Mobile Edge Computing for 5G: The Communication Perspective Kaibin Huang Dept. of Electrical & Electronic Engineering The University of Hong Kong Hong Kong Joint Work with Yuyi Mao (HKUST), Changsheng You (HKU), Jun Zhang (HKUST), Khaled B. Letaief (HKUST)
Next-Generation Mobile Networks Communication, Computing and Content Delivery 5G Networks Tactile-scale latency (milliseconds) Ubiquitous Gigabit wireless access Secure Device-to-device High mobility 2
Mobile Edge Computing (MEC) MEC provides cloud-computing capabilities at the edge of the mobile network (base stations, access points, set-top boxes ) in close proximity to mobile subscribers. The environment of MEC is characterized by low latency, proximity, high bandwidth, real-time, location and context awareness. 3
MEC Application: Face Recognition Offload computation-intensive components to MEC server MEC server Offloading Should be computed locally Image acquisition Digital image Classification results Face detection Classification Face image Pre processing Normalized Face image Database (training samples) Feature extraction Feature vector MEC server
MEC Application: Augmented Reality Offload computation-intensive components to MEC server MEC server Offloading Resolution Image format Should be computed locally Video Source Renderer Tracker Mapper Object recognizer # Features MEC server
MEC/Fog Network Architecture Increasing propagation latency vs. reducing computation latency 6
Mobile Computation Offloading (MECO) Mobile computation offloading can save energy when the following inequality is satisfied p m w s m >p c d i B + p i Local computing w : workload s m : computation speed of the mobile d i : data size B : bandwidth p m : computation power of the mobile p c : transmit power of the mobile p i : idle state power of the mobile s s : computation speed of the cloud w s s Mobile computation offloading Communication overhead Never offload Depending on bandwidth B Always offload Computation load 7
Integrating Computer Science and Wireless Communication Mobile Devices Radio Access Network Computer Network Computation Throughput Wireless Communication Radio resource management Multiple access control Physical layer techniques (Massive MIMO, mmwave ) Small-cell networks Computer Science Network function virtualization Computing offloading Cloud computing Software architecture Parallel computing Live migration 8
Resource Management in MEC Systems 9
Joint Radio-and-Computational Resource Allocation Radio resources Channel? What to compute? Computational resources Save energy Centralized resource allocation Distributed resource allocation 10
Joint Radio-and-Computational Resource Allocation Centralized framework: TDMA system, limited-capacity MEC server T Offloading/local computing T Cloud computing and downloading Edge cloud Offloading based on TDMA Mobile 1 Offloading priority function on local computing energy consumption, channel gain and data size Mobile 2 Optimal policy structure: Threshold-based binary offloading decision Mobile K Policy Structure Descending order Offloading Priority Resource allocation Minimum offloading threshold Full offloading Offloading time-sharing duration Cloud computation capacity C.You, K.Huang, H.Chae,and B.-H.Kim, Energy-efficient resource allocation for 11 mobile-edge computation offloading, IEEE Trans. Wireless Commun., 2016
MEC Server Scheduling Asynchronous arrival Varying Latency requirements Sensitive Tolerant Time Task Dependency for Individual Users Buffer 12 Image recognition Source: http://ochoa.pc.cs.cmu.edu/wschu/project_fr.html Tools: queuing theory, integer programming
Multiuser Cooperative Edge Computing Balance Detection D2D Computational resources Overload! 13
Peer-to-Peer Mobile Cooperative Computing Motivation: Scavenging nearby real-time idling computation resources # of bit Helper s CPU-idling profile CPU idling profile CPU profiling User Computation input-data Helpers 0 s0 s 1 s 2 s 3 s 4 s 5 s 6 Time An energy efficiency dilemma: Longer time for transmission reduce TX energy but scavenge less peer CPU resources, and vice versa Offloaded Data Buffer size # of bit Offloading feasibility tunnel Energy-efficient TX by String-pulling Buffer size Maximum offloaded bits Primary CPU idling profile 0 Computation Deadline Time C.You and K.Huang, Mobile cooperative computing: Energy-efficient peer-to-peer 14 computation offloading, submitted to Trans. Wireless Commu., 2017.
