ABSTRACT: The expectations on 5G networks is quite high in terms of data rates, latency, mobility, energy efficiency, and service agility. Optimizing wireless/mobile networks in any one of these dimensions falls short of meeting the peak targets in other dimensions. Mobile network programmability will be a key component in the 5G world to provide the right level of trade offs for an individual service and/or subscriber, while attaining a good degree of statistical multiplexing over the physical substrate and resources. In my talk, I will first overview the architectural journey we took from 2G to 4G mobile data networks in light of 3GPP standards. After highlighting some of the key pain points in the current LTE architecture towards next generation services, I will focus on two narrow problems in the context of ultra dense small cell architectures, where small cells share the same wireless front-haul. The first problem investigates whether the network programmability brings about any performance benefits over the smart end point solutions. The second problem looks into the problem of small cell topology control based on peak demand estimation.

BIO: Ulas C. Kozat is a senior member of technical staff at Huawei R&D in Santa Clara, CA. At Huawei, he leads several R&D and open source projects in the areas of Software Defined Mobile Networks and Network Function Virtualization. Previously, he worked as principal research scientist at Argela USA in Sunnyvale, CA and as principal research engineer atDOCOMO Innovations (formerly DOCOMO USA Labs) in Palo Alto, CA. He received his Ph.D. from the University of Maryland, College Park, his M.Sc. from the George Washington University, Washington DC, and his B.Sc. from the Bilkent University, Ankara, all in Electrical & Electronics/Computer Engineering. Dr. Kozat is a senior member of IEEE. He has published extensively and holds several patents in the areas of network modeling; cross-layer optimization; and architecture, protocol, algorithm design.

• current GPRS: fixed mob. magmnt,
• fix scheduling policy
  • machine learning
• LTE EPC issues;
  • north south traffic, internet bound
  • ignoer capabilit of new transport proto: MP TCP, DASH…
  • ignore context of user, network, app
  • wrong coupling of DP, CP, hinder cloudification

• New platform: services are vertically integrated: VR IoT..

Connect with networks experts, hear from them

• zeitgeist: Mobile cloud native
  • 1-2 years: 5g trials IoT, C-ran, slicing, hosting UP/CP functions onf SW/routers/GW
  • 3-5 year: MC native ap, context based comm., SON big data, dsitributed policy ctrl&enforcement, per app/user customizaiton

    • Techno: OPNFV, Ostack, MCORD, OCP, MEC
    • Context is important
    • actions based on cell capacity, quality: static map, historical data.
      • same location different users, same user diff. locations
      • (diagram, simple use case –> good talking)
      • flatting out network peaks

    Concreate problems

• Programming hetnet flows
  • Q: does network programmability bring any benefit over end-to-end solutions?
  • Fix scheudling policy (PFS) –

    > programmable inteer group (PFS) / intra group (Max C/I) **policy**

    • no splitting –> flow splitting, demux
    • Mininet testbed (missing LTE emulator), emulator only emulate rates based on releases (7.2 Mbps), only emulate CQI, channel state is being impl.
      • emulate links with capacities.
    • Average throughput distribution over 100 Markovian scenarios
      • y-axis: CDF, x-axis: through put (Mbps)
      • gain with Multi path (30mbps) vs programmable scheduler (40mbps)
      • no gain w/o programmability
    • Real LTE traces in tokyo
      • net gain/loss (in %)

    • Small cell topology control
      • Motivation: dense small cells
      • scenario: macro provide coverage, smacell provide capacity
      • issue: fronhaul capacity, energy eff. interference –> solution: selective power control of small cell
      • Tracking pak demand locations
        • issue: peak demand location shift quite unpredictably.–> historical data not reliable.
        • approach:
          • iterate over smallcell technolog → system of linear equation#
          • ..
        • Estimating demand
        • y: hit ratio, x: iteration number