reputation of a technology becomes tainted, recovery is often very difficult. Plasma-screen TVs could even become unfashionable. This development would play into the hands of many competing technologies, especially now that LCD TVs are available in sizes greater than 40 inch. All things considered, PDP TVs will probably lose market share very quickly this year.
Survival of the fittest: can plasma displays (top) fight off the challenge of technologies such as LCDs (bottom) in the large-TV flat-panel display market?

Bad news

The past few months have seen an unprecedented amount of bad press concerning plasma displays. While some of these reports are true, others are based on problems relating to early PDP displays and do not account for improvements that manufacturers have made. Key areas of criticism include poor image-quality, low productlifetime and high power-consumption. These problems are not new, but are a surprise to many consumers. Makers of PDP TVs have addressed many concerns with their new models, but the probable truth is that the early PDP TVs con-
sumers bought 45 years ago are now developing major problems. Early PDP consumers are highly likely to buy replacements that use other flat-panel technologies like LCD TVs because of their bad experiences with plasma-screen TVs. Several companies that produce PDP TVs, such as Panasonic and Samsung, have acted quickly to Robert Thomas is principal at SRI rebuke this criticism. Consulting Business Intelligence, a business and technology research consultancy spin-off Energy issues from the former Perhaps the biggest area of critiStanford Research cism that PDP TVs are currently Institute. See www. facing is power consumption. For sric-bi.com or e-mail example, one recent report, which rthomas@sric-bi.com. used data from the UK govern-

T E C H N O LO GY

APPLICATIONS R&D PATENT S 2 2
ULTRASHORT PULSES MATERIALS PROCESSING
Optics aids ultrafast pulse amplification
A new type of optic that overcomes the troublesome issue of gain narrowing when amplifying ultrashort pulses with a Ti:sapphire laser has been developed by Alpine Research Optics (ARO), US. Thanks to a transmission spectrum that is opposite in shape to the Gaussian gain of the laser, the optic ensures that the shape of a pulse is unaffected by the amplification process. According to ARO, the availability of the new fused silica optic means that researchers with ultrafast systems can now take a 20 fs pulse and amplify it all the way from nanojoules up to millijoules, while still ending up with pulses that are 20 fs wide. The gain-flattering optic is designed for operation at a wavelength of around 800 nm, and tilting it tunes its centre wavelength of transmission. This optic allows for amplification without gain narrowing, and has enormous advantages over the pellicle-type etalons currently in use, said ARO. Unlike pellicles, it is virtually immune to thermal shifts and so can operate at any amplifier pulse rate. It is also robust, with a long-lived coating that can be cleaned like any other high-performance optic. DEFORMABLE OPTICS

LICENSING UDC licenses phosphorescent OLED patent to Samsung SDI
Samsung SDI of South Korea and Universal Display Corporation (UDC) have signed an OLED patent licensing agreement. Samsung intends to introduce UDCs phosphorescent OLED (PHOLED) materials and associated technology into its initial
Ultraviolet LEDs help to create a low-voltage disinfection lamp
Electronics giant Philips is attempting to patent a mercury-free, low-voltage disinfecting lamp that
To search for recently published applications, visit http://www.wipo.int/pct/en/ and http://ep.espacenet.com.
Metalbearbejdningens verden The world of metalworking Le monde de lusinage des mtaux El mundo de la elaboracin de metales
INFO: VDW Generalkommissariat EMO Hannover 2005 Verein Deutscher Werkzeugmaschinenfabriken Corneliusstrasse 4, 60325 Frankfurt am Main, GERMANY Telefon +49 69-81-0, Fax +49 69-emo@vdw.de www.emo-hannover.de
400+ Watts of Nanosecond Pulse Power
.Cost effective DPSS industrial solutions
Ablation of coatings and thin films Micro drilling and cutting Machining high strength alloys 35-1kW in a range of lasers M2 low to 30 3-50 kHz 1064 and 532nm Attenuator and fibre delivery option
Powerlase Limited Imperial House, Link 10, Napier Way, Crawley,West Sussex, RH10 9RA, United Kingdom
Tel: +44(0)Fax: +44(0)1293 456233

www.powerlase.com

INNOVATION
Technium energizes Welsh entrepreneurs
Photonics in North Wales is being given a new lease of life, thanks to the creation of an innovative technology centre with bold plans. Oliver Graydon decided to pay a visit.
The optoelectronics industry in Wales received a welcome boost earlier this year with the official opening of the OpTIC Technium, a 15 m (721.8 m) technology and incubation centre in St Asaph, near the north coast. The aim of OpTIC is to act as an engine for entrepreneurship in the region that will keep optics alive and kicking in Wales. The idea came out of the Welsh Optoelectronics Forum, which wanted Wales to be responsible for its own sustainable growth, explained Dave Rimmer, managing director of OpTIC. We needed to find a way to keep expertise here and allow it to flourish, rather than losing it. First impressions suggest that OpTIC is doing just that. Since it opened for business, a dozen high-tech start-ups, developing everything from optical-fibre sensors to three-dimensional (3D) image scanners (see box below), have moved in. An undeniable buzz of innovation surrounds you as you walk through the front door. OpTIC is effectively divided into three parts: an incubation centre with 24 incubation units each offering 50 m2 of space; a business support centre; and a technology centre with

Lasers Components

Lasers & Laser Systems
CO2 - Nd:YAG - Excimer - Femtosecond

Fiber Lasers

CW & Pulse up to kW
Laser Diodes & Modules
Blue - Green - Red - Infrared
Scanners & Modulation www.bfioptilas.com
Galvanometers - Q-Switches - AO devices

Laser Safety Eyewear

Optics & Opto-Mechanics
BFi OPTiLAS European Offices : Belgium: info.be@bfioptilas.com Denmark: info.dk@bfioptilas.com - France: info.fr@bfioptilas.com Germany: info.de@bfioptilas.com - Italy: info.it@bfioptilas.com Spain: info.es@bfioptilas.com - The Netherlands: info.nl@bfioptilas.com Sweden: info.se@bfioptilas.com - UK: info.uk@bfioptilas.com

LASER in Action

Our strengths:

Laser light sources

In house development Own production Over 25 years experience Total system solutions Flexibility Long term partnership
We solve your problems with the use of laser technology, optical sensors, optics, opto-mechanics and laser measurement.

I.L.E.E.AG

Postfach, CH - 8902 Urdorf Tel: +11 / Fax: office@ilee.ch www.ilee.ch

Laser light barriers

Laser systems
eagleyard Photonics Perfectly Suited for Various Racetracks Broad Area Lasers (BAL)
New market introduction: A 7000 mW laser at 808 nm with a single 200 m stripe. BAL outstanding results due to top parameters: - wavelengths between 735 and 1120 nm - output powers between 1000 and 5000 mW - stripe width from 60 m to 400 m available Applications in industry and science: solid-state laser pumping | material processing | medical applications (e.g. photodynamic therapy) | scientific research

www.eagleyard.com

We focus on power.

