Resource Allocation and Pricing for QoS
Management in Computer Networks
Errin W. Fulp
July 14, 1999
Bastille Day
North Carolina State University
Department of Electrical and Computer Engineering
Raleigh, NC
NC State University
Resource Allocation and Pricing
Resource Allocation & QoS
• Quality of Service (QoS) achieved with the proper
allocation of network resources
– Processor time, buffer space, link bandwidth
• Allocation view point
– Single-user allocation - Efficient allocation to
provide QoS for one user
– Multi-user allocation - Efficient and fair
allocation to all users to provide QoS
E. W. Fulp
1
NC State University
Resource Allocation and Pricing
Thesis Contributions
Single User Allocation
• Dynamic Search Algorithm (DSA+)
Refereed publications - RTSS’96, JCIS’97, ICNP’97
Multi-User Allocation
• Competitive Market Fairness Proofs
• Spot Market Approach
Refereed publications - IWQoS’98, ICNP’98, ICATM’99
Patents pending - US and Japan No. 08/971,127
• Multi-Market Approach
Conference submission - EC’99
E. W. Fulp
2
NC State University
Resource Allocation and Pricing
Multi-User Resource Allocation
Allocation Goals
• Efficient - High utilization
• Fairness - Network and economic oriented
Allocation Classifications
• Centralized or distributed
• Static or dynamic
• Stateless or state-maintaining
• Microeconomic-based
E. W. Fulp
3
NC State University
Resource Allocation and Pricing
Resource Allocation & Microeconomics
Microeconomics — The study of the allocation of scarce
resources among competing ends. Nicholson
• Model
consumer
M
producer
N
producer
consumer
– Users ≡ consumers
– Switches ≡ producers
– Link bandwidth ≡ resource
• Advantages
– Maximize utility
– Optimal distributions
– Many models and methods
E. W. Fulp
4
NC State University
Resource Allocation and Pricing
Previous Microeconomic-Based Methods
Method
Limitations
Constrained max.
Jiang [51]
centralized
ATM VC pricing
Ferguson [32]
CBR only
Eff. bandwidth
Kelly [55]
Smart-market
MacKie-Mason [70]
stat. models
implement.
Want a microeconomic-method that,
• Distributed
• Little a priori info
• Allows demand changes
• Low implementation cost
E. W. Fulp
5
NC State University
Resource Allocation and Pricing
Competitive Market Model
Priced-based model proposed by Lêon Walras in 1874
• Price influences behavior
• At equilibrium the allocation is optimal
Economy consists of multiple competitive markets
• Markets are separate and independent
• Consumers can participate in multiple markets
• Used in the spot and multi-market approach
E. W. Fulp
6
NC State University
Resource Allocation and Pricing
Optimality and Fairness
In an economy consisting of multiple competitive
markets, {a} (allocation array) is
,
• Pareto-optimal if no one can increase their utility
without decreasing the utility of another.
• Weighted Max-Min Fair if, for any other feasible
j
j
âk
ak
aj
allocation {â}, ∃j : â > a =⇒ ∃k : wk < wk ≤ wj
• Equitable if, for any other feasible allocation {â},
∃j : uj (âj ) > uj (aj ) =⇒ ∃k : uk (âk ) < uk (ak ) ≤ uj (aj )
E. W. Fulp
7
NC State University
Resource Allocation and Pricing
Spot Market Approach
user
NB
• Switch - Each link is a
competitive market
• User - Seeks network
resources
switch
• Network Broker (NB) Represents the user
switch
NB
user
E. W. Fulp
Unique Properties
? Demand changes allowed
? Edge calculations
? Immediate availability
8
NC State University
Resource Allocation and Pricing
Switches
• Bandwidth priced (non-storable resource)
• Each output link is an independent dynamic
competitive market
• Price for link i is determined using a
modified tâtonnement process
– Seeks equilibrium price
– Allows demands to change dynamically
– Stateless
• Immediate availability and no reservation overhead
E. W. Fulp
9
NC State University
Resource Allocation and Pricing
7
300
5
250
4
Maximum link bandwidth
95% of link bandwidth
200
Total allocated
Price
3
150
2
100
1
50
0
new
price
Example Allocation and Prices
price
bandwidth (bps)
6
x 10
30
40
50
60
time (seconds)
current
price
R p =Rp
i
n+1
E. W. Fulp
70
0
aggregate
demand
i
n
·
α·s din
i
link
capacity
10
NC State University
Resource Allocation and Pricing
User
• Requires link bandwidth for their application
• Represented in the economy via a Network Broker
QoS Profile
5
User information
• Budget wj
• QoS profile, utility
curve
3.5
good
3
2.5
poor
2
1.5
1
0
E. W. Fulp
excellent
4
QoS score
• Bandwidth desired
4.5
QoS profile
0.2
0.4
0.6
0.8
allocated bandwidth ratio (allocated/desired)
1
11
NC State University
Resource Allocation and Pricing
Network Broker
• Agent for the user
• Located at the network edge
• Performs
– CAC, policing, purchasing decisions
• It knows switch prices and user information
• Determines the amount of bandwidth to purchase
j
w
j j
i
j
j
j
max{u (a )}, p · a ≤ w =⇒ a = min
pi
i∈Rj
E. W. Fulp
12
NC State University
Resource Allocation and Pricing
Spot Market Performance
Steady State (proofs)
• Achieves optimal and fair allocations
– Pareto-optimal
– Weighted max-min fair
– Equitable, must distribute wealth appropriately
Algorithm 5.1
• Price equation always moves towards equilibrium
Network Dynamics (changing demands)
• Use simulation to measure performance
E. W. Fulp
13
NC State University
Resource Allocation and Pricing
Spot Market ABR Rate Control
• RM-cells are used to obtain network feedback
?
price
switch
switch
user NB
NB user
6
RM-cell
Spot Market Approach
• Price is distributed using RM-cells
• A switch inserts the link price in the RM-cell if it is
higher than what is currently stored.
