2014 IEEE 22nd International Symposium of Quality of Service (IWQoS)
Online Resource Scheduling under
Concave Pricing for Cloud Computing
Rui Zhang Kui Wu
City University of Hong Kong University of Victoria
Hong Kong Canada
zhangrui.ray@gmail.com wkui@uvic.ca
AbstractWith the booming growth of cloud computing indus-
try, computational resources are readily and elastically available
to the customers. In order to attract customers with various
demands, most Infrastructure-as-a-service (IaaS) cloud service
providers offer several pricing strategies such as pay as you go,
pay less per unit when you use more (so called volume discount),
and pay even less when you reserve. The diverse pricing schemes
among different IaaS service providers or even in the same
provider form a complex economic landscape that nurtures the
market of cloud brokers. By strategically scheduling multiple
customers resource requests, a cloud broker can fully take
advantage of the discounts offered by cloud service providers.
In this paper, we focus on how a broker may help a group of
customers to fully utilize the volume discount pricing strategy
offered by cloud service providers through cost-efficient online
resource scheduling. We present a randomized online stack-
centric scheduling algorithm (ROSA) and theoretically prove the
lower bound of its competitive ratio. Our simulation shows that
ROSA achieves a competitive ratio close to the theoretical lower
bound under a special case cost function. Trace driven simulation
using Google cluster data demonstrates that ROSA is superior
to the conventional online scheduling algorithms in terms of cost
saving.
I. I NTRODUCTION
In the past few years, we have witnessed the tremendous
development of cloud computing, with more and more cloud
service providers jumping on the cloud bandwagon. Along
with the stable growth of large scale public cloud providers
like Amazon EC2 [2], Windows Azure [7] and Rackspace
[8], small scale cloud providers such as ReadySpace [9] and
GoGrid [4] have vigorously emerged. Despite the hype about
cloud computing, however, the actual adoption rate of cloud
computing is still behind expectation [12], especially outside
the United States. Clearly, to the entire cloud industry, it is
crucial to stimulate end users participation in cloud comput-
ing. From an individual cloud service providers perspective,
it is important to keep its competitiveness among peer cloud
service providers. As analysed in [18], the only way to cloud
computing success is to develop adequate pricing techniques.
In an Infrastructure-as-a-Service (IaaS) cloud, the cloud
provider dynamically segments the physical machines, using
virtualization technologies, to accommodate various virtual
machine (VM) requests from its customers. In principle, the
customers only need to pay for the resource they actually
consumed. Nevertheless, the pay-as-you-use pricing model
978-1-4799-0913-1/14/$31.00 2014 IEEE
978-1-4799-4852-9/14/$31.00 2014 IEEE 51
Jianping Wang
City University of Hong Kong
Hong Kong
jianwang@cityu.edu.hk
is presently only ideological due to the high complexity in
monitoring and auditing resource usage, such as network
bandwidth, virtual CPU time, memory space, and so on.
Consequently, real-world charging schemes in IaaS cloud
have become absurdly complicated [13]. For instance, cloud
providers usually adopt an hourly billing scheme, even if the
customers do not actually utilize the allocated resources in the
whole billing horizon [17]. In the current cloud market, many
cloud providers offer big discount for reserved and long-term
requests [2] [1] [6]. In addition, cloud providers usually give
volume discount to customers with requests of large quantity,
e.g., Amazon EC2 cloud [2] gives 10% discount for customers
spending $25, 000 or above on reserved instances and 20%
discount for customers spending $200, 000 or above. The
diverse pricing schemes and various discount offers among
different IaaS service providers or even within the same
provider form a complex economic landscape way beyond
the control of individual end users. This leaves opportunities
for the cloud brokers to emerge as mediators between the
customers and the providers.
Following the above trend, dedicated cloud brokers are
emerging to help customers make better purchase decisions.
Recent work shows that cloud brokers who mediate the trading
process between the customers and the cloud providers can
significantly reduce the cost for the customers while helping
the cloud providers to reshape or smooth out the burst in the
incoming VM requests [17]. Recent market study expects that
the global cloud services brokerage market will be worth $10.5
billion US dollars by 2018 [3].
A cloud broker can help reduce the cost of customers
through temporal multiplexing and spatial multiplexing of
resources. By temporal multiplexing, the broker takes advan-
tage of providers hourly billing cycles to use a customers
unused resource for executing other customers tasks [16]
[15] [17]. The goal is to maximize resource utilization so
that more customers can be accommodated and in return
each can pay less. By spatial multiplexing, the broker takes
advantage of volume discount by packing multiple customers
resource requests to meet the providers high threshold for bulk
resource purchase, thus, the total cost can be reduced and each
can pay less consequently. While the advantages of temporal
multiplexing have been thoroughly investigated before [17],
the benefit of spatial multiplexing remains less explored.
