Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND

Page created by Jacob Gibbs
 
CONTINUE READING
Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND
LE PIATTAFORME AI-ON-DEMAND
COME FATTORE DI INNOVAZIONE NELLE PMI

Marco Aldinucci
Dipartimento di Informatica, Università of Torino

      fortissimo2                                   1
Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND
OUTLINE

•   Motivations

•   HPC4AI vision
    •   Cloud federation + Artificial Intelligence + BigData Analytics

•   HPC4AI: A one-stop-shop for AI and BDA
    •   Data + Processes + Performance
    •   Novel blockchain-based federated accounting structure

•   The universities, the PMIs and the Innovation
Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND
BRAIN DRAIN: HIGHEST OUTBOUND/
INBOUND RATIO IN SCIENTIFIC RESEARCHERS

                                                                                                Inbound
                                                                                           Outbound

  Hannah Yan Han, “Analyse the migration of scientific researchers”. Toward Data Science, 2018 https://towardsdatascience.com/analyse-the-migration-of-scientific-researchers-5184a9500615
Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND
MOTIVATIONS
Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND
MOTIVATIONS: AI/BDA SERVICES ARE OPERATED BY MICROSOFT, AMAZON, ETC.

•   Proprietary solutions produce lock-in. And continue to exact a rent
    •   The long-tail of this rental economy model increases inequalities [J.E. Stiglitz]

•   AI require GPUs/TPUs. Their cost is high in commercial clouds
    •   Buying-vs-renting break-even point occurs very early for AI

•   AI and BDA need annotated datasets that require human effort for curation
    •   Local availability of these dataset is an enabling feature for training/analysis

•   Datasets contain sensitive information, they should be stored appropriately
    •   Compliant with European regulations (including privacy, e.g. GDPR) if within EU borders
Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND
Possible                      Arm     Cisco       IBM       NVidia        E4Engineering     Loquendo        Consoft
Partners
                    BlueReply        CSP       AizoOn     O.R.S.     Reply     Exemplar     NetValue       Celi     LabInf

                           Comau        FCA      Leonardo      Prima Industrie     General Motors       IREN       TIM

                       Intesa SanPaolo        Reale Mutua      Banca Sella       TopIX     TorinoWireless      INRiM       SITI
                                                                                                                                  AI-in-demand platform
Technological                            Collegio          Città Salute        ISI                       IIT          Human
                GARR     INFN-TO                                                            ISMB
Partners                               Carlo Alberto        e Scienza       Foundation                 Genova       Technopole

Universities,   Informatics                                    ICxT              HPC                               Control and
                                                                                                                                  Facts
Departments,                                                Innovation
                                                                                                                                  • INFRA-P call Nov. 2017
                                       UNITO                                    Center                              Computer
and             Economy                                                                        POLITO              Engineering
                                       (coord)
Inter-                                                    MEDHIUM
departmental
Centers         Philosophy
                                                         Digital Media
                                                                                SmartData                    Electronics and      • Ranked 1st on ~30
                                                                                                                Telecom
                Maths          Law
                                                      C3S Scientific
                                                       Computing
                                                                                 Center
                                                                                                               Engineering          submitted projects
Data Centers                        C3S            PdF                                    SmartData          HPC
                                                                                                                                  • Kick-off mid apr 2018
                                                                                                                                  • 4.5M€ funding
                                                                                                                                  • 2 partners
Hardware
                          free        experimental
                                                           Federated cloud islands
                                                                                                   experimental
                                                                                                                       free       • 8 associated partners
                         cooling                          GARR cloud “Piemonte zone”                                  cooling
                                        islands
                                                                                                                                  • Coord. M. Aldinucci
                                                                                                     islands
                                    (non federated)                                              (non federated)

                                                                                                                                  • Many industrial
National/EU            INDIGO                                         GARR
Networks              DataCloud
                 DEEP-HybridDataCloud
                                                 INFN-TO         national network         cloud bursting
                                                                                                                   Amazon
                                                                                                                  Azure, etc        stakeholders
                                                                   for research
Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND
HPC4AI: THE TURIN’S CENTRE ON HIGH-PERFORMANCE COMPUTING FOR
ARTIFICIAL INTELLIGENCE

