Computing Power Network
for MEC
Yunpeng Xie
China Telecom
2019.10.15
Computing changes the world, computing drives the future
2
Micro world Daily life Macro world
Computing is a pivotal driver of the intelligent world
Steam power Electric power Computing power
12 years >> 1 day 1 day >> 7 days 2.2 M >> 13.7 Bn
human whole gene sequencing weather forecast cycle cosmic observation distance(light-years)
MEC will be the focus in future network
MEC solves the problem of long delay and bandwidth
consumption caused by centralized cloud computing and
provides high bandwidth, low latency, and strong security
for vertical industries.
MEC is considered as a combine point of 5G, industrial
internet and IOT, which will grow rapidly , the global edge
network economy will exceed 4.1 trillion dollars by 2030.
Source:USA ChetanSharmaConsulting
Car networking Life Entertainment
Smart cityIndustrial internet
3
4
Connected
Network
The network is for connection service
The value is connecting people and things
Cloudification
Network
The network is for cloud services
The value is connecting the cloud and the network
Computing
Power Network
The network is for computing service
The value is the release of computing power
MEC+AI
Cloud computing
+ Virtualization
MEC+AI promote Computing Power Network
What is Computing power ?
5
➢The Computing power refers to the ability to provide services for the calculate
tasks of users.
① Application is computing power :Users
directly call applications in cloud DC, such
as VR rendering
② Algorithm/module is computing power :
Users call some algorithms or modules in
cloud DC to serve their own business
③ Basic resources is computing power:
Users only need the basic vm resources
provided by cloud DC
User
StorageCompute Network
COTS
StorageCompute Network
COTS
VM containerBare metal
APP APP
Module Module Module
①
②
③
Cloud DC1 Cloud DC2
StorageCompute Network
COTS
StorageCompute Network
COTS
VM containerBare metal
APP APP
Module Module Module
The Computing Scheduling Problem Faced by MEC
6
Edge Cloud A
Edge CloudB
Idle Occupied
The characteristics of edge application put
forward more requirements for flexible
computing scheduling crossing-edge cloud.
Thus a corresponding solution is needed
to be provided by carriers.
Edge Application Characteristics:
➢ High Mobility: The applications location often
changes. Flexible adjustments to user access point
are required.
➢ High time variability: Th e app l i c a t io n traffic
changes significantly with time.Flexible adjustments
to computing resource are required.
➢ Requirement enhancement: The application has
obviously requirement enhancement for low latency
and high bandwidth. User SLA guarantee is required.
AS-IS: The Computing Scheduling in Cloud Computing
StorageCompute Network
COTS
StorageCompute Network
COTS
VM Container Bare-Metal
Co
re C
lou
d
Man
ag
em
en
t Syste
m
APP
Backup 3
StorageCompute Network
COTS
StorageCompute Network
COTS
VM ContainerBare-Metal
Lig
htw
eig
ht C
lou
d
Man
ag
em
en
t Syste
m
APP
Backup 1
Edge Cloud
Core Cloud
7
➢ Solution description
⚫ Elastic capacity expansion and contraction: In a resource
pool, an elastic capacity expansion and contraction
strategy is made according to VM's load.
⚫ Defect:It is a computing adjustment within one resource
pool, and it lacks computing scheduling and service
coordination between resource pools.
⚫ Load-balancing : In one or more resource pool(s), it
creates multi-app backups and makes load-balancing
policy to adjust the service computing allocation among
multi pools.
⚫ Defect: It lacks network cooperation. The best service
processing node can not be determined without user
access information. Hence, it is difficult to meet the low
latency requirement.
user
APP
Backup 2
TO-BE: Computing Power Network
Fixed Access(Family)
Fixed Access(Enterprise)
MobileAccess
NB-IoT Access
Wi-Fi Access Edge
Computing
Resource
Edge Cloud Core Cloud
Core Computing Resource
User
Provides multi-access edge
network to achieve flexible service
access.
vBNG/UPF
Provides high-performance
service node, to meet service
computing requirement.
Meets the service requirement on low-
latency, high bandwidth and security
by moving down the service gateway.
Edge Cloud collaborates with Core
Cloud to adjust computing allocation
flexibly.