 2014 IEEE 22nd International Symposium of Quality of Service (IWQoS)
Job1 4 4
Resource
Job2 3 3
2 2
Job3 Job1
1 Job3
0 1 2 3 4 5 6 7 8 9 0 1 2
Time
(a)
Fig. 1: An example shows that a conventional scheduler gives a cost suboptimal schedule.
When being offered with volume discount from a cloud
service provider, end customers may be willing to adjust
the execution speed of their jobs, especially those time-
flexible and interruption-tolerant tasks, so that a higher volume
discount can be enjoyed due to the higher amount of total
requested resource of the jobs from a group of customers.
We use an example to illustrate that conventional scheduling
may not lead to the optimal cost under volume discount. As
shown in Fig. 1 (a), we have three incoming jobs. Job 1
arrives at time 0 with a deadline of 5, a workload (which
is measured by the amount of requested resource) of 6 and
a maximum processing speed1 of 3. Job 2 arrives at time
3 with a deadline of 7, a workload of 3 and a maximum
processing speed of 1. Job 3 arrives at time 6 with a deadline
of 9, a workload of 6 and a maximum processing speed of
2. Suppose that the threshold for volume discount is 2, a
conventional scheduler may schedule a job with its maximum
processing speed starting from the instant when the job is
submitted, as shown in Fig. 1 (b). Under this schedule, no
volume discount can be enjoyed during the execution time
period of job 2. We can observe that postponing the starting
time for processing job 1 to 3 and dividing the execution of
job 2 into two segments give better opportunity in enjoying
volume discount, as shown in Fig. 1 (c).
Though different cloud service providers may offer different
pricing strategies with volume discount, a pricing strategy
with volume discount can be modelled as a concave function
in general, i.e., the total cost of two separate purchases for
resource amounts r1 and r2 , respectively, should be no less
than the cost of a single purchase of the same total resource
amount r1 + r2 .
To discover cost efficient online scheduling algorithm under
a concave cost function, we make the following contributions
in this paper:
Under a generic concave cost function, we investigate
the basic features that a cost optimal scheduling should
possess.
We propose an offline scheduling algorithm under a
special case cost function and theoretically prove its
offline global optimality.
We propose an online request reshaping algorithm, called
1 We assume that the processing speed of a job equals its instantaneous
resource consumption which is charged accordingly.
Resource
Job2
Job2
Job1
1 Job3
Job2
3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
Time Time
(b) (c)
randomized online stack-centric scheduling algorithm
(ROSA), under a generic concave cost function. We the-
oretically prove the lower bound of its competitive ratio
and evaluate its performance with trace-driven simulation
using Google cluster data. Experimental results show that
ROSA achieves a competitive ratio close to the theoretical
lower bound under the special case cost function and is
superior to the conventional online scheduling algorithm
in terms of cost saving.
The rest of the paper is organized as follows. In Section II,
we formulate the resource scheduling problem. In Section III,
we first analyse the properties that an optimal schedule should
possess. We then study a special type of cost functions and
develop an optimal offline algorithm under such a special
type of cost functions. In Section IV, we propose and study
a randomized online algorithm, ROSA, which achieves low
competitive ratio with a linear complexity. Section V presents
our experimental results using Google cluster data. Section VI
concludes the paper.
II. P ROBLEM F ORMULATION
We consider the resource scheduling problem for IaaS
clouds, where the tasks of customers may arrive at random
instants with random workload that should be fulfilled before
a given deadline. Assume that n tasks are submitted during the
time interval [0, T ], indexed by J1 , J2 , . . . , Jn based on their
arrival order. Associated with each task Ji , let tai , tdi , wi denote
its arrival time, deadline, and the workload, respectively.
[tai , tdi ] is denoted as the interval of task i. In practice, we
normally set the upper limit on the resource that could be
allocated to task Ji at any time instant, denoted by ui where
ui wi . We introduce ui to reflect the case that a task
may not be fully executed instantly even if sufficient resource
is allocated due to the limited parallelism among processes
and/or threads of the task. Therefore, a task Ji can be denoted
by a tuple < tai , tdi , wi , ui >.
We assume that the cloud provider has abundant computing
capacity, which is higher than the largest possible total work-
load at any instant t. Let I(t) denote the set of tasks remaining
in the system at time t. The broker can schedule the computing
resource allocated to each task Ji I(t), denoted by ri (t).
