Managing Cancellations and No-shows of Reservations with

Managing Cancellations and No-shows of
Reservations with Overbooking to
Increase Resource Revenue
Anthony Sulistio1, Kyong Hoon Kim2, and
Rajkumar Buyya1
1Dept. of
Computer Science and Software Eng.
The University of Melbourne, Australia
2Dept. of
Information Science
Gyeongsang National University, Korea
Proceedings of the 2008 Eighth IEEE International
Symposium on Cluster Computing and the Grid (CCGRID)
Presented by: minglung
 Introduction
 Overview of Revenue Management
 Overbooking Policies
 Capacity Allocation with Overbooking
 Reservation, Penalty, Denied Cost
 Performance Evaluation
 Conclusion and Future Work
Introduction (1/2)
 In Grid, advance reservation (AR) is proposed to ensure the
specified resources are available for application’s
consumption when needed
 Problem
 In reality, users may cancel their reservations before starting
time or by not submitting at all (no-show)
 This causes income loss and lower system utilization
 Overbooking offers a solution
 Allowing the resource provider to accept more reservations
than the capacity
Introduction (2/2)
 This paper consider two challenging issues
 Determining the appropriate number of excess reservations
 Minimizing total compensation cost
 The contribution of this paper is
 Introducing three overbooking models to find an ideal
overbooking limit that exceeds the maximum capacity
 Introducing three strategies for selecting which bookings to
 Investigating the impact of the models and strategies on the net
resource revenue
Overview of Revenue Management
 Why use Revenue Management (RM) ?
 RM has been widely adopted in various industries with limited
and perishable capacity, such as airlines, hotels, and car rentals
 Grid computing is also suitable since computing powers or
nodes are considered to be perishable
 Market segmentation in Grids (Table 1)
 VO means Virtual Organizations
 Price of class i: pi, and p1 > p2 > p3
Overview of Revenue Management
 We assume that class 3 users reserve before class 2 users
before class 1 users (Figure 1)
 yi: protection level
 Protect reservations of class 1, 2
 bi: booking limit
 maxCN: maximum capacity to be reserved
 The data structure used to administering reservations is a
time-slotted data structure
Overbooking Policies(1/10)
 Overbooking can protect a resource provider against
cancellation and no-shows (Figure 2)
 Cancellation: a reservation that is terminated by a user before
the service or starting time ts
 No-show: a reservation that fails to arrive and run on the
resource on ts
Overbooking Policies(2/10)
 ob: the overbooking limit
 Needs to be updated and evaluated frequently as ts approaches
 The ideal overbooking limit is calculated using…
 A probability-based policy
 A risk-based policy
 A service-level policy
 And some assumptions should be made
 Cancellations and no-shows are independent of the number of total
 The probability of a cancellation only depends on the current time
 No-shows are treated as cancellations on ts
 Hence, we can define q(t) as a show rate that reservations show up from the
time remaining until ts
Overbooking Policies(3/10)
 A probability-based policy
 ob is determined statistically based on the probability of shows
 A risk-based policy
 Aims to balance the expected cost of denied service with the
revenue by accepting more bookings
 The cost of denied service refers to the compensation money
given to users who got rejected at the service time
Overbooking Policies(4/10)
 The cost of denied service is denoted as costds and is usually
higher than the resource price p
 A(x): the probability that the demand of users is less than or
equal to x, the number of bookings
 Fx(y): the probability that the number of bookings that will
show up at the time of service is less than or equal to y
 Assuming the current booking limit and the capacity are b and
C, the expected revenue change by increasing the booking limit
from b to b+1 can be derived:
Only if demand > b
causes revenue change
The case that the
number of shows ≦ C
The case that the
number of shows > C
Overbooking Policies(5/10)
 Algorithm 1 calculates the booking limit using risk-based policy
 The booking limit increases while the expected revenue change is positive