Build a computing scheduling solution based on
cloud, network and edge deeply collaboration. 8
Key Capability of Computing Power Network
9
➢ Computing Resource Modeling: It needs to abstract the computing resource,
constructs resource model with computing resource, network resource, storage
resource and even algorithm resource, then provides them as product to users.
➢ Computing Control and Management: It needs to control the computing resources
of core cloud and edge cloud centralized, allocates computing resources to services
from a global perspective, to achieve computing resource control and management.
➢ Computing Power Scheduling: It needs to have the capabilities to allocate and
schedule computing resources flexibly and elastically, thus quickly responding to and
adjusting to the computing requirement.
➢ Service Guarantee: It needs to have the capabilities to predict the change of service
computing requirement, and adjust the computing resource, guarantee the user SLA in
the case of service requirement changing.
Architecture of Computing Power Network
10
• analyzes user service requirement, turn into computing resource requirement , determine deployment location and resourcesbased on requirement
• selects AI service for user service in AI Empowerment Platform,
to determine the deployment specification of user service
• deploys the service which in AI Empowerment Platform on the
specified node
• manages the computing resources, allocates corresponding computing, storage and network resources, and adjusts flexibly the service deployment location and service computing resources on the basis of policy
• manages the networks of users, edge cloud and core cloud, deploy service gateway in the same location with user service, to route service traffic to the processing node
11
Process Flow of Computing Power Network
Edge Node
StorageCompute Network
COTS
ComputingResource
ComputingResouce
vBRAS/UPF
Step1Service Request
APP
Step3Function Select
Step4Gateway and Function node
Deploy
Step5Data Input
Step7Result Output
F1
F2
F3
Edge Node
StorageCompute Network
COTS
ComputingResource
ComputingResource
vBRAS/UPF
Core Node
StorageCompute Network
COTS
ComputingResource
ComputingResource
vBRAS/UPF
ComputingResource
MAN MAN
Access Network
Step6Computing Task Distribute
and Data Write Back
Step2Intelligence Requirement
Analyze
Computing Network Control and Management Orchestrator
Requirement Intelligent
Analysis module
Platform capability scheduling module
Lifecycle management
module
Em
po
werm
en
t pla
tform
12
IDLE
Latency-Sensitive Data
ChinaNet CN2
Data Flow Direction
Regional DC
Computing Network C&M
Orchestrator
AI Platform
5GC
COTS
Container Bare-MetalVM
vBRAS UPF
AI
Pla
tform
COTS
Container Bare-MetalVM
vBRAS UPF
COTS
AI
Pla
tform
VM ContainerBare-Metal
Core DC
Edge DC
Integrated Access Office
OLT
MSE MSE
CR CR
A A
B B
ER ER
CPEeNB/gNB
IDLE
Face Recognition Service
Traffic Prediction
Face Recognition Service
Traffic Prediction
AI
Pla
tform
Cloud Management
System
LightweightCloud
Management System
Non-Latency-Sensitive Data
e.g.- Flexible Computing Power Scheduling
Cloud Management
System
Image
Recognition
Image
Recognition
Image
Recognition
Data Annotation
Data Annotation
Data Annotation
13
Latency-Sensitive Data
ChinaNet CN2
Data Flow Direction
Regional DC
AIPlatform
5GC
COTS
Container Bare-MetalVM
vBRAS UPF
Image
Recognition
Data Annotation
AI
Pla
tform
COTS
Container Bare-MetalVM
vBRAS UPF
COTS
AI
Pla
tform
VM ContainerBare-Metal
Core DC
Edge DC
Integrated Access Office
OLT
MSE MSE
CR CR
A A
B B
ER ER
CPEeNB/gNB
Traffic Prediction
AI
Pla
tform
Non-Latency-Sensitive Data
e.g. - Flexible Computing Power Scheduling
BUSY Traffic Prediction
Face Recognition Service
Face Recognition Service
Computing Network C&M
Orchestrator
BUSY
Cloud Management
System
Cloud Management
System
LightweightCloud
Management System
Data Annotation
Image
Recognition
Image
Recognition
Data Annotation
THANKS