Naturally, ri (t) = 0 indicates that no resource is assigned to Ji
at time t. If Ji has already been partially processed, it is paused
at time t. This assumption is theoretically reasonable and
52
 2014 IEEE 22nd International Symposium of Quality of Service (IWQoS)

practically feasible. Theoretically, as long as each task meets that an optimal schedule should have. We then present an
its deadline, the scheduler should have the freedom to assign optimal offline scheduling algorithm under a special concave
the resources in order to reduce the cost. Practically, there cost function.
are many approaches of dynamically assigning the resources
for each job, e.g., CPU time implemented by Xen, memory A. Scheduling with General Concave Cost Functions
ballooning by VMWare, live VM migration by most of the Lemma 1. Assume that f is a positive, non-decreasing, and
Hypervisors. Formally, we require that ri (t) be piecewise con- concave function. Then for any I = [t1 , t2 ], any positive
function x(t), and any positive value u suptI x(t), the
P R(t) as the
stant with finitely many discontinuities. We define
total allocated resource at time t, i.e., R(t) = iI(t) ri (t). following inequality holds:
Associated with the allocated resource R(t) is a cost C(t), Z Z Z
which can be approximated by a non-decreasing function f , f (x(t))dt f (0)dt + f (u)dt, (7)
I I I
i.e.,
C(t) = f (R(t)). (1) where I and I are two subintervals of I such that
R
We assume a proportional cost sharing scheme, i.e., the cost x(t)dt
I = I + I , kI k = I
.
to pay for a task is proportional to the amount of resource the u
task uses. Therefore, the cost for task Ji at time t is calculated Proof. We have
as: Z
ri (t) f (x(t))dt
Ci (t) = f (R(t)). (2)
R(t) ZI
f (x(t)) f (0)
Given a set of n tasks J1 , J2 , . . . , Jn over the interval [0, T ], = x(t) + f (0)dt
x(t)
a feasible schedule consists of resource functions ri (t), i = ZI
f (u) f (0)
1, . . . , n, defined over [0, T ] that satisfy: x(t) + f (0)dt
u
R tdi ZI
f (u) x(t)
Z
ta i
r (t)dt wi , i = 1, . . . , n, (3)
i = x(t)dt + f (0)(1 )dt
0 ri (t) ui , t [tai , tdi ], i = 1, . . . , n, u u
(4) ZI Z I

/ [tai , tdi ], i = 1, . . . , n.
ri (t) = 0, t
The optimal resource scheduling problem is to find a
feasible schedule that minimizes the total cost:
Z T
min C = f (R(t))dt (6)
ri (t) 0 The practical meaning of Lemma 1 is as follows. If the
Significantly different from previous work on speed scal-
ing [10], [11], [20], the cost function is not assumed convex in
our case. Instead, it is approximated as a concave function. The
optimal task scheduling problem turns out to be minimizing a
concave function, which is hard to solve. The lack of convexity
in the cost function invalidates all existing solutions such as
those in [10], [11], [20]. The assumption that the service
providers have unlimited resource to provision differentiates
our work from existing works with explicit resource con-
straints [21].
III. O FFLINE R ESOURCE S CHEDULING
Minimization with a concave cost function usually falls into
the class of NP-hard problem, for example, the concave net-
work flow problem [14]. This partially suggests the hardness of
our scheduling problem. Though we have not formally proved
its NP-harness, we have discovered the properties of opti-
mal scheduling with a general concave cost function. These
properties provide us with valuable insights on making cost-
efficient decisions in offline and online resource scheduling.
Furthermore, these properties have inspired us to find an
optimal offline scheduling algorithm for a special concave
cost function. In this section, we first present the properties
(5) = f (u)dt + f (0)dt.
I I
Note that the first inequality is because f is a positive, non-
decreasing, and concave function and u suptI x(t). The
last equality is based on the definitions of I and I .
cost function is positive, non-decreasing, and concave, the
cost optimal way to schedule a job in an interval where no
other jobs are scheduled is to allocate the maximum possible
resource to the job to reduce the processing time for that job.
With Lemma 1, it is easy to prove the following lemma.
Lemma 2. Assume that the cost function is positive, non-
decreasing, and concave. Assume that there is a feasible
schedule, denoted by S, for a given set of n tasks J1 , . . . , Jn .
For any time interval [tk , tk+1 ], if S allocates to task Ji
the resource ri (t) < ui , t [tk , tk+1 ], then there exists
another feasible schedule, S , that behaves the same as S
for t
/ [tk , tk+1 ] but allocates resource ui to Ji in a sub-
interval of [tk , tk+1 ]. In addition, the total cost of S is no
larger than that of S, i.e., C(S ) C(S).
