1. Enabling Low Latency and High Reliability for IMS–NFV
Muhammad Taqi Raza and Songwu Lu
Computer Science Department
University of California, Los Angeles

2. Emerging Multimedia Services
There exists a number of interactive multimedia services
HD Voice Service over IP
Rich Communication
Service (chat)
UHD Streaming Service
Smart
City
Smart
Home
Interactive Gaming
eMBMS Service
Connected Car Services

3. IP Multimedia Subsystem – IMS
IMS is an architectural framework for delivering IP multimedia services
IMS Core
Telephony Network
IMS Signaling (SIP)
Proxy Server
Serving Server
Media Gateway
IMS Service (RTP/RTCP)
Data Service
Internet
Base Station
LTE Gateway
IMS Client
Radio Access
4G PS Core

4. Current / Future Trend: NFV of IMS
Network Function Virtualization (NFV) of IMS is a reality
A number of operators are using virtualized IMS
The Network Function (NF) logic is moved from dedicated boxes to commercial off-the-shelf servers
NFV lowers CapEx and OpEx
Dedicated
Proxy NF
Dedicated
Serving NF
Dedicated
Media NF
Commodity
Commodity
Commodity
Proxy NF
Serving NF
Media NF

5. Our Questions How well virtualized implementation of IMS can:
Meet low latency multimedia requirements, and
Achieve carrier grade high availability of five nines (99.999%) ?
System downtime cannot be more than msec / day

6. Empirical Study Reliability in virtualized IMS (vIMS)
Deployed a testbed of OpenIMS over OpenStack
Modified OpenIMS source code to meet 3GPP standard

7. Higher Media-Plane Latencies
Finding
Media latencies exponentially increase after certain call rate
vIMS fails to meet QoS requirements on media traffic

8. Higher Media-Plane Latencies
Analysis
Frequent interactions between different modules of different NFs
Media Controller and Processor modules interact frequently AS
Media Policy
MRFC
Media Control
Policing input
Low level media focus
MRFP
Mixer
Notification
Floor CTR
Processor
. . .
Modules
Network Functions
AS NF defines policy
MRFC NF converts policy into control plane signals
MRFP NF processes media traffic as instructed by control plane signaling

9. Higher Media-Plane Latencies
Analysis
Frequent interactions between different modules of different NFs
Script Document and Interpreter modules form a loop
Modules
Network Functions AS
Script Document
MRFC
Interpreter
Document
Request
XML Interpreter Context
MRFP
Script execution
Implementation Platform
AS dynamically generates XML scripts
Scripts provide media behavior
MRFC fetches scripts from AS over HTTP protocol

10. Traffic Forwarding for Aborted Connection
Finding
Callee devices keep generating uplink media packets for caller device
Speech mute issue for callee devices

11. Traffic Forwarding for Aborted Connection
Analysis
Control-plane termination does not stop data-plane flow
P-CSCF and MRF keeps two different dialogue finite state machines
183 Session in Progress
202 Accepted
ACK
Bye
Timeout
Confirmed
Terminated
Start
Preparative
Early
Established
Moratorium
Mortal
Morgue
200 OK
<Created>
Connection progressing
Connection
connected
Start
Connected
Created
Progressing
Dialogue state machine at P-CSCF
Dialogue state machine at MRF

12. Traffic Forwarding for Aborted Connection
Analysis
Frequent interactions between different modules of different NFs
Module failure causes hanging state machine
Failure of these bridging modules results in control-plane termination.
Data-plane being decoupled from control-plane stays connected
AS-ILCM
AS-OLCM
Application Server
ILCM
OLCM
S-CSCF
ILSM
OLSM
Incoming Leg
Outgoing Leg
P-CSCF
MRF
PGW
Device
Data Plane
Control Plane

13. Our Solution
Pipelining media plane processing and media control commands to reduce latencies
Quickly isolating faulty module by reconfiguring its neighboring modules to improve fault tolerance

14. Design Overview
Predicts future metadata values and request control instructions for all predicted values. (red color in fig.)
Reconfigures each module by adding back-up path with each of one-hop neighboring module. (blue color in fig.)
Serving-CSCF
Incoming Leg Control
Session Controller
Media Resource Controller
Media Control
Resource Controller
Media Resource Processor
Media Processor
Media Update
Application Server
Media Policy
Proxy-CSCF
Registrar
Authentication
PDN-Gateway
(LTE)
Backup Link
Primary Link
Pre-fetch control info
Update
Pre-fetch

15. Pipelining Control and Media Planes
Converting serial operations of processing of media packets at MRFP and fetching control instructions from MFRC into parallel operations.
Predicting future metadata generated by future media packets
When MRFP processes the packets it finds metadata (e.g. jitter, payload, etc.)
Prefetching control instructions for these packets
Media Resource Controller
Media Control
Resource Controller
Media Resource Processor
Media Processor
Media Update
Application Server
Media Policy
Incoming Leg Control
Pre-fetch control info
Update
Pre-fetch

16. Pipelining Control and Media Planes
Optimization using batch prefetching
Parallel operations do not reduce control instructions fetching loop
Use exponential smoothing model to take historical metadata values into account
Generate a batch of predicted metadata and prefetch them
Media Resource Controller
Media Control
Resource Controller
Media Resource Processor
Media Processor
Media Update
Application Server
Media Policy
Incoming Leg Control
Pre-fetch control info
Update
Pre-fetch

17. Fault Isolation and Module Refactoring
Failure detection
Tracking IMS operations through Finite State Machine (FSM)
Failure detection starts in Fast Retry state
Propose 5 retries of the failed SIP message before declaring failure
1. Start timer A
SIP Received
SIP Forwarded
Fast Retry
2. SIP Sent
3. Timer A expires
5. SIP Resent
Failover
4. Start timer B
6. Timer B expires
7. Modules reconfigured
Replay

18. Fault Isolation and Module Refactoring
Failover procedure
The module that detected the failure deactivates the link with failed module
It loads the failed module’s executable through preloaded configurations
It announces module failure to failed module’s neighbors and declares itself as in-service module

19. Implementation
Pipelining control instructions with media plane processing
Break the dependency between Media Processor and Media Update modules and setup two interfaces among them
Failure detection procedure
Detect failure by observing state transitions in an FSM
Fail-over procedure
Keep record of on-going device session (before failure) through hash table
Replay the failed operations

20. Evaluation Reducing Latencies We increase the call rate
60% of the traffic remains below 100 msec (state-of-the-art case)
Frequent interaction between Media Processor and Media Control modules

21. Evaluation Improving fault tolerance
Failure detection time is1 msec for just 7 calls per seconds
Our fail-over procedure is 10X quicker than that of Cloud Platform’s

22. Conclusion
IMS software modules interaction can cause higher latencies and bring failures
Propose re-factoring of different modules to contain latencies and improving fault tolerance
In future, we consider fail-stop failure model as well as data-plane failure cases