Cloud-Aided Wireless Networks with Edge Caching: Fundamental Latency Trade-Offs in Fog Radio Access Networks Ravi Tandon Osvaldo Simeone ISIT 2016, Barcelona 1
Introduction Content delivery, e.g., video, is driving growth in wireless traffic 2
Introduction Content delivery, e.g., video, is driving growth in wireless traffic Edge- vs cloud-based solutions 3
Introduction Cache-aided wireless network (or edge caching): storage of popular content at wireless edge nodes [Golrezaei et al 12] EN EN EN EN EN EN: Edge Node Cache Reduces latency due to backhaul usage
Introduction Information-theoretic analysis of cache-aided interference channels - Achievable 1/DoF for 3 3 system [Maddah-Ali and Niesen 15] - Bounds on 1/DoF for more general models with caching also at the receiver [Naderializadeh et al 16] [Hachem et al 16] [Xu et al 16]
Introduction Cloud-aided wireless network (or C-RAN): Centralization of baseband processing at the cloud cloud fronthaul EN EN EN EN EN EN: Edge Node
Introduction cloud fronthaul EN EN EN EN EN (Digital) fronthauling approaches: - Hard fronthaul transfer [Patil and Yu 14] - Soft fronthaul transfer: Fronthaul compression [Simeone et al 14] Centralized interference management
Fog-RAN (F-RAN) Fog Radio Access Network (F-RAN): Cloud and cacheaided wireless network for content delivery cloud fronthaul EN EN EN EN EN EN: Edge Node Cache
Fog-RAN (F-RAN) Optimal operation of an F-RAN: complex design problem over fronthaul, cache and spectral resources Fundamental trade-off between delivery latency and system resources
System Model 10
System Model 11
System Model 12
System Model Quasi-static channel model with continuous distribution Power constraint P 13
System Model 14
System Model 15
System Model 16
Cache-Fronthaul-Edge Policy Tx interval Time
Cache-Fronthaul-Edge Policy Tx interval Caching interval Time Cache storage policy: What to cache - No knowledge of instantaneous users requests and CSI - No inter-file coding (intra-file coding allowed) file cached content at EN k,
Cache-Fronthaul-Edge Policy Tx interval Time Cache storage policy: What to cache Fronthaul policy: What to transmit on the fronthaul links - Based on instantaneous users requests and CSI
Cache-Fronthaul-Edge Policy Tx interval Time Cache storage policy: What to cache Fronthaul policy: What to transmit on the fronthaul links Edge transmission policy: What to transmit on the wireless channel - Based on instantaneous users requests and CSI
Normalized Delivery Latency Serial fronthaul-edge transmission Tx interval Fronthaul TF Wireless Delivery time per bit (e.g., [Liu and Erkip 11]) TE (, C, P) max lim F user's requests L T F T L E Time
Normalized Delivery Latency Serial fronthaul-edge transmission Tx interval Fronthaul TF Wireless Delivery time per bit (e.g., [Liu and Erkip 11]) TE (, C, P) max lim F user's requests L T F T L E Time Normalized Delivery Time (NDT): (, r) lim P (, rlog P, P) 1/logP Ideal system: interference-free and unlimited caching
Normalized Delivery Latency Pipelined fronthaul-edge transmission Tx interval Wireless Fronthaul T Delivery time per bit and NDT (, C, P) max lim F user's requests L T L and (, r) lim P Time (, rlog P, P) 1/logP Practical implications in, e.g., [Leconte et al 16]
Main Result: NDT for 2 2 of F-RANs Theorem: The minimum NDT for the 2 2 F-RAN with is given as 1 2 max 1, 2 for 0 r 1 * r (,r) 1 1 for r 1 r 24
Main Result: NDT for 2 2 of F-RANs 25
Main Result: NDT for 2 2 of F-RANs 26
Main Result: NDT for 2 2 of F-RANs 27
Main Result: Achievability Full caching: Cooperative zero-forcing beamforming at the ENs 28
Main Result: Achievability No caching: Zero-forcing beamforming at the cloud + softtransfer fronthauling (compression with bits/ sample) 29
Main Result: Achievability Caching of half file: Interference alignment on an X-channel [Motahari et al 14] [Cadambe and Jafar 09] 30
Main Result: Converse 31
Main Result: Converse 32
Main Result: Converse 33
Main Result: Converse 34
Main Result: Converse Information cut 1: Information cut 2: Information cut 3: Linear combinations of the inequalities above yield the desired result 35
Conclusions and Outlook F-RAN leverages the synergy and complementarity of cloud processing and edge caching Definition of NDT as high-snr worst-case latency relative to an ideal system Characterized the NDT for a 2 2 system Extensions (see arxiv w/ Avik Sengupta): - General lower and upper bounds - Characterization of NDT for a general F-RAN within a multiplicative gap of 2 - Extension to pipelined model Open problems: partial connectivity, imperfect CSI,