•   Facilitate scientific research and engineering in the areas of
    Artificial Intelligence and Big Data Analytics
    •   Support large scale experimentation of applications
    •   Engage regional industry in joint research projects, also boosting their R&D
        capabilities
    •   Gather and store dataset with specific local/EU value (medical, business, code, …)

•   Focusing on methods for the on-demand provisioning of AI and
    BDA cloud services
Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND
HPC4AI: A ONE-STOP-SHOP FOR AI AND BDA

•   On-demand Services - Exploitation through cloud abstractions to target different users
    •   Vertical: Rapidly prototyping + possibly to move the solution down for tuning/optimisation
    •   Horizontal: Marketplace+ federation at each levels of solutions and data

•   Data - to store data & datasets
    •   Securely, preserving ownership, always available

•   Processes - to develop new protocols
    •   Cross-pollinating computer science, other sciences and engineering
    •   Automation and continuous improvement of processes

•   Performance - to ease the access to HPC
    •   For Machine Learning and BigData on-demand
Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND
Users                    Kind of service             Services                          Artifacts

Domain experts with no skills on ML and    Service-as-a-        SaaS for ML and BDA            Market place for ML and BDA
BDA.                                       Service (SaaS)       designed within HPC4AI         services: Dashboards, trained
                                                                partners                       models in several domains (NLP,
Training set not required. Off-the-shelf                                                       Vision, …)
algorithms/networks.

Domain experts skilled on ML and BDA.      Platform-as-a-       PaaS solutions for ML and      Market place of VMs and
Not expert in parallel computing.          Service (PaaS)       BDA directly designed within   Platforms realising software
                                                                HPC4AI or companion            stacks for ML and BDA. Solutions
New networks or pipelines; training set                         projects                       for data ingestion, data lake, etc.
required.

Researchers, cloud engineering, ML and BDA 1) Infrastructure-   1) GARR/other cloud able to    1) Openstack, docker, VM, object
framework designers, cloud engineers, stack as-a-Service (IaaS) support federation             storage, file storage, kubernetes,
and automation designers.                                                                      etc.
                                            2) Metal-as-a-
                                            Service (IaaS)      2) Job scheduler for HPC       2) Alternative cloud, job queue,
                                                                resources                      Big Data Stack (Spark, …).
Researchers, run-time designers.            Hardware            Bare Metal                     Multicore, GPU, storage, network,
                                                                                               switch, UPS, cooling, etc.
Marco Aldinucci Dipartimento di Informatica, Università of Torino - LE PIATTAFORME AI-ON-DEMAND
IaaS   PaaS
EXAMPLE PAAS ON BARE METAL: KUBERNETES-AS-A-SERVICE
MANAGEMENT OF FEDERATED MARKETPLACE WITH BLOCKCHAIN

                                  Researchers                        Marketplace

                                                                      services &
                                                                       datasets

                                                                                   BlueReply     CSP    AizoOn
                       Projects
      Legal entities                            UNITO             POLITO                 Comau    FCA     Leonardo

                                                                                    Intesa SanPaolo    Reale Mutua
                                       hpc4ai
                         HPC4AI         coins

                                                          cryptocurrency
                             hpc4ai free                    exchanger
                               coins

      Data Centers                         C3S           PdF         SmartData     HPC

                                                        hpc4ai blockchain
EXAMPLE: NEXT GENERATION DATA MANAGEMENT FOR LIFE SCIENCE

1. Start from existing data
  •   Store, share, sell: under the full control of owner

2. Imagine AI support
  •   For scoring, support and automation - not for diagnosis

3. Evaluate annotation, improve the processes
  •   Imagine how to improve protocols/processes and annotation to be better automatised and AI-
      supported

4. Develop novel analysis techniques and GoTo 3
  •   Continuous improving of processes by cross-pollinating computer and life science
THE BURNOUT OF ITALIAN RESEARCH ECOSYSTEM

•   Italy has no universities in the top 100, good/excellent
    groups are patchily spread across different universities

•   Italian universities are becoming a selection machinery of
    intelligent youngsters to be sent to foreign PhD and
    research centres

•   Industry is sliding toward assessed technologies and
    becoming progressively unable to innovate
WHY ITALIAN UNIVERSITIES ARE BECOMING A SELECTION MACHINERY OF
INTELLIGENT YOUNGSTERS TO BE SENT TO FOREIGN PHD AND RESEARCH CENTRES?