Proof. Denote I = [tk , tk+1 ]. Define any two time instants
tk , tk+1 such that
R
ri (t)dt
tk tk < tk+1 tk+1 , ktk+1 tk k = I .
ui
We can compress the execution time of task Ji into the
time period I = [tk , tk+1 ] with less total cost, by assigning
Ji the maximum resource ui in I and no resource in the

53
 2014 IEEE 22nd International Symposium of Quality of Service (IWQoS)
time period I = I I . This version of schedule is called a
compressed schedule,Pdenoted by S .
Denote R (t) =
jI(t),j6=i rj (t), i.e., R(t) = R (t) +
ri (t). Define f (x(t)) = f (x(t) + R (t)). Since the cost
function f is positive, non-decreasing, and concave, f is also
positive, non-decreasing, and concave. Based on Lemma 1, we
have
Z Z Z R1
f (ri (t))dt f (0)dt + f (ui )dt, (8)
I I I Fig. 2: Auxiliary diagram for Lemma 3
equivalently,
Z Z Z there does not exist a scheduled job J such that a segment
f (R(t))dt f (R (t))dt + f (R (t) + u)dt. (9)
I I I another time interval which has a higher scheduled resource
In other words, C(S) C(S ). In addition, since the schedule
S is feasible, the schedule S is also feasible because it does
not impact any deadlines.
The practical meaning of Lemma 2 is that whenever
resource is allocated to a job Ji , the optimal solution should
allocate its maximum resource ui .
Lemma 3. Assume that the cost function is positive, non-
decreasing, and concave. Given a job J =< ta , td , w, u >
which can be scheduled into two time intervals I1 and I2 with
allocated resource R1 and R2 , such that w w satisfies the condition of Corollary 1 with a cost of 2 +
u kI1 k, u kI2 k
and R1 R2 , then the cost of allocating resource u to J
for a duration of w
u within interval I1 leads to a smaller total
cost compared to allocating resource u to J for a duration of
w
u within interval I2 .
Proof. According to Lemma 2, we only need to consider the
cases that J is scheduled with the maximum possible resource,
u. Therefore, it is sufficient to prove the following inequality:
f (u + R1 ) y + f (R1 )(kI1 k y) + f (R2 )kI2 k
f (u + R2 ) y + f (R2 )(kI2 k y) + f (R1 )kI1 k, (10)
w
where 0 < y u.
By concavity of f (x) and the condition that R1 R2 , the
following inequality holds:
(f (u + R1 ) f (R1 )) (f (u + R2 ) f (R2 )) 0,
This can be easily explained using Fig. 2. As shown in Fig. 2,
we let a = f (u+R1 )f (R1 ) and b = f (u+R2 )f (R2 ). The
gradient of a positive, non-decreasing and concave function is
a positive and non-increasing function. Therefore, we have
a b. This implies that,
y (f (u + R1 ) f (u + R2 )) y (f (R1 ) f (R2 )) 0
Therefore, (10) holds.
We note that Lemma 2 and Lemma 3 are true for any
concave cost function f (x). Connecting Lemma 3 to the re-
source scheduling problem, we immediately get the following
corollary.
Corollary 1. Assume that the cost function is positive, non-
decreasing, and concave. If a schedule S is optimal, then,
Concave function f(x)
b
a
R1+u R2 R2+u
of the allocated processing time for J can be reassigned to
(excluding the resource allocated to J) than the interval this
segment was allocated to.
Unfortunately, the converse of Corollary 1 is not true. We
have the following counter example. Let J1 =< 0, 1, 1, 1 >,
J2 =< 0, 2, 2, 2 >, J3 =< 1, 3, 2, 2 > and J4 =< 2, 3, 1, 1 >
be four jobsto be scheduled. We assume that the cost function
is f (x) = x. Fig. 3 (a) shows a globally optimal schedule
with the optimal cost 2 3 3.46 which satisfies the property
of Corollary 1. Fig. 3 (b) shows a non-optimal schedule which
also
4 = 4.
B. Scheduling under a Linear Function with a fixed cost
In this section, we propose an optimal offline scheduling
algorithm under a special concave cost function which is a
linear function with a fixed cost.
f (x) = Hx + K (x) (11)
where H, K > 0 and (x) {0, 1}. H stands for the unit
resource price, K stands for the fixed cost and (x) is a binary
function of x ( = 1 when x > 0 and = 0 when x = 0). A
linear function with a fixed cost is commonly used in practice,
which has a straightforward business justification. In many
cases, whenever some service is provided, it will incur a fixed
cost that is independent of the service quantity, for example,
the cost incurred for system monitoring and administration,
or a minimum revenue asked by the service provider for any
single transaction. Let S be an arbitrary schedule, the cost can
be written as,
X
C(S) = KkI k + H wi (12)
i
where I is the union of the
P time intervals that are occupied
by at least one job. As H i wi is constant when all jobs are
scheduled, the problem becomes minimizing the total length
of I . We replace Lemma 3 for general concave functions with
the following two lemmas for the special case cost function.