TERAHERTZ TEC HNOLOGY

Studying drug coatings is just one use the pharmaceutical industry has found for terahertz radiation. John Paul Cerroti, vice-president of product development at Teraview, tells Jacqueline Hewett about its current research and potential markets for the technology.
UK-based firm Teraview professes to be the worlds first company solely devoted to the commercial exploitation of terahertz (THz) radiation. Founded in April 2001 as a spinoff from Toshiba Research Europe, the firm has raised 10 m (714.6 m) in venture-capital funding since its inception.
JH: What application areas is Teraview targeting?
JPC: We have three application groups: analytical, medical and security. The analytical group is primarily focused on the pharmaceutical industry and is producing spectrometers and imagers. The medical group is concentrating on identifying cancers, particularly skin cancers. This is at the preclinical stage and we have not yet started clinical trials. We are working with excised tissue just now and we can differentiate between healthy and cancerous areas. The security group is studying the use of THz radiation in applications such as scanners at airports. In THz spectroscopy, the contrast mechanism is the refractive index. This means we can see things such as ceramic knives, which cannot be picked up by metal detectors or X-ray machines. We can also get a spectroscopic fingerprint of plastic explosives, so we can tell they are there and, generally, identify what kind they are. These three distinct groups are at different commercialization stages. The most advanced in terms of products is the analytical group. We have a spectrometer and an imager on the market, and have just launched a versatile version of the spectrometer. We also have a number of systems in well known pharmaceutical companies. The next-most-advanced is the security group, although we are still some way from selling commercial products. We are working in partnership with Smiths Detection and are at the proof-of-principle stage at the moment. Once this stage is completed, we will develop prototypes that can be tested in airports, for example. After that, there will be another iteration to take the product to final sale.

Weinberger

Crash-test image taken with a Weinberger visario g2.
the exact level determined by the inherent transitor noise (kTC noise) of the pixels photodiodes (light-sensitive elements of a pixel). The achievable dynamic range is around 60 dB or more, depending on the depth of the A-D conversion.

Purchasing advice

As is common when purchasing scientific equipment, it is the application that determines which parameters to consider when buying a high-speed camera. The primary factors are sensitivity, resolution, frame rate,
research, and aerodynamics in wind tunnels. CMOS cameras are so versatile that only very high-speed applications are outside their reach. Applications such as spark analysis, ballistics of space debris and magnetic field effects may require fps, which, today, can only be achieved using special mechanical solutions that deliver exposure times in the range of 35 ns. However, continual improvements in semiconductor and computer technology mean that the performance of CMOS cameras is likely to get faster in the future. This will open the door to the high-quality, cost-effective visualization of even more events.
Gerhard Holst is head of the science and research department at PCO AG, a provider of highperformance CMOS and CCD cameras based in Kelheim, Germany. See www.pco.de.
Acknowledgements Joost Seijnaeve, FillFactory (Cypress) and Christian Backert, Weinberger GmbH.
PCO & Scientific Documentations
Super-slow-motion recording of single water drops.
repeatedly turned on and off. This can cause flicker in imaging at high frame rates. It is advisable that, in addition to spending time selecting the right camera system for an application, the user also thinks about the most appropriate illumination system (DCdriven LEDs might be one solution).

Emerging applications

What are the most popular applications for sophisticated CMOS camera systems? Historically, aside from scientific research, the automotive industry has used high-speed cameras to analyse the deformation of cars as they crash. As a result, it has been a strong driver in improving high-speed imaging. In crash tests, G-stable camera systems are mounted inside the car and external cameras also record the crash sequence. In order to get exact measurements, the precision of the triggering of the cameras is crucial. Another increasingly important application area is the life sciences, where highspeed cameras can help to analyse fast biological events, such as muscle contractions. In this field, sensitivity is an important factor, as many applications require imaging via a microscope. In TV and broadcasting, high-speed systems are often used for super-slow-motion sequences in sports, movies or adverts. Many readers will be familiar with advertising video sequences showing beverages being poured into glasses in slow motion. In industry, high-speed systems can help to perform materials testing imaging the moment when a piece starts to break, or investigating malfunctions in production machines that cannot be seen with the naked eye. Fluid and gas flow can also be visualized by high-frame-rate imaging techniques such as particle image velocimetry (PIV). PIV systems are used to study many phenomena, including the injection processes in engines, water-wave development in geochemical

Stabilized He-Ne-Laser SL 03-Series
Nominal wavelength 632,9910 0,0002 nm Single-mode output power 0,8 mW Frequency stability 110-9 after warm-up Warm-up period 12 min Frequency or amplitude stabilization Internal thread for installing optics and fiberoptic couplers
Power meter Spiricon Power Products
A universal, portable, dual-channel power and energy meter is now available from Spiricon Power Products. Called the MPE-2500, the meter interfaces to thermopile, pyroelectric, semiconductor and temperature probes. The MPE-2500 features digital and analogue read-outs, and a graph-plotting facility on a highresolution backlit LCD screen. Interfaces include an SD-card slot and a USB port to connect to LabVIEW on Windows, Mac and Linux operating systems. Other features include a dual-battery option for extended operation, simultaneous measurement of average power, and individual
Am Vogelherd 46 D-98693 Ilmenau / GERMANY Fax: -6447-8 Tel: +49-(0)3677-6447-0 Internet: www.sios.de Email: info@sios.de

SIOS Metechnik GmbH

linewidth limited to 50 MHz. The combination of an SLM laser system with Xitons OPG-series nonlinear module is said to generate tunable, narrow-bandwidth light in the 400 nm to 5 m range. Applications include spectroscopy, holography, LIDAR and nonlinear frequency conversion. www.gwu-group.de Stand B1 161
Customized detectors Electron Tubes
Visitors looking for customized lightdetector modules should visit Electron Tubes booth in Hall B2. The company specializes in photon counting, pulsed light and analogue detection and can adapt its detectors to suit particular applications. According to Electron Tubes, its modules can incorporate any combination of photomultiplier, power supply, voltage divider, electromagnetic and electrostatic screening, countertime/RS232 connection and signal-processing electronics. www.electron-tubes.co.uk Stand B2 531

Photoreceivers FEMTO

The FWPR-20 series of photoreceivers from FEMTO can detect lowlevel light-signals down to the femtowatt range. Two models are available, with either a built-in silicon or an indium gallium arsenide photodiode, to cover the 3201700 nm wavelength range. The photoreceivers incorporate a transimpedance amplifier with very low noise and extremely high gain of up to 1012 VA. The resulting noise equivalent power is 0.7 fW / Hz, which permits the direct detection of optical powers down to 50 fW. When combined with an optional lock-in amplifier, FEMTO says that sub-femtowatt sensitivity can be achieved. The FWPR-20 series is said to be a compact, low-cost replacement for APDs and PMTs as no high-voltage power supply or cooling system is required. www.femto.de Stand B2 540

