This document summarizes a paper presented at IEEE GLOBECOM 2010. The paper proposes a control-theoretic approach called Redundant Retransmission Ratio Control (RRRC) algorithm to model and mitigate SIP server overload as a feedback control problem. The RRRC algorithm uses a PI controller to regulate the message retransmission rate to maintain a target redundant message ratio and overcome server overload conditions. OPNET simulations of two overload scenarios validate that the RRRC algorithm can effectively prevent overload propagation and cancel overload.
2. Wh t i SIP?
What is SIP?
Session Initiation Protocol
protocol that establishes,
Internet
manages (multimedia)
sessions [RFC 3261]
used for VoIP presence &
VoIP,
video conference
Proxy
Server
Proxy
Server
SIP consists of two basic
elements
l
t
UA (User Agent) and P-Server
(Proxy Server)
About 1000 companies produce
SIP products
Microsoft’s Windows
Messenger (≥4 7) i l d SIP
M
(≥4.7) includes
UA
UA
Simplified SIP Network Configuration
2
4. IMS SIP Server Overload – A
f
h ll
Performance Management Challenge
3GPP has adopted SIP
as the basis of IMS architecture
Problem: Server(s) cannot complete
the processing of requests under
overload conditions
Multiple causes: Insufficient
p
capacity, Component Failures,
Unexpected traffic surges, DOS
attacks [RFC 5390]
Impact: Performance degradation,
drop in throughput, revenue loss,
network collapse
Simplified
Si lifi d IMS C t l L
Control Layer O
Overview
i
4
5. Why Worry About SIP Message Retransmission?
Why Worry About SIP Message Retransmission?
built in
Retransmission built-in to maintain SIP reliability
against message loss
Loss is detected as long delay in acknowledgment
Surge in user demand can cause SIP server
overload and long delay to acknowledge SIP
messages
Long dela s ma trigger more retransmissions and a
delays may
positive feedback exacerbating server overload
5
6. Contributions of This Paper
Using control-theoretic approach to
model an overloaded downstream server and its
upstream server as a feedback control system
Proposing Redundant Retransmission Ratio Control
(RRRC) algorithm (a PI rate control algorithm) to
mitigate the overload
achieve satisfactory target redundant message ratio
by controlling retransmission rate
Performing OPNET simulations under two typical
overload scenarios to
Validate RRRC (implicit SIP overload control) algorithm
6
7. Outline
SIP Retransmission Mechanism Overview
Related Work on SIP Overload Control
Queuing Dynamics of Overloaded Server
Control-Theoretic Design for Overload Control Based
g
on Redundant Retransmission Ratio
Performance Evaluation to Validate RRRC SIP
Overload Control Algorithm
Conclusions
7
9. Retransmission Mechanism
Purpose: Confirmation of successful transmission
between UA and UA via P-servers
Two Types:
Hop by Hop
First retransmission after T1 , subsequent one is 2
times previous interval. Total intervals up to 64 x T1
(maximum 6 retransmissions). Default T1 = 0.5 s.
(
)
f
End-to End
First retransmission after T1 , subsequent one is 2 times
previous interval up to a maximum of T2 . Total
intervals up to 64 x T1 (maximum 11 retransmissions).
Default T2 = 4.0 s.