Definition 1. We first define the term latest possible start time
for job Ji =< tai , tdi , wi , ui > as Li = tdi wui where wi
i
stands for the remaining workload of Ji during the scheduling
process. Let = mini {Li } indicate the smallest latest possible
start time.
54
 2014 IEEE 22nd International Symposium of Quality of Service (IWQoS)

4 4
3 3 J3

Resource

Resource
2 J2 J3 2
1 1 J2
J1 J4 J1 J4
0 1 2 3 0 1 2 3
Time Time
(a) (b)
Fig. 3: (a) A globally optimal schedule (b) A non-optimal schedule which satisfies the condition of Corollary 1

Lemma 4. Let SJ = {Ji |i = 1, . . . , n} be a set of n jobs to be

scheduled and = mini {Li } be the smallest latest possible 10
start time among all jobs. Assume that the cost function is
a linear function with a fixed cost, as defined in (11). There J2

exists an optimal schedule in which no jobs are scheduled
before .
Proof. Suppose we have a set of n jobs, SJ = {Ji |i =
1, . . . , n} and Ji =< tai , tdi , wi , ui >, to be allocated in time
interval [0, T ], where T is the maximum time instance. Let
g be an arbitrary schedule of SJ . We will show that we can
reallocate the workload which is scheduled in [0, ] by g to
[, T ] without incurring the total cost.
Let I0 (g) be the union of intervals with at least one job
scheduled in g. Let I denote the entire time space, i.e. I =
[0, T ].
Step 1: We reallocate the workload of Ji to the time intervals
(I0 (g) {t| < t < tdi }) as much as possible. We call the
resulting schedule g 1 . According to (12), C(g 1 ) C(g). Let
xi be the amount of workload from Ji that is allocated in [0, ]
by g 1 .
Step 2: Let SJ = {Ji |xi > 0, Ji SJ } and Jm =<
tm , tdm , wm , um > be such that tdm = minJi SJ (tdi ). We
a before = 5. Since J1 has the smallest latest possible start
reschedule the workload of Jm which is scheduled in [0, ]
(with an amount of xm ) by g 1 into the intervals ((I I0 (g))
{t| < t < tdm }). This is feasible as the entire workload of Jm
can be scheduled in [, T ]. After the reallocation, all workload
of Jm is scheduled in [, T ].
Step 3: For convenience, we let I(Jm , g) and I(Jm , g 1 )
be the sets of time intervals where g and g 1 allocate Jm ,
respectively. We reallocate the workload of Ji SJ from
(I(Jm , g)[0, ]) to (I(Jm , g 1 )[, T ]). This is also feasible
as tdi tdm . We call the resulting schedule g 2 .
We repeatedly perform Steps 2-3 on g 2 and any subsequent
schedules until all jobs are reallocated to [, T ]. We call the
resulting schedule g k . Let Ik (g) denote the union of intervals
with at least one job scheduled in g k . By Step 2 and 3, it is
obvious that kIk (g)k kI0 (g)k. Hence, C(g k ) C(g). The
same procedure can also be applied to an optimal schedule.
Hence, there exists an optimal schedule such that no jobs are
scheduled before .
We demonstrate the proof of Lemma 4 using an example
shown in Fig. 4. We assume that the cost function is in
the form of (11). A cost-optimal schedule allocates resource
Resource
5
J1 J1
J2 J3
J2 J2 J1
0 1 2 3 4 5 6 7 8 9
Time
Fig. 4: An example which demonstrates the reallocation
procedure in the proof of Lemma 4
(denoted by the shaded boxes in Fig. 4) to three jobs J1 =<
1, 7, 4, 2 >, J2 =< 0, 8, 6, 3 > and J3 =< 7, 9, 12, 6 >. The
latest possible start time for these three jobs are L1 = 5,
L2 = 6, and L3 = 7, respectively. We now apply the
reallocation procedure in the proof of Lemma 4 to obtain an
alternative optimal schedule in which no jobs are allocated
time, according to Step 1 of Lemma 4, we first reallocate 3
units of J2 s workload from time interval [1, 2] to [7, 8]. Then,
according to Step 2, J1 is selected. We reschedule 4 units
of J1 s workload from [1, 2] and [3, 4] to [5, 7]. According to
Step 3, we reallocate 3 units of J2 s workload from [3, 4] to
[5, 6]. Now, all jobs are allocated after time instance 5. We can
observe that the total length of intervals with at least one job
scheduled remains to be 4 after the reallocation. Therefore,
the cost remains the same. According to Lemma 4, we can
optimally schedule a job, Ji , in [, tdi ] if Li = .