109 Smith Place, Cambridge, MA 02138 USA Tel: (617) 441-0600 Fax: (617) 497-8800 www.cambridgetechnology.com

UV diode laser Omicorn

Omicron has developed an ultraviolet (UV) diode laser that can write structures below 200 nm in UV photoresist. The thermoelectrically cooled module emits at 375 nm and features 350 MHz analogue modulation. According to the firm, its high-speed lasers are an alternative to gas lasers for applications such as DVD mastering, photolithography and reprography. The lifetime, price, form factor and power consumption are all said to compare favourably with UV gas lasers. www.lasersystem.de Stand B2 179
Fizeau interferometer ESDI
Optical-metrology manufacturer ESDI has released a phaseshifting Fizeau interferometer that it says is vibrationinsensitive and can capture events of 10 s duration. The Intellium H1000 suits a shop-floor environment and can accommodate machined parts and wafers measuring up to 24 inches. The device uses its patented I-Phase module to produce three phase-shifted interferograms that are acquired simultaneously and processed to give a threedimensional surface map. www.engsynthesis.com Stand B1 413
Compact cooler Helix Polycold Systems
Helix Polycold Systems will introduce the PCC compact cooler the next generation of its Cryotiger technology at LASER 2005. According to the firm, the PCC is a robust cooling system for highperformance CCDs and other detectors used in life-sciences research, analytical chemistry, materials analysis, semiconductor inspection and astronomy. Helix says the PCC cold end can be used in any orientation and is remote enough from the compressor to provide temperatures of as low as 150 C at the point of use. The PCC uses patented, non-flammable gas blends and cools down in around 20 min using a standard electrical outlet. www.helixtechnology.com Stand B2 445
Femtosecond fibre laser Toptica Photonics
The FemtoFibre femtosecond fibre laser from Toptica Photonics will make its debut at LASER 2005. The system features a mode-locked ring oscillator as the master laser and up to two amplifier arms producing sub-100 fs pulses with an actively stabilized repetition rate of 100 MHz. The wavelength can be automatically tuned between 500 nm and 2 m. The average output power is 200 mW per amplifier arm at a central wavelength of 1550 nm; 80 mW at 780 nm; 30 mW in the near-infrared; and more than 1 mW in the visible spectral range. An additional module can generate a supercontinuum extending from 950 to 2000 nm. www.toptica.com Stand B1 278
Nanopositioning system Physik Instruments
PI will display its most advanced six-axis piezo nanopositioning and scanning system: the P-587.6CD. Offering six-axis digital control over six degrees of freedom, the system is said to suit applications such as nanomanufacturing, nanometrology, scanning, microscopy, optics and mask alignment. Specifications include long travel distances of up to 800 m, nanometre resolution, active trajectory error compensation and millisecond response times. The P-587.6CD features a parallel metrology design in which all sensors measure the position of the same moving platform against the same stationary reference. This is said to improve multiaxis precision. The system also comes with LabVIEW drivers and DLLs for easy set up and integration. www.pi.ws Stand B1 450

Standard of Excellence

Paul Hoess KG P.O. Box 950240, 81518 Munich, Germany Phone: +49 (0)Fax: +49 (0)E-mail: phoess@attglobal.net U.S.A.: Stanford Computer Optics, Inc. 780 Cragmont Avenue, Berkeley, CA 94708, USA Phone: +1 (510) 527-3516 Fax: +1 (510) 558-9582 E-mail: info@stanfordcomputeroptics.com http://www.stanfordcomputeroptics.com
Buy from the technology leader
Laser-safety eyewear LaserVision
Interference coating mso jena
German firm mso jena is offering a range of customized interference coatings. Suitable for satellite-based multispectral cameras, the spectral filters performance can be adjusted to the sensitivity of Si-photodiodes. This is achieved using an all-dielectric coating with no absorbing glasses. The company manufactures its optical components by plasma-ion-assisted deposition under cleanroom conditions. According to the firm, the facility supports prototype manufacturing, as well as serial production. The coating process is supported by thin-film design software and the companys team of experienced engineers. www.mso-jena.de Stand B2 446
Picosecond laser Lumera Laser
Lumera Laser of Germany has launched a picosecond laser that produces 10 ps pulses with a pulse energy of up to 40 J and a peak power of up to 4 MW. Dubbed RAPID, the compact, diode pumped Nd:YVO4 laser delivers 2 W average power at 1064 nm. With a beam quality of M2 less than 1.2, the laser is able to focus on very small areas and reach peak-power densities in the TW/cm2 range. Hermetically sealed, the laser head features closed-loop internal water cooling. The compact unit supports single-pulse operation and burst-mode TTL triggering. Applications include micromachining along with the generation of terahertz waves, satellite ranging and picosecond spectroscopy. www.lumera-laser.com Stand B3 527
Beam profiler Ophir Optronics
Ophir Optronics is releasing its BeamStar FX v1.10 beam-profiling system at LASER 2005. Based on a userfriendly interface, the package includes new features such as automated report generation (text, graphic or mixed output), elliptical fit, point-to-point distance measurement and automation using Active X controls. A feature dubbed M2 Mentor is available to guide the user through the process of M2 measurement. www.ophiropt.com Stand B2 620/B1 421

Power meter Gentec-EO

Canadian firm GentecEO supplies a range of power meters for use with high-energy pulsed YAG lasers. The UP19K-VM/VH series combines a broadband high-energy volume absorber with the firms Ultra-disk technology, to detect emission over the deep-ultraviolet to near-infrared range at peak powers of up to 100 GW/cm2. OEM versions are available and units can be configured with several output-signal and connector options. www.gentec-eo.com Stand B1 459/B2 441

Osram supervisory board elects Goetzeler as CEO
Osrams supervisory board has elected Martin Goetzeler to succeed Wolf-Dieter Bopst as the lighting firms CEO. Bopst, who has led the firm since 1991 and been with the company for over Goetzeler: leading role. 37 years, doubled Osrams market share to 20% and almost tripled sales to 74.2 bn. Goetzeler joined in 1999 and was most recently chief financial
gren secures strategic position at Comlase
Swedish semiconductor laser and optoelectronic device specialist Comlase has hired industry veteran Nils gren as chief

CALENDAR

For a more comprehensive list of events, including links to websites, visit optics.org/events. DATE EVENT LOCATION ORGANIZER CONTACT

June 2023

12th International Metrology Congress

Lyon, France

Collge Franais de Mtrologie University of Crete

www.cfmetrologie.com

June 20 July 1
4th Summer School in Visual Optics

Crete, Greece

www.ivo.gr/summerschool/en/ index.php leos.cres.it/wfopc/ www.osa.org/meetings/ topicals/ISOM_ODS

June 2224 July 1014

4th IEEE/LEOS Workshop on Fibres International Symposium on Optical Memory and Optical Data Storage
Palermo, Italy Honolulu, Hawaii, US

