Optics for the Cloud

Optics for the Cloud

Publications

Events

Opportunities

Overview

Optics for the Cloud is a programme of research to advance and enable the adoption of optical technologies in the rapidly growing field of cloud computing.  The scope of this opportunity includes storage, network, compute, memory and specialized accelerators. The research also spans the entire vertical stack, ranging from optical device fabrication, to optical system and sub-system design, to co-design and integration with the rest of the data center infrastructure and will include new applications enabled by optical technologies.

The world is moving to the cloud at an ever-increasing rate. Microsoft is at the core of this transformation and is continuously building new data centres to keep up with demand. However, most of the core technologies used in these centres today were designed or conceived before the cloud existed. Many of those technologies represent compromises, encumbered by historic design choices pertinent to an antiquated application, and encrusted with features appealing to the idiosyncrasies of many different scenarios. At the cloud scale, there is the scope to create cloud-specific technologies that free us from decades of legacy thinking.

At Microsoft Research Cambridge we are building a team to explore Optics for the Cloud. We believe that by taking a holistic view of the needs of the cloud we can create disruptive technologies for the cloud. To achieve this, we are assembling a highly inter-disciplinary team to ask the simple question – how can optics for the cloud change the cloud? We hope to invent optical end-to-end systems, across the three primary resources of storage, network and compute, that will underpin the next generation of the cloud.

Explore careers in optics research

Projects

Project Iris

This project explores novel designs of regional and Wide Area (WAN) cloud networks from the ground-up.

Project Silica

Project Silica, a collaboration with the University of Southampton is a ground-breaking cloud storage system that uses femtosecond lasers to write data on to quartz glass which, due to its durable properties, provides long-lasting, easily-readable storage.

Project Sirius

This project aims to develop an all-optical, data-center-wide network that is completely flat, in contrast to the hierarchy of electrical switches used today.

Storage

Investigating new optics-based approaches to storage

Data storage in the cloud is deployed in multiple tiers based on the type of customer workload that it is best suited for. There are usually at least three or four tiers, and the important workload characteristics are the frequency of datum access (equivalent to the ratio between operations and stored capacity), the latency (delay) permissible for providing data when requested, and the size of the individual data units. Tiers are given names which draw parallels with temperature, as shown below:

TierAccess frequencyLatencyCurrent media
Hothighsub-millisecondflash
Standard and Coolmediumup to tens of millisecondshard disk drive
Coldlowminutes to hourstape

Cloud providers present pricing models based on abstracted marginal costs, which enable customers to engage in informed rational behaviour when selecting the appropriate tier. Due to the extremely large data collections used in modern cloud applications, data storage is a very cost-sensitive application.

The rapid expansion of cloud computing, and the corresponding data volumes, is also driving the demand for storage. To date, for both HDD and tape, this demand has been met by advances in the achieved information storage density of the magnetic material. However, the storage density is now very close to fundamental physical limits imposed by the size of the magnetic domains that can be manipulated and successfully read. In the Cold tier, the limited storage lifetime of tape (3-5 years) is also problematic, as data must be regularly transferred from old media to new to ensure data integrity.

To keep up with the increasing demand for archival storage, it is necessary to investigate new approaches to storage that provide properties that are better than or equal to today’s technology, provide greatly extended data lifetimes and reduce the cost per byte. The read/write throughput should be around 100M-1GByte/s to be competitive with existing technology and support the demand. In the Standard and Cool tiers, HDD-based systems are sometimes limited by the number of operations that can be performed per second, so approaches to storage that can improve this rate will be advantageous.

Project Silica

Project Silica, a collaboration with the University of Southampton is a ground-breaking cloud storage system that uses femtosecond lasers to write data on to quartz glass which, due to its durable properties, provides long-lasting, easily-readable storage.

Project Silica video

In September 2017, Mark Russinovich, CTO of Azure, announced a number of new collaborations with MSR Cambridge including Project Silica. See Mark present the impressive early results (watch from 1:01:13 – 1:03:18).