Under the linear cost function (11), it is trivial that, if a
job has to be allocated to a particular time interval, allocating
other jobs as much as possible to this time interval will not
incur additional cost. Combining with Lemma 4, we find a
deterministic procedure which leads to an optimal schedule
under the cost function (11). This procedure is presented in
Algorithm 1. As there are a pair nested loops which iterate over
n jobs, Algorithm 1 has an complexity of O(n2 q), where q
is the maximum number of reallocation to be performed on
one job in each step.
Theorem 1. Assume that the cost function is in the form of

55
 2014 IEEE 22nd International Symposium of Quality of Service (IWQoS)

1 Input: SJ , the set of all jobs; tasks in advance and has to make decision with information
2 while SJ 6= do only available so far.
3 Calculate Li for all Ji SJ using Definition 1; We denote by CON the total cost induced by an online
4 = mini Li ; scheduling algorithm and by C the total cost induced by
5 tdmax = 0 the offline optimal scheduling algorithm. We use competitive
6 for Ji SJ do ratio which is defined below to evaluate how close the online
7 if Li = then scheduling is to the offline optimal algorithm.
8 Schedule Ji in [, tdi ]; Definition 2. An online scheduling algorithm is -competitive
9 SJ = SJ {Ji }; if there exists a constant such as
10 tdmax = max(tdmax , tdi );
11 end CON C + (13)
12 end
holds for any input. We call the competitive ratio.
13 for Ji SJ do
14 if < tdi then If the online is not deterministic, then CON is a random
15 Schedule the remaining workload of Ji as variable and in this case, the competitive ratio is defined as:
much as possible in interval [, tdmax ] with
Definition 3. A randomized online scheduling algorithm is
resource ui ;
-competitive if there exists a constant such as
16 Update wi as the amount of the
unscheduled workload of Ji ; E[CON ] C + (14)
17 if wi = 0 then
18 SJ = SJ J i ; holds for any input, where E is the expectation taken over the
19 end random choices made by the online algorithm.
20 end It is clear that the competitive ratio is higher than 1, and the
21 end smaller the competitive ratio, the closer the online algorithm
22 end to the offline optimal solution.
Algorithm 1: Optimal scheduling algorithm with a
linear cost function plus positive fixed cost B. Randomized Online Stack-Centric Scheduling Algorithm
(ROSA)
To present ROSA formally, we first define the concept of
(11). Scheduling by Algorithm 1 is optimal. task density of an time interval.
Proof. The proof directly follows Lemma 4. Definition 4. Assume that there are a set of n tasks Ji =<
tai , tdi , wi , ui >, i = 1, . . . , n. Let S be a schedule of a subset
In this section, we introduced a special case cost function of {Ji }. The task density of a time interval I is defined as
and an algorithm to perform cost optimal job allocation under n
such cost function. In Section V, we use this offline algorithm d(I) =
X
(I Ii (S)) ui , (15)
to evaluate the randomized online algorithm introduced in i=1
Section IV.
where Ii (S) stands for the union of the intervals where Ji is
IV. A R ANDOMIZED O NLINE A LGORITHM scheduled in S (Ii (S) = if Ji is not scheduled), (x) = 1
if x 6= and 0 otherwise.
In this section, we introduce an efficient online scheduling
algorithm with a positive, non-decreasing and concave cost The basic idea of our online algorithm, shown in Algo-
function f (x). The basic idea of our online algorithm is to rithm 2, is to sequentially allocate jobs in the order they
stack the processing times of multiple jobs whenever possible are submitted. Algorithm 2 makes local optimal schedule on
and run the jobs with the maximum possible resource in order allocating each job, using information available so far. The
to reduce the total cost. We prove the lower bound for the local optimality of Algorithm 2 can be verified using Lemma 2
competitive ratio of the proposed online algorithm against the and Lemma 3. When the scheduler allocates the processing
optimal schedule. time for a job, J =< ta , td , w, u >, it always allocates
the job with its maximum possible resource, u. Also, when
A. Online Scheduling and Expected Competitive Ratio scheduling the workload of J, we consider the time intervals
The online resource scheduling problem assumes that, at within the range of [ta , td ] in the order that the interval with
any time instant t, the scheduler only knows the tasks which the highest scheduled workload comes first. After allocating
arrive upon or before t. Other than that, the scheduler does as much workload as possible to the current time interval,
not rely on any knowledge of future information. Online task we go on to the next interval until all of the workload of
scheduling is required in many cases, because the cloud service J is accommodated. In this way, the resulting schedule after
provider or service broker may not have information of all allocating each job always complies with the conditions of

56
 2014 IEEE 22nd International Symposium of Quality of Service (IWQoS)

both Lemma 2 and Lemma 3. Algorithm 2 selects a random At time 0, one task, denoted as task 1, with u1 = n +
start time for (remaining) jobs in time intervals with task 1, w1 = 2(n + 1), and deadline of instant 2n arrives.