CRES, IEEE, LEOS OSA

ADVER TISERS INDEX
AF Opticals Co www.afoptical.com 67 Alpine Research Optics www.AROCORP.com 28 EMO Hannover www.emo-hannover.de 23 Avantes BV www.avantes.com 36 Bandwidth Semiconductor www.bandwidthsemi.com 34 Berliner Glas KGaA www.berlinerglas.com 62 BFI Optilas International www.bfioptilas.avnet.com 31 Breault Research Organization www.breault.com IFC B&W Tek Inc www.bwtek.com 66 Cambridge Technology Inc www.cambridgetechnology.com 63 Canadian Photonic Laboratory www.cplab.com 45 Cerac Incorporated www.cerac.com 50 Cobolt AB www.cobolt.se Cristal Laser SA www.cristal-laser.fr Crystal Systems Inc 71 www.crystalsystems.com Crystran Ltd www.crystran.co.uk 57 CVI Laser www.cvilaser.com 4 Diode Laser Concept Inc www.diodelaserconcepts.com 42 Duma Optronics Ltd www.duma.co.il 73 Eagleyard Photonics GmbH www.eagleyard.com 32 Edmund Industrial Optics www.edmundoptics.de 72 EKSPLA Ltd www.eksma.com 50 ELCAN Optical Technologies www.elcan.com 19 Electron Tubes Ltd www.electron-tubes.co.uk 77 Engineering Synthesis Design Inc 28 www.engsynthesis.com Excel Technology Europe GmbH www.continuumlasers.com 24 Femto Messtechnik GmbH www.femto.de 68 Fujian Castech Crystals Inc 61 www.castech.com GPD Optoelectronics Corp www.gpd-ir.com 42 GWU Lasertechnik Vertriebs GmbH www.gwu-group.de 12 Hamamatsu Photonics www.sales.hamamatsu.com 13, 57 HC Photonics Corporation 30 ILEE AG www.ilee.ch 32 ILX Lightwave Corporation www.ilxlightwave.com 70 Image Science Ltd www.image-science.co.uk 20 IMT Masken und Teilungen AG www.imtag.ch 14 Intelligent Photonics Control Corp www.photonicscontrol.com 76 ITF Optical Technologies www.itfoptical.com 76 Jenoptik Laserdioder GmbH www.jold.com 65 Daitron Inc www.i-chipstech.com 10 Kaleido Technology www.kaleido-technology.com 73 Kapteyn-Murnane Laboratories www.kmlabs.com 76 Kentek Corp www.kentek-laser.com 34 Labsphere Inc www.labsphere.com 48 Lambda Research Corporation www.lambdares.com 27 Laser Components (UK) Ltd www.lasercomponents.co.uk 9 Laser Lines www.laserlines.co.uk 64 Laser Quantum Ltd www.laserquantum.com74 Lasermet Ltd www.lasermet.com 42 Leister Process Technology www.leister.com 62 LIMO Laser Systems www.limo.de 44 Lumina Power Inc www.luminapower.com 57 Master Bond Inc www.masterbond.com 50 Melles Griot, US www.mellesgriot.com OBC Metrolux Optische Messtechnik GmbH www.metrolux.de 32 Micro-Controle SA 42 Microtec GmbH www.microtec.de 77 Mller-Wedel Optical GmbH www.moeller-wedel-optical.com 46 MSO Jena GmbH www.mso-jena.de 68 New Focus Inc www.newfocus.com 38 Newport Corp Photonics Division www.newport.com 11 nLight Photonics www.nLight.net 7 Northrop Grumman Synoptics www.northropgrumman.com 39 NUFERN USA www.nufern.com 58 Ocean Optics BV www.oceanopticsbv.com 71 Ophir Optronics Ltd www.ophiropt.com IBC Optarius Ltd www.optarius.co.uk 50 Optec SA www.optec.be 76 Optical Research Associates Inc www.opticalres.com 8 Optikos Corp www.optikos.com 59 Paul Hess KG www.stanfordcomputeroptics.com 66 PCO AG www.pco.de 60 Photonic Products www.photonic-products.com 75 Photop Technologies Inc www.photoptech.com 67 Physik Instrumente (PI) GmbH & Co KG www.pi.ws/oleb 21 Piezosystem Jena www.piezojena.com 46 Powerlase Limited www.powerlase.com 23 Primes GmbH www.primes.de 67 Quantel www.quantel.fr 65 Scitec Instruments Ltd www.scitec.uk.com 34 Sill Optics GmbH & Co KG www.silloptics.de 6 SIOS Metechnik GmbH www.sios.de 60 Southampton Photonics www.southamptonphotonics.com 35 Spectrogon AB www.spectrogon.com 46 Spere Optics-Hoffman www.sphereoptics.com 61 Spiricon Laser Beam Diagnostics Inc www.spiriconpower.com 69 Stanford Research Systems www.srsys.com 26 StockerYale Canada Inc www.stockeryale.com 46 StockerYale Ltd (IRL) www.stockeryale.com 71 TeraXion www.teraxion.com 49 Testbourne Limited www.testbourne.com 16 Toptica AG www.toptica.com Trioptics GmbH www.trioptics.com Umicore Coating Services www.coatingservices.umicore.com 72 Unaxis Balzers AG www.optics.unaxis.com 49 Wahl Optoparts GmbH www.wahl-optoparts.com XenICs www.xenics.com Xmark Media Ltd www.photonex.org 18

IEICE TRANS. COMMUN., VOL. E83-B, NO. 2 FEBRUARY 2000
An Architecture Supporting Quality-of-Service in VirtualOutput-Queued Switches

Rainer SCHOENENy, Member

SUMMARY Input bu ered switches most e ciently use
memory and switch bandwidth. With Virtual Output Queueing (VOQ), head-of-line blocking can be avoided, thus breaking the throughput barrier of 58.6%. In this paper a switch architecture based on VOQ is proposed, which o ers deterministic and stochastic delay bounds for prioritized tra c. This is achieved by a hybrid static and dynamic arbitration scheme, which matches ports both by a precalculated schedule and realtime calculations. By using weighted dynamic arbitration algorithms 100% throughput with lowest delays under all admissible tra c can be achieved. An integrated global priority scheme allows the multiplexing of realtime and data tra c. Following the arbitration decision, a cell scheduler decides locally in the input ports upon the next connection from which a cell is forwarded. Cell scheduling based on earliest-deadline- rst (EDF) is shown to perform similar to its behaviour in an output-queued switch. key words: ATM, VOQ, arbitration, allocation, scheduling
PAPER IEICE/IEEE Joint Special Issue on Recent Progress in ATM Technologies

1. Introduction

ATM networks will have to consist of switches supporting the individual per-connection Quality-of-Service (QoS) ATM promised to deliver. Even IPv6 switches o ering Di serv/Intserv need architectures with builtin QoS features. Among the switching architectures, input queued switches are most powerful because the access rate of crossbar and bu er memory is not higher than the line rate of the connected links. With Virtual Output Queueing (VOQ) 1], where each input manages a separate queue for each output, it has been shown that a throughput of 100% can be achieved 2]{ 4]. Arbitration algorithms resolve the contention for output ports in each time slot. The achievable throughput and delay performance heavily depends the used algorithm. Arbitration can principially be performed by two ways: Statically by assigning time slots to speci c connections or connection groups in advance or dynamically by resolving the contention for the same output port in each time slot anew. The static arbitration (also called allocation) reserves time slots for speci c connections in advance. This xed schedule o ers bandwidth guarantees, deterministic delay bounds for worst-case tra c and actively shapes (smoothes) the tra c. As shown in this paper,
Manuscript received May 28, 1999 Manuscript revised August 26, 1999 y The author is with the Institute of Integrated Signal Processing Systems (ISS) at Aachen University of Technology (RWTH) (www.ert.rwth-aachen.de).
static arbitration can guarantee delay bounds for trafc accepted by the connection admission control (CAC) of ATM by assigning a constant service bandwidth to each connection. This immediately supports CBR and VBR services, whose tra c parameters are known at CAC time. Similarily, IPv6 guaranteed service 5] with RSVP 6] is supported. Due to the boundedness of the departure process, bounds can as well be given end-toend. Analysis and results are in this paper. On the other hand, dynamic arbitration distributes the left-over bandwidth to contending ports. Weighted algorithms 3] o er the best delay performance compared to unweighted algorithms and they operate much better in more di cult than symmetric load con gurations 7]. Additionally, a static priority distinction between service classes must be integrated in a global arbitration method. A local priority decision within each input port alone is not su cient. In this paper a method for prioritized arbitration and its performance are treated. After the the centralized arbitration decision, the input ports are noti ed, which output port they have to send a cell to. A local cell scheduler within each input port must select the rst queued cell of a suitable connection for transmission to the given output port. Unlike other papers assuming FCFS service it is necessary to provide cell scheduling in the input ports within each priority class. Thus problems of individual QoS, ow separation and fairness can be addressed. The di culty is the complexity of the system involving the hierarchical decision of both arbiter and scheduler. In this paper EDF scheduling is shown to perform similar in the VOQ con guration as it does in outputqueued switches. Together with an approximative analytic model of the arbitration, the delay performance for FCFS and EDF can be obtained in a closed form. The paper is organized as follows. Section 2 discusses VOQ and arbitration algorithms and related work. The static arbitration is explained and analysed in Sect. 3. Dynamic methods are treated in Sect. 4. Section 4.3 contains performance results for the dynamic arbitration method. Based on an approximative analytic model of the dynamic arbitration in Sect. 5, cell scheduling is treated in Sect. 6.