9
10. Related Work on Overload Control
All the existing overload control solutions adopt
push-back mechanism
cancel the overload effectively
by introducing overhead to advertise upstream servers to
reduce message sending rate
d
di
t
produce overload propagation from sever to server
until end-users
block a large amount of calls unnecessarily
cause revenue loss of service providers
• Our Proposal: Reduce retransmission rate only to
mitigate overload
by maintaining original message rate to
keep the revenue of service providers
10
12. Queuing Dynamics of Overloaded Server
Q
g y
100Trying response
Invite request 1(t)
r1(t)
r2' (t )
Timer fires
Message buffer
q1(t)
Reset timer
Timer expires
qr1(t)
Invite request
Server 2
2(t)
2
1
r2(t)
Server 1
2(t)
q2(t)
1(t)
100Trying response
Timer starts
Ti
Timer buffer
Queuing dynamics of Server 2
Queuing dynamics of Server 1
q 2 (t ) 2 (t ) r2 (t ) 2 (t ) 2 (t )
q1 (t ) 1 (t ) r1 (t ) r2' (t ) 1 (t ) 1 (t )
(1)
(2)
Notation: 1(t) original message rate, r1 (t) message retransmission rate,
2(t) service rate 1 (t) response rate q1 (t) queue size
rate,
rate,
Overload Scenario: Server slowdown at Server 2 due to routine maintenance
Overload Collapse: 2(t) 2(t) > 2(t) q 2 (t ) 0 (see Eq. (1)) q2(t)
ti
trigger r''2(t) r2(t) i
increases q2(t) more quickly
i kl
Overload Propagation: r'2(t) enter Server 1 q 1 ( t ) 0 (see Eq. (2)) q1(t)
12
13. Feedback Overload Control System
g
g
y
Figure 4.Block diagram of feedback SIP overload control system
Root Cause of Overload Collapse
• Retransmission for loss recovery is non-redundant
• Retransmission caused by the overload delay is redundant
Solution
• PI controller C(s) regulates retransmission rate r'2(t) to mitigate the
overload
• achieve desirable target redundant message ratio 0
• redundant message ratio is the ratio between redundant
espo se essage ate
a d total espo se essage ate
response message rate 1r and tota response message rate 1
• P(s) represents the interaction between an overloaded downstream
receiving server and its upstream sending server
13
14. Overload Controller Design
g
Redundant message ratio (t)=1r(t)/1(t) r'2(t)/1(t)
PI controller regulates retransmission rate r'2(t)
r
t
'
r2 (t) KPe(t ) KI 0 e()d
t
KP ( 0 (t )) KI 0 ( 0 ())d
Control plant P(s)=(s)/r'2(s)=[r'2(s)e-s/1]/r'2(s)=e-s/1
PI controller C(s)=KP+KI/s
Open-loop overload control system G(s)=C(s)P(s)=(KP+KI/s)e-s/1
Positive phase margin m of G(s) can guarantee control system stability
PI controller gains can be obtained based on phase margin m
KP
1
2
KI
1 (3 / 4 m )
2
14
15. SIP Overload Control Algorithm (RRRC)
SIP Overload Control Algorithm (RRRC)
Overload Control Algorithm
When each retransmission timer fires or expires
if < 1
Retransmit the message
else
Retransmit the message with a retransmission
probability corresponding to a retransmission
rate r'2 calculated by a PI controller
Adaptive PI control algorithm:
(1) Specify target redundant message ratio
and phase margin m; Set the initial values for ,
and ; Obtain PI controller gains using Eq.
(13).
(2) Measure ,
and
r and calculate
upon response message arrivals.
>1.5
or
<0.5 , self-tune PI
(3) If
controller gains using Eq. (13) and update = ;
g
Otherwise, PI controller remains unchanged.
(4) Calculate the retransmission rate r'2 using
Eq. (7); Go to Step (2).
Varying parameter:
: Round trip delay
: Response message rate
1r : Redundant response message rate
: Redundant message ratio
1r/ 1
KP : Proportional gain of PI controller
KI : Integral gain of PI controller
r'2 : Message retransmission rate
: Monitoring parameter
/ 1
Fixed parameter:
T1 : First-time retransmission timer
: Target redundant message ratio
m : Phase margin
15
16. Scenario to Validate Overload Control Algorithm
Scenario to Validate Overload Control Algorithm
• Poisson distributed message generation rate and service rate
• Two typical overload scenarios
Scenario 1
Initial overload at
Server 1 due to
demand burst
• Mean arrival rate 1=800 messages/sec (emulating a
short surge of user demands) from time t=0s to t=30s
• Mean arrival rate 1=200 messages/sec (emulating
regular user demands) from time t=30s to t=90s
• Mean service capacities of two proxy servers were
C1=C2=1000 messages/sec
Scenario 2
Initial overload at
Server 2 due to
server slowdown
• Mean arrival rate 1=200 messages/sec
• Mean server capacity C1=1000 messages/sec
• Mean server capacity C2=100 messages/sec (emulating
server slowdown) from time t=0s to t=30s, and C2=1000
messages/sec from time t=30s to t=90s
16
17. SIP Network Topology For OPNET Simulation
SIP Network Topology For OPNET Simulation
17