Network

Research background: Network

Network requirements, both in terms of throughput and latency, are expected to increase significantly in the next few years, driven by new emerging workloads such as large-scale machine learning and resource disaggregation.  These will demand throughputs beyond 100 GB/s and latency below 100 ns.  Hitherto, the industry has relied on exponential scaling of silicon to keep up with increasing traffic demands, but with the expected slowdown of Moore’s Law, novel network technologies with new growth curves that can match the requirements of future workloads must be investigated.

Our research aims to re-invent the network that will underpin the cloud in the year 2025 and will meet these key demands via emerging optical technologies. Optics already plays a major role in our technical network infrastructure as the underlying technology used to interconnect racks and data centres across the globe. Our goal is to further improve optics technology by innovating across the whole stack, from new network architectures to new transceivers and optical fibers to deliver higher performance at a lower cost as well as extending its use to new areas such as optical switches and board interconnects to take advantage of their ultra-low and predictable latency, very high bandwidth, and low cost.

This effort requires a cross-disciplinary approach combining skills and expertise across the hardware, physics, network, and software domains. If successful, however, this has the potential to revolutionize the cloud infrastructure by providing predictable and uniform high performance (bandwidth and latency) within and across data centres, thus breaking today’s silos (e.g. a single server or a rack) with significant benefits in terms of fault tolerance, resource management and application performance.

Project Iris

This project explores novel designs of regional and Wide Area (WAN) cloud networks from the ground-up.

Project Sirius

This project aims to develop an all-optical, data-center-wide network that is completely flat, in contrast to the hierarchy of electrical switches used today.

Compute

Research background: Compute

The rapid expansion of data centres creates a new research opportunity, in that techniques which are only relevant at large scales, and technologies which require exotic form-factors, become realistic and even attractive options in a data centre setting.

For example, the optics community has long known that lenses apply a Fourier transform to the light waves travelling through them.  Thus, a lens computes a function over its input data.  In a traditional von Neumann computer, such a Fourier transform would be computationally expensive to calculate yet is almost instantaneous with light: the advantages to computing with optics are clear.

Optical computing, in a limited form, is close to reality: as a matter of fact, all-optical matrix multiplication was first demonstrated in the 1970s.  More recently, in the field of machine learning and artificial intelligence, deep neural nets (DNNs) are becoming widely adopted.  These link simple processing steps together into layered networks: could these restricted yet popular AI kernels be computed using light?  Even non-linear functions required for DNN’s such as ReLU and tanh may have optical analogues which would allow DNNs to scale and be trained efficiently; computing with light is naturally a parallel processing model, and of course heat dissipation is much less of an issue.  While many hard challenges need to be addressed to get these whiteboard ideas closer to reality, it is also important to think about other functions

We seek to spur cross-disciplinary research between the optics community and computer scientists at this exciting interface between physics and computing.  What new computation models make sense?  Are there new learning frameworks more amenable to implementation using optics?

Talks and Workshops

Talks

Benn Thomsen on “Optics for the cloud: Opportunities for Light” at UCL Optics Summit, June 2017.

Paolo Costa on “Bridging the Last Mile for Optical Switching in Data Centers” at Photonics West (January 2018) and OFC (March 2018).

Ariel Gomez Diaz on “Optics for the Cloud” at Downing College, May 2018.

Austin Donnelly on “How would you store a Zetta-Byte of Cold Data?” at vETC, June 2018.

Hitesh Ballani on “Optics for the Cloud: Opportunities and Challenges” at OSA Advanced Photonics Congress (July 2018), Microsoft Faculty Summit (August 2018), and ECOC (September 2018).

Ant Rowstron on “Rethinking Data Storage for the Zettabyte Cloud Era: The Journey from Metal to Glass” at OSA Advanced Photonics Congress, July 2018.

Ioan Stefanovici on “Glass: A New Media for a New Era?” at Usenix HotStorage, July 2018.

Benn Thomsen on “Challenges and opportunities for photonics in the cloud” at University of Southampton, Future Photonics Hub Industry Day, September 2018.

Scarlet Schwiderski-Grosche on “Microsoft Optics for the Cloud – A New Approach to Data Centre Technology” at PASC19, June 2019.

Events

September 23, 2018 | Benn Thomsen workshop at ECOC 2018 WS05 Data Centres – 1, Session 1: “Has the time come for coherent optics in the data centres?