Graph G

Arbiter Input 0 Switch Fabric
Matching M Outputs O Inputs I Outputs O 2

Inputs I 2

Output 0
Arbiter request for PORT CID PRIO

y cell ? n

3 wM-1,2

cell ?

scheduler for queue 0 M-1

Output M-1

scheduler

send cell

Input M-1

Fig. 1 VOQ architecture.

Fig. 2 Bipartite matching.

scheduler block

Fig. 3 Input scheduling.
2. VOQ System Description
The VOQ con guration 1] shown in Fig. 1 consists of a nonblocking switch fabric, an arbitration unit and M input and output ports. Each port has a link rate of rlink. Arriving cells on input port i are placed into the corresponding queue for their destination port o. This queue does not necessarily maintain cells in FCFS order, as often assumed, it can as well serve cells in weighted round-robin or earliest-deadline- rst (EDF) order 8], as shown in Sect. 6. Let its current queue size be qi o. The process of arrivals to this queue is characterized by a mean rate i o. Let the input and output loads be
M ;1 out = X i o Tslot o i=0 This is admissible 1] if 8o : M ;1 in = X i o Tslot (1) i o=0 out < 1 and 8i : in < 1. o i
In each time slot the arbiter selects unique pairs of input and output ports (a \match" (i o)) either by a lookup in the allocation table (if there is an entry) or by realtime calculations based on weights wi o sent to it from the input ports. This task is equivalent to a bipartite graph matching problem 1] as shown in Fig. 2. Static arbitration (allocation) has been used in traditional circuit-switched TDM systems, where the arrival and departure instances of frames are known in advance. TDM switches only have to precompute a periodic schedule to control the crosspoints in a switching network. For packet-switched networks (ATM) there is no frame reference time. Due to the asynchronous nature of packet tra c a periodic structure cannot immediately be exploited. Ideas to integrate precompute schedules for packet switches appeared in 9]{ 11]. As shown in Sect. 3, the constant bandwidth which is guaranteed per connection combined with the bounded trafc model in Sect. 3.1 o ers delay bounds per switch and end-to-end. Figure 3 shows that when the connection indenti er (CID) is allocated and a cell of that connection is available, no more decisions have to be taken. However, if there is no cell at all, the inputoutput port match cannot be changed and this slot is

3. Static Bandwidth Allocation
The delay guarantee concept of allocation is based on the individual (per-VC) guarantee of a service rate. This guarantee is rm, i.e. the probability Prfd > dmax gy to exceed a delay dmax is zero. This is in contrast to statistical delay bounds, where Prfd > dmax g is never zero but only tends to zero for higher dmax (typically with an exponential decay 18]) and the performance heavily depends on the total load of all other connections and their statistical properties (an undesired property). The requirement for rm delay bounds with allocation is that the tra c of the connection-ofinterest is bounded, as shown now. 3.1 Bounded Tra c Tra c bounds are common in integrated services networks. For ATM, the known tra c descriptor (PCR, CDV T ) bounds the peak cell rate over an interval, (SCR, MBS ) bounds the average rate over a longer interval. The contract conformance can easily be policed by the GCRA 19] algorithm. IPv6 o ers a TSpec in its RSVP message 20] during connection admission, which is very similar. The parameters are tted into the tra c bound model of Cruz 21], 22] which bounds the amount of tra c A (given by its rate function R) in an interval (t1 t2 ) by

n(CID) = ! r rslot

R(t)dt = A(t1 t2 ) < 1minnf i + i (t2 ; t1 )g(2) = <i< ==
This can be written R (~ ~) for n pairs ( i i ). Each pair bounds the tra c rate to i with a burst tolerance i. ATM o ers n = 1 for CBR and n = 2 for VBR in the tra c contract at the UNI 19]. Using TPCR = 1=PCR and Tlink = 1=rlink we give the following mapping of parametersyy
1 = BS = 1 + bCDV T=(TPCR ; Tlink )c 1 = PCR 2 = MBS 2 = SCR
To calculate rm delay bounds, a worst-case tra c model 23] using these parameters must be applied. It consists of periodically repeated bursts of length BS (MBS f. VBR) cells followed by silence such that the average rate during the period length is PCR (SCR). The bene t of this model is that delay bounds derived for the worst-case cannot be exceeded by any other trafc which also obeyes the tra c contract.

COVs = V ar( s )=E ( s )

CDF, a complementary distribution function. In units of cells and cells=second.
yyy Corresponds to the smallest supported rate, e.g. 64kb=s.

connection number

0.35 0.3 0.25 0.2

COV of egress traffic

0.15 0.1

S slots

COV of allocation
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Fig. 4 Allocation table.

Fig. 5 COVo vs. COVs.
Fig. 6 Allocation jitter. Table 1 Properties of Allocation.
It is a measure of the quality of the allocation in the sense of a smooth service time distribution. It is desirable to have COVs = 0 (G=D=1). The actual COVs depends on S , n(CID), the total number of other allocated (blocked) slots, the allocation algorithm and max (see Fig. 6). The output tra c stream is shaped such that it assimilates the service process. Thus COVs and the output tra c variation COVo are strongly correlated. COVo also depends on !(CID): Even with ideal allocation (COVs = 0), unused slots lead to gaps in the output tra c stream (Fig. 6) which adds one component to the COVo : Let for simplicity the random process slot used be i.i.d. and the probability of an unused slot be p = 1 ; (CID). Then the assumed geometrically distributed interdeparture time has p (11) COVo = pp = 1 ; (CID) For example, with !(CID) = 1:05, (CID) 0:95 the contribution is COVs = 0:218, which can be seen in Fig. 5. Here the tra c arrived with COVa = 0:4. The graph has been obtained by allocation of 240 connections. Each dot represents the values for one connection. The COVs due to quantization is
2 COVs quant = Tlink =12T 2
isolated, per-VC treatment (no in uence of other tra c) pre-determined constant delay bound ! QoS output tra c is shaped and bounded self-policing (output conforms to GCRA) QoS is adjustable with a single parameter ! simple model for stochastic tra c: G=Ek =1 can be used together with dynamic arbitration ! 100% utilization non work-conserving globally ! mean delay is higher hybrid arbitration o ers lower mean delay but higher miss ratio
the queue is empty (busy time) is less than the period of the periodic worst case tra c bounded by the given tra c parameters. Thus delay bounds can be given. A bound for CBR is derived here in continuous time. The delay a cell of bounded stream experiences is composed of the sum of two statistically independent terms de =delay until service if queue is empty and db =delay due to cell's position in burst. With a random phase, de is equally distributed (rect-shaped) over Q 1 the interval 0 T ]: PDF (de ) = pe (de ) = (de =T ; 2 ). The distribution PDF (db ) is a function shaped like a sequence of dirac ( ) impulses (kk): BS ;1 X p (d ) = 1 (d ; i (T ; T )) (14)
which is 0.015 here. Allocation jitter due to max (Fig. 6) is at most COVs = 0:33 here, which can be seen in Fig. 5. It is bounded by p COVs < max Tlink =( 3 T ) (13) = 3.4 Allocation Performance Allocation has a number of positive properties (Table 1) which can be quanti ed for given tra c bounds, i.e. during CAC the model parameters are known for connections demanding real-time QoS. When bounded (deterministic) tra c (section 3.1) is applied, an explicit delay bound is guaranteed (Fig. 8). Assuming a fully equidistant service (COVs = 0) with a service rate 1=T higher than PCR for CBR or SCR for VBR, the time between instances where