Issues, Challenges, and Research Opportunities 15
Cache Enabled MEC Spatial/Temporal popularity driven caching Service Caching Heterogeneous resources allocation CPU-hungry Memory-hungry Storage-hungry Data Caching Reduce the computation latency 16
Mobility Management for MEC Mobility-aware Offloading Using D2D Communication Random Waypoint Model for Mobility Task Task K+1 Task K Mobility-aware Live Prefetching Mobility-Aware Server Scheduling Task K-1 Task K s candidate p(1) p(2) p(3) p(l) CT 1 CT 2 =Task K CT 3 CT L Task Prediction Prefetch Task Prediction Prefetch S.-W. Ko, K. Huang, S.-L. Kim and H. Chae, Live prefetching for mobile 17 computation offloading, IEEE Trans. Wireless Commun., 2017.
Green MEC: Dynamic Right-Sizing VM VM VM VM Dynamic Right Sizing Energy efficient, but Switching and VM migration cost Degradation of QoS High maintenance costs VM VM VM Accurate profile of computation workload at edge cloud should be forecasted. VM VM Increasing cell coverage Turn-off VM Migration 18
Green MEC: Renewable Powered MEC Random Energy Arrival Process Random Task Arrival Process Green energy-aware resource allocation & computation offloading - Channel & energy & workload information Large-scale renewable powered MEC system - Spatial diversity of the available renewable energy 19
Green MEC: Geographical Load Balancing Geographical load balancing for MEC: leverages the spatial diversities of the workload patterns to make workload routing decision among different edge-clouds. Park Park Advantages: Improve the energy efficiency of the lightly-load edge servers and user experience. Prolong the battery lives of mobiles. Some efficient resource management techniques at edge-cloud are used, such as VM management, dynamic right-sizing, etc. Restaurant 20
Green MEC: Wirelessly Powered MEC Cloud Mobile (e.g., wearable computing device) MPT Offloading Mode 2 MPT Offloading Controllable CPU cycles Mode 1 1/f 1 1/f 2 1/f N MPT 0 T Local Computation t T time time Motivation: Using wireless energy to power MEC Objective: Maximize energy savings given computation deadline constraint Offloading or not Energy beamforming Double near-far problem Multiuser scheduling C. You, K. Huang and H. Chae, Energy Efficient Mobile Cloud Computing 21 Powered by Wireless Energy Transfer, IEEE Journal Sel. Area Comm., 2016
MEC for 5G 22
5G MEC Platform MEC Platform Overview Each edge-cloud server consists of a hosting infrastructure and an application platform. 23
5G MEC: Challenges and Requirements Application Portability Operation Network Integration MEC System Performance Security Resilience Regulatory and Legal Consideration 24
5G Applications of MEC MEC for video stream analysis MEC for auto-driving Tracker Mapper Object Recognizer MEC for augmented reality service 25
5G Specifications for MEC Traffic selection UPF selection Session and service continuity Information exchange with AF QoS control and charge for routing 26
How does 5G enable MEC? Support Service Requirement MEC Latency-sensitive application Mission-critical service Small packet error rate 5G Advanced Mobility Management Strategy (AMMS) 5G Core Network Determine Monitor & update Mobility Pattern Decide AMMS 27
How does 5G enable MEC? Capability of Network Slicing Business Layer User and industry business applications Entertainment Slice Network Function and Application Store Service Layer Network Applications Network Applications Unified Management and Orchestration IoT /Smart grid Slice Network Functions Dedicated User Plane Cloud Infrastructure Layer Network Functions Dedicated User Plane ehealth/public safety Slice RRH Network Storage 28
Is PHY Layer Dead? Asked by Mischa Dohler, Robert Heath, Angel Lozano, Constantinos Papadias, Valenzuela in 2011. Not yet it is an exciting information age of C s Computing Communication Control Content delivery. 29
Reference: Y. Mao, C. You, J. Zhang, K. Huang, K. B. Letaief A Survey on Mobile Edge Computing: The Communication Perspective Online: https://arxiv.org/abs/1701.01090 30