Overbooking Policies(6/10)
 In multiple fare classes, the increased revenue from having an
additional booking can not be easily calculated
 A suitable approach is to determine a weighted average price
based on the mean demands μ in each user class
 Assuming that each customer’s showing probabilities are
independent, the show distribution Fx(y) at time t is as follows:
Overbooking Policies(7/10)
 A service-level policy
 Consider the negative impact from the users who got rejected
 Define a service-level threshold and try not to exceed the
 B(x): the number of shows for a given x bookings
 s(x): the service level of x bookings, defined by EQ5
( B( x)  C)  max(0, B( x)  C)
 If we use a binomial distribution for show demands…
E[(B(x) – C)+]
Overbooking Policies(8/10)
 Algorithm 2 calculates the booking limit using service-level
 The booking limit increases while the service-level not exceeds the
Overbooking Policies(9/10)
 Overbooking limit calculation
 Price, p = 100
 Capacity, C = 50
 Denied cost, costds: 125 ~ 175
 Show-rate, q: 0.60 ~ 0.95
 Expected net revenue (ENR) is calculated by EQ7
Overbooking Policies(10/10)
 The service-level policy
produces a lower deniedservice rate compared to
the probability- and riskbased policy
Capacity Allocation with Overbooking
 The capacity allocation problem in RM is to decide the
booking limit for each class user, in order to maximize the
overall expected total revenue
 Expected marginal seat revenue (EMSR) heuristic is used to
determine the booking limits of the three user class
(algorithm 3)
Capacity Allocation with Overbooking
 Fi(x) denotes the probability that the demand of a class i user is less
than or equal to x
 BookingLimit finds the booking limit of a lower class user
Capacity Allocation with Overbooking
 BookingLimit algorithm
Expected revenue increase from bl-1 to bl
Reservation, Penalty, Denied Cost (1/5)
 Initial Cost of A Reservation
 For non-AR jobs
 dur: duration
 numCN: the number of CNs used
 bcost: the base cost of running a job at one time unit
 For AR jobs
 numSlot: the total number of reserved slots
 bcostAR: the cost of running the AR job at one time slot
 τ: a constant factor (τ>1) to differentiate the pricing
 δ: the length of a slot
Reservation, Penalty, Denied Cost
 Penalty Cost
 Penalty rate: αp
 Penalty cost: αp * (the price for each reservation)
 0 ≦ αp ≦ 1
 αp = 0: reservation is fully refundable
 αp = 1: not refundable
 In multiple fare classes, αp1 < αp2 < αp3
Reservation, Penalty, Denied Cost
 Denied cost
 We use EQ9 to determine costds
 τds depends on the agreement or policy set by the resource provider to a
particular user class, τds > τ
 Denied-booking strategies
 Lottery (algorithm 4)
Reservation, Penalty, Denied Cost
 Denied Cost First (DCF) (algorithm 5)
Reservation, Penalty, Denied Cost (5/5)
 Lower Class DCF (algorithm 6)
Performance Evaluation (1/7)
 Simulation by GridSim toolkit
 Simulation setups
Performance Evaluation (2/7)
 Objective: examine the impact of the overbooking policies and
the denied-booking strategies on the net revenue of a resource,
where cancellations and no-shows are allowed
Performance Evaluation (3/7)
 Experimental result
 Negative effect of unanticipated cancellations and no-shows
Performance Evaluation (4/7)
 6-9% increase in net profits using overbooking strategies
 On average, DCF gives the highest total amount of income,
followed by LC-DCF and then Lottery
Performance Evaluation (5/7)
 The SL policy is the most conservative of all
Performance Evaluation (6/7)
 Lottery strategy works best in reducing total denied bookings
 From the compensation cost’s point of view, Lottery is the least
 DCF seems to be a better choice than LC-DCF
Performance Evaluation (7/7)
 LC-DCF has the lowest number of denied Premium users
 To minimize the negative effects from high-paying users who
have been denied access, the combination of SL and LC-DCF
policies is a better solution in the long run
Conclusion and Future Work
 Conclusion
 DCF is best since it has the lowest costds and the highest net
revenue associated with the SL policy
 Future work
 Cancellations and no-shows are dependent of the number of
total booking
 Handling a group cancellation for batch reservations

similar documents