intensity of 0 (i.e., no overlapping tasks). Such randomization The first group of n 1 tasks, tasks 2, . . . , n with ui =
offers an opportunity for the current task in consideration 1, wi = 1(i = 2, . . . , n), arrive randomly during the time
to overlap with future unknown incoming tasks. We refer to interval (0, n 1], all having the same deadline of instant
this algorithm as a randomized online stack-centric algorithm n.
(ROSA). The second group of n 1 tasks, tasks n + 1, . . . , 2n 1
with ui = 1, wi = 1(i = n + 1, . . . , 2n) arrive randomly
1 Initialization: an ordered set of time instants during the time interval [n + 1, 2n 1], all having the
I = ; same deadline of instant 2n.
2 while an task Ji arrives do First, we derive the optimal total cost on the random
3 Insert time instants tai and tdi into I; instance. Since the first group of tasks and the second group
4 Find all subintervals [tai , tdi ], each of tasks have no overlap in time, tasks in the first group have
representing a time period in between two no way to be scheduled with any task in the second group.
adjacent time instants in I, and mark them as Obviously, the optimal schedule on this random instance is to
unprocessed; equally split the workload of task 1 into two parts, and then
5 while wi > 0 do schedule the first half and the second half of task 1 with tasks
6 Select the unprocessed subinterval of the first group and tasks of the second group, respectively.
[tai , tdi ], denoted by [t1 , t2 ], that has the The optimal total cost is constant and equals 2f (2n).
highest task density (randomly select one We only need to consider reasonable deterministic online
if there is a tie); algorithms. We call an online algorithm reasonable if it has
7 if the task intensity of [t1 , t2 ] equals 0 the following properties:
then 1) The algorithm makes schedules only with information
8 if t1 > t2 w ui then
i
available so far, and when the schedule of a job is de-
9 Return task Ji is infeasible; termined, the algorithm should not change the schedule
10 end at a later time.
11 Randomly select an instant 2) Whenever resource is allocated to a job Ji , the algorithm
t2 [t1 , t2 wui ];
i
should allocate its maximum resource ui .
12 Allocate ui to Ji in [t2 , t2 + w ui ];
i
3) When there is not enough information to make a better
wi
13 Insert t2 and t2 + ui into I; schedule for a task, the algorithm should not split the
14 else workload of the task.
15 Allocate ui to Ji in The first property is because the algorithm needs to be online;
[t1 , min{t2 , t1 + w ui }];
i
the second property is because of Lemma 2; the third property
wi
16 Insert t1 + ui into I if it is smaller is because ROSA works in the same way (refer to Algorithm 2:
than t2 ; line 7 to line 13).
17 end Any reasonable deterministic online algorithm will have to
18 Update wi = max{0, wi ui kt2 t1 k}; start scheduling task 1 at some point in time before time 2n2
19 Mark this subinterval as processed; (otherwise the deadline cannot be met). Consider an algorithm
20 end that makes a schedule to execute task 1 at time t [0, 2n 2].
21 end There are three possible scenarios:
Algorithm 2: ROSA- Randomized online stack- 1) Case t [0, n 1]: In this case, task 1 has to be
centric algorithm scheduled with jobs in the first group. Clearly, the cost
of any online algorithm on scheduling task 1 and the
Theorem 2. ROSA has the competitive ratio no less than 12 + jobs in the first group is no less than the cost of the
f (n1)+f (n+1) best solution, which is f (n + 1) + f (2n). That is, the
2f (2n) , where n is the total number of jobs and f is
minimum cost for executing these jobs is to stack all jobs
a positive, non-decreasing, and concave cost function.
together, resulting in the cost of f (n+1+n1) = f (2n)
Proof. We prove the theorem based on Yaos minimax prin- for the overlapping period and the cost of f (n + 1) for
ciple [19], i.e., to establish a lower bound on the performance finishing the rest workload of job 1.
of a randomized algorithm, it suffices to find an appropriate Similarly, the cost of any online algorithm on scheduling
distribution of inputs, and to prove that no deterministic algo- jobs in the second group is no less than the minimum
rithm can have the cost smaller than the lower bound against cost, which is f (n 1).
that distribution. As such, we specify a random instance of Therefore, the total cost of any online algorithm is no
the problem and analyse what any algorithm could attain in less than f (n 1) + f (n + 1) + f (2n).
expectation on this random instance. 2) Case t [n 1, n + 1]: Any online algorithm has a cost

57
 2014 IEEE 22nd International Symposium of Quality of Service (IWQoS)
no less than
f (n 1) + f (n + 1) + f (2n). (16)
For t [n1, n+1], Equation (16) represents the best
possible solution that stacks all jobs in the first group
into the time period [n 1, n] and stacks all jobs in the
second group into the time period [n + 1, n + 2].