September 23, 2018 | Benn Thomsen workshop at ECOC 2018 WS12 Data Centres – 2, Session 2: “Optical technologies for distributed edge and exascale core data centres?

November 1, 2018 | Scarlet Schwiderski-Grosche, Paolo Costa, Ariel Gomez Diaz session at Future Decoded 2018Microsoft Optics for the Cloud – Rethinking Data Centre Technology

March 3, 2019 | Paolo Costa workshop at OFC 2019  “Opportunities and Challenges for Optical Switching in the Data Center

March 26, 2019 | Austin Donnelly at Tech Storm 2019 “Glass:A New Media for a New Era”

Optics for the Cloud Research Alliance

Microsoft Research Cambridge has also brought together its collaborations with several university research groups in Europe and the US to create the Optics for the Cloud Research Alliance. This Alliance will investigate and develop optical technologies for the cloud. The Alliance will take a multi-disciplinary approach to tackle big and open challenges, with an aim to enable greater adoption of optical technologies in cloud infrastructure, including cloud storage, network and compute.

The research will span a significant part of the cloud infrastructure ecosystem, ranging from photonics fabrication, via optical system and sub-system design, to co-design and integration with the rest of the data centre stack. The Alliance will provide a vibrant collaborative research environment, involving experts across optical and cloud research disciplines, with an open publications policy.

The Alliance funds a cohort of PhD scholars. The three-year PhDs target long-term problems and aim to advance the state of the art through forward-looking fundamental research that will be published in top technical venues. Each PhD student is hosted and supervised by a university faculty member and co-supervised by a Microsoft researcher. The students will also benefit from discussion, insights and feedback from practitioners at Microsoft and, more broadly, through greater interaction with partners in industry and academia.

Goals of the Research Alliance

The primary goals of the alliance are to:

  • Leverage and enable true cross-disciplinary research
  • Support high impact pre-competitive research
  • Aspire to generate disruptive, not sustaining, ideas

The secondary goals are to:

  • Help to expand the ecosystem
  • Communicate the value of optics and photonics research in UK and Europe
  • Send a message that we view our members as world leaders in this space

Founder Members of the Research Alliance

  • Aston University
  • Cambridge University
  • Ecole polytechnique fédérale de Lausanne
  • Eindhoven University of Technology
  • Southampton University
  • University College London
  • University of California, Santa Barbara

People

Group Members

Interns, Visitors, and Collaborators

  • Portrait of Dan Blumenthal

    Dan Blumenthal

    Visiting Researcher, Summer 2017

    UCSB

  • Portrait of Kari Clark

    Kari Clark

    Intern, Spring 2016

    UCL

  • Portrait of Chris Dainty

    Chris Dainty

    Consultant

  • Portrait of Adam Funnell

    Adam Funnell

    Intern, Summer 2016

    UCL

  • Portrait of Thomas Gerard

    Thomas Gerard

    Intern, Fall 2016

    UCL

  • Portrait of Hussein Kassir

    Hussein Kassir

    Intern, Summer '15

    École Polytechnique Fédérale de Lausanne

  • Portrait of Qi Li

    Qi Li

    Intern, Summer 2015

    Columbia University

  • Portrait of Alana  Marzoev

    Alana Marzoev

    Intern, Summer 2018

  • Portrait of Jayashree  Mohan

    Jayashree Mohan

    Intern, Summer 2018

  • Portrait of Theofilos Petsios

    Theofilos Petsios

    Intern, Fall 2017

    Columbia University

  • Portrait of Nadesh  Ramanathan

    Nadesh Ramanathan

    Intern, Spring 2017

  • Portrait of Philip Watts

    Philip Watts

    Visiting Researcher, Fall 2016

    UCL

  • Portrait of Aaron Zhao

    Aaron Zhao

    Intern, Summer 2017

    University of Cambridge

Microsoft Azure

  • Portrait of Jeff Cox

    Jeff Cox

    Partner Dir Software Eng

    Microsoft

  • Portrait of Jamie Gaudette

    Jamie Gaudette

    Principal Network Eng Manager

    Microsoft

  • Portrait of Dave Maltz

    Dave Maltz

    Partner Group Software Engineering Manager