Mobile Edge Cloud Services in 5G Yanyong Zhang WINLAB, Rutgers University yyzhang@winlab.rutgers.edu
Edge clouds, edge applications MOTIVATION
Mobile Edge Clouds WINLAB
Edge Applications Mobile Edge Cloud Services Emergency Service Safety Smart Factory Automatic driving AR navigation Real-time alert Automatic translation Field machine control Real-time Emergency stop Scalability for response devices and objects WINLAB
Challenges Complexity Terminal Mobility Distributed, Parallel, Data- Multi-sensory and Pipelined - Device/Object ID Intensive fusion - Sensed data Computation Query Object Valu Object Valu ID e (1) Sensing Object Valu GUID e GUID e 1 X1 (2) Computation (2) Computation 1 X1 1 X1 2 X2 2 X2 (3) Actuation Cloud 2 X2 … … Service … … … … . . . Network 10^12 Response Directories 10^12 domain 10^12 - Actuation command Large # of NW domains Real-time Scale <100msec response time in application <100msec response time in application Trillion-order edge devices & objects are Trillion-order edge devices & objects are properly handled properly handled WINLAB
Network virtualization, ASR, task mapping, spatial and temporal scheduling SMART-EDGE
SmartEdge Vision WINLAB
Global Name Resolution Globally unique name Sue’s_mobile_2 (GUID) for network attached Server_1234 Taxis in NB Media File_ABC objects Sensor@XYZ John’s _laptop_1 device, content, context, AS name, sensor, and so on Host Sensor Context Content Multiple domain-specific naming Naming Naming Naming Naming Service Service Service Service services Globally Unique Flat Identifier (GUID) Global Name Resolution Service Global Name Resolution Service for GUID NA mappings Hybrid GUID/NA approach Network Both name/address headers in PDU “Fast path” when NA is available GUID resolution, late binding option Net2.local_ID Network address Net1.local_ID WINLAB
Mobility Service via Name Resolution Service API capabilities: Register “John Smith22’s devices” with NCS - send (GUID, options, data) Name Certification Options = anycast, mcast, time, .. Services (NCS) - get (content_GUID, options) GUID assigned Options = nearest, all, .. GUID lookup from directory GNRS update NA99 (after link-layer association) MobilityFirst Network (Data Plane) Send (GUID = 11011..011, SID=01, data) NA32 GNRS GUID <-> NA lookup GNRS query GUID = 11011..011 Represents network Send (GUID = 11011..011, SID=01, NA99, NA32, data) object with 2 devices DATA GUID SID NAs Packet sent out by host WINLAB
Overview of Virtual Mobile Cloud Network (vMCN) Name resolution layer (3) Service Anycast & Dynamic Migration smartEdge slice (1) Name-based virtual network (2) VN & vBS on VN Routing layer the move ID Query Packet Virtual base station (vBS)
MF Extended to Support VN (3) GNRS Service Plane (1) Virtual GUID as an identifier of a VN (1) Virtual GUID as an identifier of a VN (2) Ingress router identifies VNs (2) Ingress router identifies VNs GUID Locator VN Type (3) GNRS is exploited to capture VN (3) GNRS is exploited to capture VN A1 19 No membership and access control membership and access control VNID_1 VR1, VR2, … Yes VR1 1a Yes VR2 1b Yes Central Coordinator (1) Network 19 A1 VR2 VR6 A2 VR1 S (2) Ingress VR3 VR5 Router B1 Network 53a
vBS: A Technique of Building a Service-specific VN in WiFi Isolated BS resources assigned to a target service Isolated BS resources assigned to a target service Elastic QoS Network-driven association and handover Network-driven association and handover Controllability Cooperative NW virtualization in backhaul Cooperative NW virtualization in backhaul Unified Policy vBS 2 vBS 3 vBS 1 Virtual BS BS Resource Pool (vBS) Physical BS The other services K. Nakauchi and Y. Shoji, “WiFi Network Virtualization to Control the Connectivity of a Target Service,” IEEE Transactions on Network and Service Management , June, 2015