The resulting delay distribution is the convolution ( ) Q of these components: p(d) = pe (d) pb (d) = kk. For the example, with BS = 4 a sequence of four diracs convoluted with a rect should have the sandcastle shape, as the simulated result in Fig. 7 shows. The maximum value leads to this delay bound: max(dCBR ) < (BS ; 1) (T ; Tlink ) + T (15) = The situation for VBR is similar. One level of burstiness more makes it more complex, however. For a smooth worst case model 23] the same holds: max(dV BR ) < (MBS ; 1) (T ; TPCR ) + T (16) = For the heavy VBR modely Eq. (17) holds. TO := BS (T ; TPCR ) max(dV BR ) < b(MBS ; 1)=BS c TO (17) = + (MBS ; 1) mod BS ] (T ; Tlink ) + T
Maximal length bursts on link and PCR rate 23].

PDF(cell delay)

Tlink arrival traffic.. time

T service slots

Fig. 8 Worst case trafc.

100 cell delay [slots]

Fig. 7 Example PDF (d).

Fig. 9 Tra c bounds.

Tra c CBR CBR VBR

Fig. 10 fcompand.

Descriptors PCR = 64kb=s, CDV T = 5ms PCR = 1Mb=s, CDV T = 1:4ms PCR = 10Mb=s, SCR = 3:3Mb=s MBS = 1000 (MPEG-2)
When the assumption COVs = 0 is relaxed, an additional delay of CDV Tmax must be added. 3.5 End-to-End Delay Bounds With allocation an end-to-end delay guarantee dN over N hops can be given (Fig. 9). Lemma 1: For any switch n on the route and given CBRy input tra c parameters Rn ( n n ) the output is shaped such that its tra c is bounded by pairs Rn+1 ( n+1 n+1 ). Of the in nitely many possible pairs, two are of interest: Eq. (18) at the service time n !n and Eq. (19) at the declared n. 0 = PCRn+1 = n !n (18) n+= 1 + b2Tlink max =(TPCR ; Tlink )c n+1 n+= PCRn+1 = n (19) n+= 1 + bCDV T 00 =(TPCR ; Tlink )c n+1 max n+CDV Tmax = ( n ; 1)(TPCR ; T ) +T + 2 maxTlink In fact, after the rst allocating switch the output trafc is much smoother due to active shaping. What results are lower delay bounds for any following switch. Overallocation ! is not necessary for switches n > 1, because overload is impossible by construction. Theorem 1: For any limited number of hops n the total cell delay dsum n and its output tra c Ri+1 is bounded. Let Ri be the tra c bounds before switch i and di (Ri ) the resulting delay bound for it. Then holds
Table 2 Delay bounds for example realtime applications (with
maxCTD = 200ms, ! = 1) for switch 1 and the following (n).
max(d1 ) 5:5ms 1:76ms 107ms
max(dn ) 5:5ms 352 s 400 s
far. dn+1 is bounded because its input tra c Ri is bounded (Lemma 1) and allocation then guarantees eqs. 15 and 17. Thus Theorem 1 holds for all n. In Table 2 example values for typical voice/video tra c show the practical suitability of allocation. CAC rules can be obtained easily 11] by using the bounds above.

4. Dynamic Arbitration

Dynamic arbitration serves as a complementary mechanism to use unallocated slots for connections which do not require strict delay bounds or require a lower mean cell delay, which cannot be provided by static mechanisms. 4.1 Priority Support for Weighted Matching Weighted matching algorithms can achieve full throughput under admissible load 3]. However, with di erent priority classes a stable operation cannot be guaranteed. In fact an admissible load of higher priority cell streams can nevertheless lead to unbounded delay if lower priority (best e ort) connections experience congestion (high queue weight wi o ). When a smaller weight on another port represents real-time cells only, they have to su er. Only a globally prioritizing arbitration can achieve the correct separation. The goal is to have no in uence of lower priority queue weights over higher priority weights. A threedimensional extension of the known weighted matching algorithms is discarded because of the complexity. We aim for a method operating in two dimensions only. Consider a queue organization as shown to Fig. 1 and assume the queues are additionally separated per priority (with P priority levels, 1=highest). Let the

dsum n =

n X i=1

di (Ri )

rameters R1 , the rst switch at the network ingress has a bounded delay dsum 1 = d1 given by Eqs. (15) and (17). This is obviously the delay for N = 1. Let the theorem be true for switch n. Switch n +1 then has also a bounded delay output. Proof: The worst-case delay of switch n + 1, dn+1, adds to dsum n , the maximum delay guaranteed so
Proof (by induction): With given bounded tra c pa-
For VBR substitute SCR ! P CR.
100 mean cell delay [cell slots] SLIP-1 PIM 10 iMCFF
1000 mean cell delay [cell slots]
quantiles of the cell delay using FCFS with SIMP [slots]

PIM SLIP-1 iMCFF

100 sim: 1e-1 sim: 1e-2 sim: 1e-3 calc: 1e-1 calc: 1e-2 calc: 1e-3

SIMP MWM

output queueing
output queueing 0.0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 offered load 1

Similar to the

used for speech coding.
mean delay for two priorities [cell slots] 1 using SIMP using MWM 2 100

priority 2 f eq

mean cell delay [cell slots] for priority 1 (constant load) using MWM 1000 bit offset=0 bit offset=1 bit offset=2 bit offset=bit offset=100 bit offset=bit offset=bit offset=7 bit offset=1.7

priority 2

f sat f compand priority 1

priority 1

and f compand

0.5 (bit offset) 8

bit offset 8

0.1 0.2

0.6 0.7 offered load
Fig. 14 Priority distinction.
4.3 Performance Results for Arbitration In this section we compare the performance results for reference algorithms output queueing, MWM, iSLIP, PIM, iMCFF with SIMP using simulations for an switch with OPNET 24]. With Bernoulli tra c and a symmetrically chosen destination (1 of M ; 1), each input port carries a total o ered load of , i.e. i o = =(M ; 1). Bursty tra c is important for many reasons. According to 25] we use a packet train model producing a burst of B (t) cells at full rate followed by empty slots such that the mean rate i o = =(M ; 1) is the desired fraction of the input rate. B is exponentially distributed here the mean burst length B is 32. This corresponds to the mean length of an Ethernet PDU segmented into ATM cells. In Fig. 11 the results for PIM and iSLIP are as in 25]. The best case is output queueing which we want to approximate with VOQ. The ideal arbitration for VOQ is MWM which most closely approximates output queueing. As we see, iMCFF performs worse than SIMP. The performance of SIMP is in between iMCFF and MWM. In Fig. 12 we see the delay performance for the bursty scenario. The most important characteristics are: (i) the absolute delay is two decades higher than for the previous scenario due to the burst scale queueing e ect and the short-term asymmetries. (ii) PIM and iSLIP perform similar with more than 300 cell transmission times above = 0:6, i.e. noticably worse than the other algorithms. (iii) As before, iMCFF, SIMP, MWM and output queueing are quite close to each other with improved performance (less delay) in ascending order. We observe that the algorithms that decide based on weights (iMCFF, SIMP, MWM) perform very well for typical tra c. When using the rst priority mechanism feq the main parameter for achieving priority separation is the priority weight factor kq = 2bq (bq shift bits). The other two degrees of freedom in the choice of trafc parameters manifest in total = 1 + 2 and sp1 (share of rst priority cells) with 1 = sp1 total and