18. Simulation Results of Scenario 1
Simulation Results of Scenario 1
4
x 10
NOLC q1
7
OLC q1
6
o
5
4
3
2
1
0
10
20
30
40
50
60
70
80
90
Time (sec)
Queue size q1 (messages) of Server 1
versus time
OLC qo
8000
6000
10
4000
2000
0
0
20
NOLC qo
0
10
20
30
40
50
60
70
80
OLC Queue size qo (messag
ges)
10000
NOLC Queue size q (messag
e
ges)
Queue siz q1 (messages
ze
s)
8
0
90
Time (sec)
Queue size qo (messages) of an originating
server versus time
• Without overload control algorithm applied, Server 1 became CPU overloaded
overload deteriorated as time evolves, leading to eventual crash of Server 1
• RRRC algorithm made queue size of Server 1 increase slowly
taking only 8s to cancel the overload at Server 1 after new user demand
rate reduced at time t=30s
18
19. Simulation Results of Scenario 2
Simulation Results of Scenario 2
4
x 10
OLC q1
4
1
3
2
10
2
1
0
10
20
30
40
50
60
70
80
0
90
Time (sec)
( )
Queue size q1 (messages) of Server 1
versus time
x 10
1.8
Queue size q2 (messages
s
s)
NOLC q1
0
4
20
OLC Queu size q (messa
ue
ages)
1
NOLC Que size q (mess
eue
sages)
5
NOLC q2
1.6
OLC q2
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
10
20
30
40
50
60
70
80
90
Time (sec)
Queue size q2 (messages) of Server 2
versus time
• Without overload control algorithm applied overload was propagated from Server
applied,
2 to Server 1 when initial overload happened at Server 2
• Persisted overload would crash Server 1 after Server 2 resumed its normal
service
• RRRC algorithm prevent overload propagation to Server 1
taking only 9s to cancel the overload at Server 2
19
20. Conclusions
Applying control-theoretic approach to model SIP
overload problem as a feedback control problem
D
Developing R d d t R t
l i Redundant Retransmission R ti C t l
i i Ratio Control
(RRRC) algorithm (a PI rate control algorithm) to mitigate
the overload by
controlling retransmission rate
achieving desirable target redundant message ratio
Simulation results demonstrate that RRRC (implicit SIP
overload control) algorithm can
prevent the overload propagation
cancel the overload effectively
Our solution does NOT require modification in the SIP
header and time consuming standardization process
time-consuming
can be freely implemented in any SIP servers of different carriers
20
21. Remarks (1)
OPNET simulation code for 3 implicit SIP overload control algorithms
(RRRC, RTDC, and RTQC) published by IEEE Globecom 2010/ICC 2011
available for non-commercial research use upon request
RRRC algorithm (proposed by this IEEE Globecom 2010 paper) has been
quickly adopted by The Central Weather Bureau of Taiwan for their early
earthquake warning system
“A Effi i t Earthquake Early W
“An Efficient E th
k E l Warning M
i Message D li
Delivery Al ith U i an i
Algorithm Using
in
Time Control-Theoretic Approach,” Ubiquitous Intelligence and Computing,
Lecture Notes in Computer Science, 6905, Springer, Berlin, Heidelberg, 2011, pp.
161-173.
http://www.springerlink.com/content/b6252x2k613rv211/?MUD=MP
http://www.ipv6.org.tw/docu/elearning8_2011/1010004798p_3-7.pdf
S
Short review and comments on RRRC (
C (implicit S overload control)
SIP
)
algorithm: "Local SIP Overload Control", Proceedings of WWIC, June 2013.
http://link.springer.com/chapter/10.1007%2F978-3-642-38401-1_16#
http://c3lab.poliba.it/images/2/2a/SipOverload_WWIC13.pdf
http://c3lab poliba it/images/2/2a/SipOverload WWIC13 pdf
21
22. Remarks (2)
Journal version discusses how to apply RTDC algorithm to mitigate SIP
overload for both SIP over UDP and SIP over TCP (with TLS)
“Applying control theoretic approach to mitigate SIP overload,”
y
( )
Telecommunication Systems, 54(4), 2013, pp. 387-404. Available at
http://www.researchgate.net/publication/257667871_Applying_control_theoretic
_approach_to_mitigate_SIP_overload
Survey on SIP overload control algorithms: “A Comparative Study of SIP
y
g
p
y
Overload Control Algorithms,” Network and Traffic Engineering in Emerging
Distributed Computing Applications, IGI Global, 2012, pp. 1-20.
http://www.igi-global.com/chapter/comparative-study-sip-overloadcontrol/67496
t l/67496
http://www.researchgate.net/publication/231609451_A_Comparative_Study_of
_SIP_Overload_Control_Algorithms
Discussion on control system design can be found in the answers to the
ResearchGate question “What are trends in control theory and its
applications in physical systems (from a research point of view)? ”
https://www.researchgate.net/post/What_are_trends_in_control_theory_and_its
https://www researchgate net/post/What are trends in control theory and its
_applications_in_physical_systems_from_a_research_point_of_view2
22