Application-Specific Routing Virtual Routing table @ VR1 Service GUID is assigned to a group of replicated Service GUID is assigned to a group of replicated <App, Link, N_Hop> servers servers DST_ GUID Application-level metric is used for routing Application-level metric is used for routing A1 <0.2, 5x, VR2> decision decision A2 <0.7, 5x, VR2> Application state is periodically announced Application state is periodically announced B1 <0.3, 3x, VR3> computation load, waiting time, etc. computation load, waiting time, etc. S <0, 1, S> Network 19 A1 VR2 VR6 Service Y A2 VR1 S VR3 VR5 B1 Network 53a
Edge Cloud Assignment – Spatial Scheduling An edge application consists of a sequence of requests each having a deadline and demanding a certain amount of resource Each edge cloud periodically exposes its available resources CPU utilization, data set, pending workload, Then we pick one that can satisfy the real-time constraint; if no server is qualified, we consider migrating existing requests (without violating their deadlines) Delayed scheduling Migration may be needed due to user mobility Mobility prediction WINLAB
Parallel, Distributed, and Pipelined Execution WINLAB
Efficient Runtime Management Data shipping or computation shipping? Approximate computing to hide network latency Co-scheduling across multiple sites WINLAB
PROTOTYPING & EVALUATION
Initial Prototype MobilityFirst MobilityFirst vBS vBS vBS remote control software is implemented in vBS remote control software is implemented in the MF stack the MF stack Common C-plane for MF and vBS Common C-plane for MF and vBS Written in Python Written in Python Service vBS Prototype vBS Prototype MF-VN Prototype MF-VN Prototype GUID 1331 MF Central VNCS & Coordinator OFS & MM CPS Terminal Python Cloud VR2 script (MF-enabled) Server1 GUID 21 vBS 172.21.0.254 VR1 Non-CPS/CPS VR3 Cloud Traffic Generator Server2 GUID 22 NA: 172.21/16 NA: 172.25/16 NA: 172.27/16 Locator: XX Locator: YY Locator: ZZ WINLAB
90%ile Response Time Measured MF-VN MF-VN vBS vBS • Routing: ASR • Mode: 802.11n/a - Reporting interval: 2s • Band: 5GHz MF VN MF VN • Transport: hop-by-hop • Tx Rate: 65Mbps reliability • Rx Rate: 65Mbps - Retrans. limit: unlimited • # of BS: 2 (Ch 36,48) - Retrans. timeout: 200ms APP-level Query & Response (mfping) response query Response Time T=t 1 T=t 0 Before After 90% CDF vBS&MF vBS only w/o vBS&MF 90%ile Response Time Response Time (RTT) WINLAB
WiFi Congestion &Dynamic Server Load Server#1 Server Load Before CPS BS#1 0. Terminals 2 T MF VN MF VN Server#2 Server Load 1. Non-CPS 0 terminals T Without vBS Shared and competed Shared and competed Forwarded to the Forwarded to the Without ASR b/w CPS and Non- b/w CPS and Non- closest server even closest server even CPS apps CPS apps if the load is high if the load is high Server#1 Server Load After CPS BS#1 0. Terminals 2 T MF VN MF VN Server#2 Server Load 1. Non-CPS 0 terminals T With vBS With ASR Isolated for a specific Isolated for a specific Forwarded to Forwarded to 20 CPS app CPS app less-loaded server less-loaded server WINLAB
Reduction of 90%ile Response Time from 1900ms to 210ms Emulated cloud app with 50KB Data, w/ Server Load and WiFi Cross Traffic vBS&ASR w/ vBS & ASR 214ms 1932ms 90%ile Due Due to retransmission w/o vBS & w/o ASR s and the 0.64 potential bogus state in the MF transport protocol protocol 0.24 100 WINLAB
Questions & Answers WINLAB
Recommend
More recommend