Fig. 15 Priority scenario.
2 = (1 ; sp1 ) total. The behavior for two priority classesy can be seen in Fig. 15, where 1 = 0:2 is held constant and 2 is varied as bursty tra c. We see that there is almost no in uence of the lower priority load on the higher priority performance when a wordlength of seven or more bits is used. To analyse quantitatively how many bits bp are needed we study the mean delay for both priority groups at a xed load total = 0:7 and a xed ratio sp1 = 0:2 as a function of di erent bp. The xed values used have revealed to be the most expressive. For bursty tra c Fig. 14 shows that about bp = 6 bits are needed to fully separate the priorities. We can verify that with enough bits dprio1 = 0:2 slots at load 1 = 0:2. This is exactly the delay for MWM with = 0:2 in Fig. 11. Thus we have an exact priority separation. The higher priority tra c performance does only depend on its total load. The necessary value of bp depends on the tra c in priority p. This is not desirable, so we apply the deterministic separation by saturation (Eq. (24)), where no overlapping is possible, i.e. wio p+1 cannot become higher than 2bp. We lose some of the weight dynamics in this priority because the values possible are in the interval 0 2bp ; 1]. With saturation we observe in Fig. 14 that now the separation is much improved for 1 : : : 5 bits. There is still some in uence typical for MWMyy. We can even choose one bit per priority and achieve a considerable separation. This, however, ignores all weights within a priority level and yields a bad performance under asymmetric workload as MSM does. With companding (Fig. 14) we nally observe the very best separation. Here three bits are su cient for a complete separation, whereas the dynamics is as good as pure MWM.
Higher priority tra c is Bernoulli, lower is bursty. The maximum matching is the one that maximizes the sum of the matched weights. This sum might be still higher for some lower priority matches (e.g. 7 + 7) than for the alternative match (e.g. 8 + 0).
probability density function of arbiter interservicetime with SIMP load rho=0.95
POISSON BURST E_6 exponential_approx

other input load

w=(M-1)/M 2 1

time slot

0.0001

T2 t3 Q_ic

t4 Q_oc

token_pool

2 w=1/M
distance [cell slots] 1e-60 80
traffic of interest correct ports connected *

token_pool_2 T4

other output load
Fig. 16 VOQ switch: PDF for interservice time of SIMP arbiter.
Fig. 17 PN system for VOQ.
tions the QoS demand per switch can be expressed as Prfd > dmax g < where and dmax are derived from ATM tra c and QoS parameters 19]. The delay quantiles are given in Eq. (28), and especially for Poisson tra c in Eq. (29). (28) d = E w ]] ln(2 E w]2 ] ) Ew Ew d Poisson = Tslot (2 +) ) ln( (23+ ) ) (29) 3(1 ; In Fig. 13 the simulated and calculated results for 2 f10;1 10;2 10;3 g are compared and show an acceptable accuracy. This has been used to obtain approximations for the FCFS scheduling performance in Fig. 19 at = 0:95.

5. Calculating VOQ Performance
In a VOQ con guration there are M virtual schedulers for each priority in each input port, one for each output. Each of these schedulers is only activated to serve a cell if induced by the arbiter. Thus the arbiter appoints the service interval, opposed to an OQ switch, where a cell is served in each time slot. In Fig. 16 the time between successive service events is shown for a number of congurations, e.g. di erent load and di erent tra c for SIMP. Observe the periodicity in multiples of M Tslot. For lower load or smooth tra c with SIMP an exponential distribution looks very similar, but assuming an M=M=1 queueing system is wrong because in the VOQ system there is no statistical independence between arrivals and service. The dependence is exactly what we want weighted matching algorithms to achieve. For determining the delay performance of the scheduler a performance model for the arbiter is needed. The delay quantiles can be obtained by simulation and analytical modelling. With a detailled stochastic PN system (Petri Net 26]) the arbitration can be modelled most accurately 8], but the stationary solution is computationally very intensive as the number of Markov states grow large. Alternatively a more abstract modelling of the reasons for a higher delay, the input and output con icts, provides quite good results. For the performance from port i to o this is modelled in Fig. 17 by having the input and output port loaded by in i and out respectively. With an independent probability o 1=M a token in Qic is transferred to Qoc. The performance is similar to two virtual M=D=1-type queues in series. Thus the calculated approximation can be obtained based on Eq. (26). Prfw < tg = 1 ; ae;bt (26) = E w 2 and b = 2E w] a = 2E w2]] (27) E w2 ] where the rst and second moment are obtained for the M=D=1=1=FCFS system 27]. For realtime applica-
6. Cell Scheduling in VOQ Switches
The reason for sophisticated cell scheduling algorithms is the global provision of an individual per-stream QoS and the decision which concrete cell to send in a slotby-slot timescale. In traditional output-queued (OQ) switches all cells for a speci c egress link are handled by one scheduler in the output port and there is almost no di erence between the multiplexer being fed by M input links with a rate of =M each or one input with a rate. For M ! 1 an FCFSy scheduler with Poisson tra c is appropiately modelled by an M=D=1 queueing system. In many cases the performance can be numerically calculated as the cell delay distribution PDF (d) in stationarity. More complex schedulers have evolved for the need to treat distinct tra c streams di erently. Here we concentrate only on the most promising algorithm EDF y 29], 30], because the principal properties of VOQ scheduling are to be shown. EDF o ers a parameter D V C ] which has the meaning of deadline. With a

y FCFS= rst come rst served 28], EDF=earliest deadline rst.

maxCTD / s 5 4

FCFS scheduling in VOQ switch
R0D0 R0D1 R0D2 R1D0 R1D1 R1D2 R2D0 R2D1 R2D2

GCIDs D2

approximation
R2 R1 R0 1/R 1/Rate 1/(4R) 1/(2R)
Fig. 18 Rate V C ] and dmax V C ].
properly chosen parameter D V C ] = dmax V C ]+ c this very successfully guarantees individual statistical delay bounds. For the graphs shown in this paper the slot time is assumed to be T = 1 s in a switch. The analyzed simulation scenario consists of 2160 streamsy with nine di erent characteristics in rate and deadline, as shown in Fig. 18. All ports are homogeneously and symmetrically loaded with = 0:95 (Poisson tra c). The results in Fig. 19 and 20 show that in principle the connection separation works the same way as in an OQ architecture. For all loads these results can be obtained using the delay quantiles in Fig. 13. For EDF, a good approximation for the CDF is given by Prfw < tg = 1 ; ae;b(t+D0 ;D0i ) (30) = using the mean deadline value 27] and the shape parameters a, b obtained from the VOQ model in Eq. (27). Using Eq. (30) and a, b from Eq. (27), the performance can be calculated. The simulated graphs in Fig. 20 show the typical behaviour: A separate performance for each delay class, the distance is exactly the di erence of the deadlines. The slope of the graphs is exactly the same as for FCFS, which validates that our approximation is applicable. As it can be seen, EDF supports deadlines as it does in an output-queued switch.

