Mobility + Streaming

Mobility Aware Server Selection for
Mobile Streaming Multimedia CDN
Muhammad Mukarram Bin Tariq, Ravi Jain, Toshiro Kawahara
{tariq, jain, kawahara}@docomolabs-usa.com
DoCoMo USA Labs.
September 29, 2003
1
Summary
• We present a mobility-aware server selection scheme for content
distribution networks.
• Our target is CDN with high density of servers, each server having a
small coverage area. Mobile users can move out of such service areas in
the duration of streaming media sessions, resulting potentially degraded
QoS.
• Server Handoff can be performed to revive QoS, but it is expensive.
• We use user’s mobility along with traditional criteria such as proximity,
server load etc., and assigns a server such that the probability of user
moving out of coverage area of the assigned server is reduced, while
meeting QoS criteria.
• Simulation results show up to 18 % reduction in number of server
handoffs.
2
Outline
• Summary
• Introduction
– Overview
– Problem Statement
• Mobility Aware Server Selection
– Assumed CDN topology
– Gathering user mobility information and estimating residence time with
servers.
– Server Selection.
• Simulation
– Mobility, Server Selection, Content Distribution.
• Results
• Conclusions
3
• Multimedia has increasing share in
overall traffic
• Fixed broadband has not harnessed
Multimedia, how will mobile
broadband?
– Mobile phones are there all the time.
– Usage scenarios: movies, songs, news,
playing video games etc, while traveling
Market Size (traffic)
Introduction
Multimedia
70-80%
Voice
20-30%
2005
Year
2010
Expected Future mobile
communication market [Yumiba01]
• CDN must meet the challenge of
mobility, wireless and streaming
media trio.
– Our focus is the (Mobility + Streaming).
4
Streaming Media In Mobile Networks
• In previous work
[Tariq02] we showed
that server handoff is
helpful for streaming
content to mobile users.
Server Handoff
R
Server Handoff
R
R
R
Server
Server
R
R
R
R
R
Server
R
Logically Non
Adjacent Subnets,
(hot spots)
• Localizes traffic, reduces
delay, jitter, and load on
the network.
Subnets in a Mobile Network
5
Naïve server handoff scheme has problems
• If the users move too fast, there would be too
many server handoffs, which are expensive for
the network.
– Signaling, Content Placement
• Our mobility-aware server selection assigns
right users to right servers, reducing the need
for handoffs.
– Reduce Number of Handoffs while meeting QoS
criteria.
6
CDN Topology
Tier 3 Server
Coverage Area
Tier 2 Server
Coverage Area
Tier 1 Server
Coverage Area
aka. server-zone
Each has a RR
Servers
More
Coverage
Area
Better
QoS
Access
Network Subnets
• Allows:
– Maximization of traffic localization
– Obtain desired QoS ↔ Number of Handoffs
tradeoff by choosing appropriate server tier.
7
Server Selection Process
Server
Tiers
Move to higher tier if
Move to lower tier if
1) Server Capacity Available
1) Server Capacity Available
2) User is Moving Fast
2) Won’t increase handoffs
3) QoS Diff is maintained
3) QoS Diff is maintained
Server Capacity
Information
RR
We introduce a Lazy Mode
where we do not move users
to lower tiers unless higher
tiers are saturated!!!
Mobility Information
8
ri
Mobility Information
A subnet
Trajectory
of the client
Client maintains its average
subnet residence time over k ri
recent moves
(tn  tn  k )

k
n
Mean residence time of all n
clients in RR’s server-zone
Residence Time
(tn  tn1 )
Client i
tn 1
tn
r 
Mean Server Residence-time
for each tier t
RR uses the information to estimate a future residence
time of client i with tier t
We can make a high granularity estimate using subnet
specific information, at cost of higher overhead.
 ri
i 1
n
Rt
ri
Ei ,t  ( ) Rt
r
ri
Ei ,t  ( ) Rt , s
rs
9
Simulation
Mobility Simulation
• Simulate realistic user movement in a large
geographical area, collect movement events –
we wrote a custom simulator for this.
• Simulate different server selection algorithms
Server Selection
Simulation
Content
Distribution
Simulation
– Baseline, clients always assigned to default tier
– Eager mode with both Low and High Granularity
Mobility Information
– Lazy Mode with both Low and High Granularity
Mobility Information.
• See how we did in terms of delay and jitter
experience by the users.
10
Mobility Simulation
• Custom simulator to
simulate realistic urban
area user movement.
– Cars, Trains, Streets,
Freeways, Public
Transport, Congestion, etc.
– San Francisco Bay area,
3575 sq. miles.
– Over laid with 189 basestations,
59 subnets
11
Simulation Parameters
• CDN topology
– 34 servers arranged in 3 tiers, 21, 8 and 4 in tiers 1, 2, and 3
respectively.
– The 3 tiers at 80ms, 160ms and 240ms respectively, from the edge.
– Server Capacity, variable {50, 75, 100, 200, 300} simultaneous
sessions
• Users
– 2500 users with three QoS class, {1, 2, 3}, users distributed across the
three QoS classes proportionately to the number of servers at
corresponding tier.
– Session Durations, variable {50, 100, 200, 1000, 1500} seconds
– Data rate per user: 64kbps, 20pps
• Selection Criteria
– Desired QoS Separation between adjacent classes: 20 ms.
– Server Overload threshold for Lazy mode. 10% of the maximum
reported load allowance.
12
Simulation Results (1/2)
Eager Mode
Lazy Mode
Percent Reduction in
Number of Handoffs
Reduction in number of handoffs
Low Granularity
20
18
16
14
12
10
8
6
4
2
0
-2
-4
High Granularity
Results for server
capacity = 300
Results for server
capacity = 100
50
100
200
500
1000
1500
Session Duration (seconds)
More results in the paper…
13
Simulation Results (2/2)
Impact on QoS
E2E Delay (seconds)
• Desired Separation is
maintained in all
scenarios
• Eager mode is
achieves
better convergence
and
at lower overall value.
• Higher server
capacity
allows us to do more.
• Accuracy of
estimation
has little impact.
0.26
0.24
0.22
0.2
0.18
0.16
0.14
0.12
0.1
0.08
100
200
300
Server Capacity (number of sessions)
QoS Class 1
QoS Class 2
QoS Class 3
Lazy Mode
Eager Mode
14
Conclusions
• We have presented a mobility aware server selection scheme.
– Up to 18% reduction in number of server handoffs.
– Simple, Largely stateless
– Relies on simple and manageable information – much of which is
already available in the network.
• If you are eager, you better be sure – with eager mode, higher
accuracy is crucial.
• Has applications beyond streaming media.
– Anywhere that you want to make tradeoff with mobility by switching to
wider-area systems.
• Open issues:
– Improving while maintaining simplicity.
– Manageability.
– Bundling with other technologies.
15
Algorithm Details
Task: Assign server to a client i of QoS class q and current/default server tier t
selectedTier := t
Find load allowance of next higher tier Lt+1
If the client is in fastest Lt+1
– true if
C
U
j:r j  ri
j
is less than Lt+1 here Uj number of sessions of a client j
If the delay separation will be maintained
– true if
Dq 1  Dq   q, q 1 . similarly for q,q-1
Assign Server Tier t+1.
End-If
Else-If Eager Mode or (Lazy Mode with Lt+1 too low)
– checking to see if we can move it to lower tier instead
Make sure client won’t increase the number of handoffs i.e., E
i ,t 1  Rt
Assign Server Tier t-1.
End-If
16
IWCW 2003
Conference Report
Muhammad Mukarram Bin Tariq
DoCoMo USA Labs.
October 8, 2003
17