E. W. Fulp
14
NC State University
Resource Allocation and Pricing
ABR Simulation
Determine
• How equitable under dynamic conditions
– Average QoS and % Good or Better (% GoB)
• Compare with perfect max-min and demand-based
WMM
Simulation
• 152 users transmitting MPEG-compressed video
traffic with random start times
• Two types of users MoD and Teleconferencing
E. W. Fulp
15
NC State University
Resource Allocation and Pricing
Average % GoB
Average QoS Scores
average QoS score
% GoB
100
50
0
Tele.
0
MoD
Demand-WMM
Max-Min
Price Method
All
Average QoS Scores for MoD Users
Average QoS Scores for Teleconferencing Users
6
6
Price ABR
Demand-based WMM
Max-min
5.5
5
4.5
4
3.5
3
2.5
Price ABR
Demand-based WMM
Max-min
5.5
average QoS score
average QoS score
2
Demand-WMM
Max-Min
Price Method
All
E. W. Fulp
4
Tele.
MoD
2
0
6
5
4.5
4
3.5
3
2.5
200
400
600
800
time (seconds)
1000
1200
2
0
200
400
600
800
time (seconds)
1000
1200
16
NC State University
Resource Allocation and Pricing
Spot Market Approach
Advantages
• Distributed
• Little a priori info required
• Low implementation cost
Stateless & simple calculations
? Efficient & fair allocations
? Calculation at network edge
? Allows demand changes
? Immediate availability
E. W. Fulp
Disadvantages
• No guarantees
17
NC State University
Resource Allocation and Pricing
Multi-Market Approach
Two markets per output link
• Spot market - Immediate availability
• Reservation market - Guaranteed bandwidth
Multi-Market Bandwidth
Link capacity
α
β
Spot
Reserved
14
12
bandwidth
Unique Properties
? Provides guarantees and
immediate availability
? User can purchase from
various markets
? User can modify choices
as prices change
16
10
8
6
4
2
0
0
2
4
6
time
E. W. Fulp
8
10
18
NC State University
Resource Allocation and Pricing
Reservation Market
• Bandwidth sold as an amount over time (segment)
• Switch will auction β percent as reserved bandwidth
• Users bid for an amount of the next segment
link i, segment (l − 1)
66666666
bid
...
...
6 66666
link i, segment l
...
time
auction
price
auction for
segment l
• Any unused reserved bandwidth sold as spot
bandwidth
E. W. Fulp
19
NC State University
Resource Allocation and Pricing
User and The Multi-Market
• Must define QoS profile, wj , desired bandwidth
• Can purchase spot or reserved bandwidth
Indifference Curves
1
Indifference curve
Describes preferences for
spot and reserved bandwidth
quantity of spot bandwidth
Prefer cheaper
Prefer reserved
0.8
0.6
0.4
0.2
0
0
E. W. Fulp
0.2
0.4
0.6
0.8
quantity of reserved bandwidth
1
20
NC State University
Resource Allocation and Pricing
NB and The Multi-Market
• Can purchase spot and/or reserved bandwidth
Indifference Curve and Budget Line
Reservation bid based on
• Indifference curve
• Spot and reservation
market prices
• Wealth
quantity of spot bandwidth
1
Indifference curve
Budget line
Equilibrium point
Reserved bandwidth bid
0.8
0.6
ureserved
preserved
0.4
pspot
=
uspot
0.2
0
0
0.2
0.4
0.6
0.8
quantity of reserved bandwidth
1
• If not enough reserved bandwidth is purchased, then
spot bandwidth is used for the remaining portion
E. W. Fulp
21
NC State University
Resource Allocation and Pricing
Multi-Market Simulation
Demonstrate advantages of multi-market economy
• Seven link parking-lot network configuration
• Each link 45 Mbps, with segment length of 15
minutes
• Each user transmitted a MPEG-compressed video
and were considered
– Long-term (120 total)
∗ 1/2 prefer reserved, remaining prefer cheaper
– Short-term (40 total)
∗ Prefer cheaper, cause sudden demand shift
E. W. Fulp
22
NC State University
7
x 10
Link 3 Bandwidth Allocation
Link 3 Bandwidth Prices
450
Maximum link capacity
90% of link capacity
Total demand
Total allocated
Reserved allocated
Reservation segment boundary
10
Spot market
Reservation market
Reservation segment boundary
400
350
300
price
bandwidth (bps)
15
Resource Allocation and Pricing
5
250
200
150
100
50
segment 2
0
2000
segment 3
3000
segment 4
segment 5
4000
5000
time (seconds)
segment 6
segment 2
0
2000
6000
segment 3
3000
Average QoS Scores
6
Prefer cheaper
Prefer reserved
average QoS score
5
4
3
2
1
2000
E. W. Fulp
3000
4000
time (seconds)
5000
6000
segment 4
segment 5
4000
5000
time (seconds)
segment 6
6000
23
NC State University
Resource Allocation and Pricing
Multi-Market Approach
Advantages
Disadvantages
? Immediate availability and
guarantees
• Guarantee duration
? Users can purchase various
types
? Users can modify choices
as prices change
E. W. Fulp
24
NC State University
Resource Allocation and Pricing
Future Work
• Price-based routing
• Connection admission control
• Internet pricing
• Price-based security
• Users selling bandwidth
E. W. Fulp
25