P i D0 = P i D0i i i

1e-delay [cell slots] 250 300
Fig. 19 VOQ switch: Prfd > tg for FCFS.
1 EDF scheduling in VOQ switch 0.1

1e-05 0

200 delay [cell slots]
Fig. 20 VOQ switch: Prfd > tg for EDF.
the performance of which depends on other connections, however. Scheduling algorithms can further separate connections and operate similar to their output-queued variant. With the performance approximations for the SIMP algorithm given in this paper the performance of more complex schedulers can be derived, as shown for EDF here.

Acknowledgements

The autor would like to thank the anonymous reviewers for their valuable comments.

7. Conclusion

It is shown that in a switch using virtual-outputqueueing (VOQ) a number of arbitration algorithms can be applied. Static allocation o ers deterministic, i.e. rm, delay bounds for bounded tra c per-VC. However, the mean delays can become quite high. Dynamic algorithms perform quite close to the ideal output queueing. With weighted algorithms priorities can be supported and statistical delay bounds can be given,

References

1] A. Mekkittikul and N. McKeown, \A Starvation-free Algorithm For Achieving 100% Throughput in an Input-Queued Switch," in Proc. of the IEEE International Conference on Communication Networks, 1996. 2] N. McKeown, Scheduling Algorithms for Input-Queued Cell Switches. PhD thesis, UC Berkeley, 1995. 3] N. McKeown, A. Mekkittikul, V. Anantharam, and J. Walrand, \Achieving 100% Throughput in an Input-Queued
16 ports 15 destinations each 9 classes.
10 Switch," IEEE Transactions on Communications, vol. 47, Aug 1999. A. Mekkittikul and N. McKeown, \A Practical Scheduling Algorithm to Achieve 100% Throughput in Input-Queued Switches," Proceedings of the IEEE INFOCOM, 1998. S. Shenker, C. Partridge, and R. Guerin, \Speci cation of Guaranteed Quality of Service." IETF RFC 2212, Sep 1997. J. Wroclawski, \The Use of RSVP with IETF Integrated Services." IETF RFC 2210, Sep 1997. R. Schoenen, G. Post, and G. Sander, \Weighted Arbitration Algorithms with Priorities for Input-Queued Switches with 100% Throughput," in Proceedings of the IEEE Broadband Switching Systems, 1999. R. Schoenen and R. Hying, \Distributed Cell Scheduling Algorithms for Virtual-Output-Queued Switches," in Proceedings of the IEEE GLOBECOM, 1999. J. Y. Hui, Switching and Tra c Theory for Integrated Broadband Networks. ISBN 0-7923-9061-X: Kluwer, 1990. T. Anderson, S. Owicki, J. Saxe, and C. Thacker, \High Speed Switch Scheduling for Local Area Networks," ACM Transactions on Computer Systems, vol. 11, Nov 1993. R. Schoenen and G. Post, \Static Bandwidth Allocation for Input-Queued Switches with strict QoS bounds," in Proceedings of the IEEE Broadband Switching Systems, 1999. C. H. Papadimitriou and K. Steiglitz, Combinatorial Optimization. Prentice-Hall, Inc., 1982. Y. Tamir and H.-C. Chi, \Symmetric Cross Bar Arbiters for VLSI Communication Switches," IEEE Transactions on Parallel and Distributed Systems, vol. 4, no. 1, pp. 13{27, 1993. C. Lund, S. Phillips, and N. Reingold, \Fair Prioritized Scheduling in an Input-Bu ered Switch," in Proceedings of the IEEE International Conference on Broadband Communications, 1996. R. Schoenen, G. Post, and G. Sander, \Prioritized Arbitration for Input-Queued Switches with 100% Throughput," in Proc. of ATM Workshop'99, 1999. R. Schoenen, G. Post, and A. Muller, \Analysis and Dimensioning of Credit-Based Flow Control for the ABR Service in ATM Networks," in Proceedings of the IEEE GLOBECOM, 1998. Vol.4 p.2399. D. Stephens and H. Zhang, \Implementing Distributed Packet Fair Queueing in a Scalable Switch Architecture," in Proceedings of the IEEE INFOCOM, 1998. L. Kleinrock, Queueing Systems, Vol. II: Applications. New York: John Wiley & Sons, 1976. The ATM Forum, Prentice Hall, Englewood Cli s, N.J., ATM user-network interface (UNI) speci cation version 3.1, 1994. S. Shenker and J. Wroclawski, \General Characterization Parameters for Integrated Service Network Elements." IETF RFC 2215, Sep 1997. R. Cruz, \A Calculus for Network Delay, Part I: Network Elements in Isolation," IEEE Transactions on Information Theory, vol. 37, p. 114, Jan 1991. H. Michiel and K. Laevens, \Teletra c Engineering in a Broad-Band Era," Proceedings of the IEEE, vol. 85, p. 2007, Dec 1997. J. Roberts, U. Mocci, and J. Virtamo, eds., Broadband Network Teletra c: Performance Evaluation and Design of Broadband Multiservice Networks nal report of action COST 242. LNCS1155, Springer, 1996. MIL 3, Inc., 3400 International Drive, NW Washington, DC 20008, USA, OPNET Release 6.0, 1999. http://www.mil3.com. N. McKeown and T. Anderson, \A Quantitative Comparison of Scheduling Algorithms

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]

for Input-Queued Switches," unpublished, 1997. available at http://http.cs.berkeley.edu/%7Etea/atm.html. M. Marsan, Modelling with Generalized Stochastic Petri Nets. Wiley, 1996. ISBN 0-471-93059-8. B. Walke and W. Rosenbohm, \Waiting-Time Distributions for deadline-oriented Serving," Performance of Computer Systems, p. 241, 1979. North-Holland Publishing Company. L. Kleinrock, Queueing Systems, Vol. I: Theory. New York: John Wiley & Sons, 1975. F. Chiussi and V. Sivaraman, \Achieving High Utilization in Guaranteed Services Networks using Earliest-DeadlineFirst Scheduling," in Proceedings of the International Workshop on QoS, 1998. S. W. Lee, D. H. Cho, and Y. K. Park, \Improved dynamic weighted cell scheduling algorithm based on Earliest Deadline First scheme for various tra cs of ATM switch," in Proc. IEEE Globecom '96, (London), pp. 1959{1963, 1996.
was born in Dusseldorf, Germany, on March 9, 1970. He received the DiplomIngenieur degree in electrical engineering from RWTH Aachen University of Technology, Germany, in 1995. Currently he is a Ph.D. candidate in the ISS department of RWTH Aachen. His research interests include tra c management for QoS supporting networks, switch architectures, credit-based ow control and stochastic Petri nets. He is a member of the IEEE.

Rainer

Schoenen