•   Con la cultura non si mangia [GT]

•   Mandare il curriculum? Meglio giocare a calcetto [GP]

•   Volevo vincere il Premio Nobel per l'Economia. Ero anche bravo, ero... non dico lì lì per farlo, però ero
    nella giusta... ha prevalso il mio amore per la politica, ed il Premio Nobel non lo vincerò più anche se ho
    buone possibilità di diventare presidente della repubblica [RB]

•   Cara studentessa, io, da padre le consiglio di cercare di sposare il figlio di Berlusconi o qualcun altro del
    genere; e credo che, con il suo sorriso, se lo può certamente permettere [SB]

•   In Italia i fondi per la ricerca non sono più bassi, a livello pubblico, della media europea [MR]

•   Alla costruzione del tunnel tra il Cern ed i laboratori del Gran Sasso, attraverso il quale si è svolto
    l'esperimento, l'Italia ha contribuito con uno stanziamento oggi stimabile intorno ai 45 milioni di euro
    [MSG]
HPC4AI & THE BURNOUT

•   HP4AI is primarily a modern, large research laboratory
    •   We don’t provide production computing, we don’t compete with Amazon, Cineca, etc.
    •   We are interested to experiment new solutions at different layers in the stack of cloud
        abstractions, and to create a protocol to move them between different layers (i.e. a
        software engineering methodology)

•   We want to keep, develop and transfer to our students and
    industrial partners the knowledge
    •   On how to create, manage, develop, innovate with the cloud/HPC/AI/BDA
    •   We will not use any technology without fully understand it
BUSINESS PLAN
PREVIOUS PROJECTS ON PARALLEL COMPUTING @UNITO
THE LAST 8 YEARS

•   Infrastructure
      •   HPC4AI (POR-FESR 2014-2020): Turin’s centre in High-Performance Computing for Artificial Intelligence
          (2018, 24 months, total cost 4.5M €).
      •   C3S: Competence Center on Scientific Computing (Compagnia di San Paolo, founding 900K €).

•   Research
      •   OptiBike (EU I4MS): Robust Lightweight Composite Bicycle design and optimisation, experiment of EU i4MS Fortissimo2 project (2017, 24 months, total cost 230K €).
      •   Toreador (EC-RIA, H2020, ICT-2015-16): TrustwOrthy model-awaRE Analytics Data platfORm (2015, 36 months, total cost 6.2M €).
      •   Rephrase (EC-RIA, H2020, ICT-2014-1): Refactoring Parallel Heterogeneous Resource-Aware Applications – a Software Engineering Approach (2015, 36 months, total cost
          3.5M €).
      •   REPARA (EC-STREP, 7th FP): Reengineering and Enabling Performance And poweR of Applications (2013, 36 months, total cost 3.5M €).
      •   IMPACT (founded by Compagnia di San Paolo): Innovative Methods for Particle Colliders at the Terascale (2012, 36 months).
      •   ParaPhrase (EC-STREP, 7th FP): Parallel Patterns for Adaptive Heterogeneous Multicore Systems (2011, 42 months, total cost 4.2M €).

•   Networking
      •   HiPEAC (EC-NoE, 7th FP & H2020) European Network of Excellence on High Performance and Embedded Architecture and Compilation (2012-now).
      •   cHiPSet (EC-COST Action IC1406): High-Performance Modelling and Simulation for Big Data Applications (2015, 48 months).
      •   NESUS (EC-COST Action IC1305): Network for Sustainable Ultrascale Computing (2014, 48 months).
      •   DIMA-HUB (EU I4MS): Regional Digital Manufacturing Innovation Hub (2016, 6 months).
EuroPar 2018
                                     Co-chairs: M. Aldinucci, L. Padovani, M. Torquati
Torino, Italy — 27-31 August 2018
You can also read