Natarajan Venkataraman, Principal Engineer, Ericsson India
Da t a pl a n e f o r SUBSCRIBER GATEWAY OVERVIEW & CHALLENGES - - PowerPoint PPT Presentation
Da t a pl a n e f o r SUBSCRIBER GATEWAY OVERVIEW & CHALLENGES - - PowerPoint PPT Presentation
Da t a pl a n e f o r SUBSCRIBER GATEWAY OVERVIEW & CHALLENGES Natarajan Venkataraman, Principal Engineer, Ericsson India Su bsc r iber Ga t ew a y Entity at edge of access-aggregation network / mobile RAN network Provides
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 2
› Entity at edge of access-aggregation network / mobile RAN network › Provides Internet access to subscribers › Enables ISPs to monitor usage for charging › Primary Role
–AAA: Authenticate, Authorize subscribers and Account their traffic (for charging etc.) –Provide IP-address to subscribers –Provide various services as per SLA
› Example Nodes
–BNG / BRAS: Fixed broadband –P-Gateway: Mobile broadband
› In talk we will take the BNG as a representative of a subscriber gateway
Su bsc r iber Ga t ew a y
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 3
Depl o ymen t Po in t s - BNG
Access Network
Aggregation Network
Core
Lower scale but higher throughput per subscriber Higher scale but lower throughput per subscriber Can be at edge of access network Can be at edge
- f aggregation
network
BNG BNG
SUBSCRI BERS SUBSCRI BERS SUBSCRI BERS
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 4
Su bsc r iber Ga t ew a y a n d Su ppo r t No d es
SUBS GW
AGGR ROUTER CORE ROUTER
PCRF/ AAA
ACCT
DSLAM
OLT
WIFI AP/AC
AGGR ROUTER
DHCP Server
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 5
› Good packet throughput (like any router) › High session setup / teardown rate (sessions/sec) › High subscriber scale
–At each dataplane and at system level
› Rich services, service orchestration › Session survivability support
Desir ed Ch a r a c t er ist ic s
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 6
› Continuous session signaling packet transfer between control plane and dataplane › Continuous provisioning messages and responses between control plane and dataplane › Continuous session keep-alives with subscribers › Continuous push of statistics counters from dataplane to control plane and thence to accounting servers › Loss of these will be impactful Poor product perception
–Low session setup rates, Session flaps, Subscriber dissatisfaction, Revenue loss
Da t a pl a n e - Co n t r o l Pl a n e In t er a c t io n
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 7
› Services made available to a subscriber can change at any time
–Pre-paid model: Once limit is reached, service needs to be deactivated –Value-added services: Subscribers can purchase at anytime and expect instant activation –Time-based services: Increase in bandwidth or usage during peak / off-peak hours
› Service changes are communicated to subscriber gateway via policy control nodes › Results in reprovisioning of an active packet processing chain for subscriber
–Will such reprovisioning result in temporary traffic impact? –Can services be added / removed with minimal to nil hit?
Ser v ic e Pr o v isio n in g
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 8
Co n c ept u a l Da t a pl a n e
RX
Subscriber Traffic Detection
Signalling Traffic Detection
To control plane
Ingress Services (ACL/QoS/…)
Route
Egress Services (ACL/QoS/…)
TX ++ ++
ACL ++ Services ACL ++ Services
Route
Subscriber Traffic Detection
Signalling Traffic Injection
From control plane
Q
Hierarchical Queuing
Signalling Traffic Detection
To control plane
Rx / Tx can be via tunnels
SUBSCRI BERS SUBSCRI BERS SUBSCRI BERS
Internet
Access / Aggr Network
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 9
Typic a l Da t a pl a n e Design
RX (+LB) Packet Processing Packet Processing Packet Processing Packet Processing
Reorder (if no LB @ Rx)
Queuing / Scheduling TX From control plane To control plane Drops
Need to prioritize between CP and line traffic
Differing service needs, tunnel encap / decap, mean higher variation in packet arrival rate at reorder / queue stage (Larger reorder queues) H-QoS can become a throughput bottleneck Queues for tunnels would need a reinject back to packet processing Reinject after tunnel decap may be needed if load- balancer @ Rx is used
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 10
› Large number of statistics – both to be collected and pushed periodically
–Session level counters –Counters for each service
› Large number of /32 (v4) [or /128 (v6)] routes
–1 per subscriber: IPv4 (/32 routes) –2 or 3 per subscriber: IPv6 (Covers link-local, IA-PD and IA-NA routes) › Link-local and IA-NA would be /128 routes
› Large number of QoS policers / meters
–Each session (and each service as well) would need to be rate-limited as per SLA › Thus, multiple policers / meters per subscriber
Sc a l e - Memo r y Impl ic a t io n s
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 11
› Session timers include
–Protocol (keep-alive) timers –Idleness detection timers –Session lifetime timers
› Higher scale can result in many timers firing simultaneously › Offloading of protocol timers to dataplane can increase responsiveness
–Lowers burden on control plane –Lowers traffic between dataplane & control plane –BUT: Impacts dataplane throughput traffic!
Sc a l e – Timer Impl ic a t io n s
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 12
› Since we create / remove subscriber sessions continuously, the H-QoS scheduler cone is dynamic and changes constantly › Bandwidth oversubscription is required
–Cannot ensure no oversubscription via configuration
› Memory requirement can be huge
–But can result in packet buffers sitting in queues and starving the pipeline
Qo S Aspec t s
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 13
› Backplane (chassis terminology)
–Path between dataplane and control plane –As mentioned earlier, design of this path is critical for effective functionality › We saw that certain kinds of traffic cannot be lost –Backplane QoS is an essential characteristic of subscriber gateway dataplane
› Fabric (chassis terminology)
–Path between dataplane and service plane –In chassis systems, designed to be high throughput and near lossless –Subscriber gateway requires the essential properties of chassis fabric to be mimic-ed › Loss of fabric traffic also results in poor SLA adherence
NFV IMpl ic a t io n s
Dataplane for Subscriber Management - Requirements and Challenges | 2017-03-30 | Page 14
› Subscriber traffic likely to be carried over L2-over-L3 tunnels
–Optimizing dataplane for tunnels may become the common case!
› Service plane via separate service VMs
–Packet processing at subscriber gateway can become simpler – hence faster –Service chain support in the dataplane › Subscriber gateway can just add service chain tags in its dataplane or interact with SDN controller to setup service chain paths