3) Case t [n + 1, 2n 2]: The analysis of this case is
similar to that in the first case. The lower bound of the
online algorithm is f (n 1) + f (n + 1) + f (2n).
To summarize, the total cost of any reasonable deterministic
online algorithm on this random instance is no less than f (n
1)+f (n+1)+f (2n). Since the total cost of the optimal offline
solution is 2f (2n), Theorem 2 follows.
From Theorem 2, we easily have the following corollaries:
Corollary 2. Assume that the cost function has the form
f (x) = n , where 0 < < 1. ROSA has the competitive
ratio no less than 12 + 21 when n .
Corollary 3. Assume that the cost function is in the form of
(11). ROSA has the competitive ratio no less than 1 when
n .
Based on Corollary 2, the lower bound of the competitive
ratio of ROSA is 1.207 when = 0.5. While the lower bound
of Corollary 3 is meaningless in the sense that the competitive
ratio has to be larger than 1, our experimental evaluation
in Section V shows that ROSA approaches this meaningless
lower bound closely, meaning that empirically ROSA is nearly
optimal.
V. E XPERIMENTAL E VALUATION
In this section, we carry out two sets of experiments. The
first set of experiments aims at evaluating the tightness of
the competitive ratio analysis introduced in Theorem 2. The
second set of experiments aims at evaluating the performance
of ROSA under a generic concave cost function in comparison
to other conventional online scheduling algorithms.
A. Competitive Ratio Analysis with the Special Cost Function
We test the competitive ratio of ROSA with the special cost
function defined by (11). As we have theoretically proved the
optimality of Algorithm 1 in Section III, we use Algorithm 1
to compute the optimal cost for each scenario. Apart from
the parameters, tai , tdi , wi and ui , of job Ji introduced in
Section II, we have some additional control parameters for
the experiment, as listed in Table I. In addition, the deadline
of Ji , tdi , is generated as follows.
wi
td = (1 + ) + ta , (17)
ui
Where is a parameter which controls the flexibility of the
incoming jobs for resource scheduling. We choose the value
of from {1, 5, 10, 20} in this experiment.
The result in Fig. 5 shows that the competitive ratio of
ROSA tends towards 1 as the number of simulated jobs in-
creases. Fig. 5 also shows that the competitive ratio decreases
58
TABLE I: Parameters for Competitive Ratio Analysis
Parameter Description
H, K Parameters for the cost function described in (11), vary
from three different settings : H = 1, K = 0.1 (small
fixed cost), H = 1, K = 0.5 (medium fixed cost) and
H = 1, K = 1 (large fixed cost).
T Number of hours of the simulation period, fixed to 1000
dmax The maximum time duration a job has to be processed (
wi
dmax , i), fix to 5
ui
umax The maximum resource allocated to a job (ui
umax , i), fixed to 0.5
N Number of jobs, vary from 500 (sparse) to 5000 (dense)
when increases. This indicates that, given more flexibility
of job execution time, ROSA can achieve a better competitive
ratio. From Fig. 5, we can also see that the competitive ratio
largely depends on the ratio between the fixed cost K and the
unit price H.
B. Trace Driven Simulation with Generic Concave Cost Func-
tions
In the second set of experiments, we conduct simulations
based on Google cluster data [5] which has been widely used
to perform cloud computing related simulations. The trace data
contains a large number of job records coming from 933 users.
It is recorded for a 29-day duration in May 2011 on a cluster of
11K physical machines. The size of the trace data is over 180
GB. Here, we consider the jobs with explicit computational
tasks which can be processed in a deterministic duration given
the precise processing power of the machine used. To align
with our problem , we preprocessed the data to eliminate the
jobs that are not naturally finished (e.g. web services). This
leaves us with 17 million job records coming from 422 users.
We first analyse the feature of the trace data. Fig. 6 (a)
shows the distribution of the processing length of the jobs. It
shows an approximate exponential distribution which is known
to be heavy-tailed with a small mean. Over 80% of the jobs can
be finished within 30 minutes when working at the requested
speed. Therefore, if we use the same approach introduced in
(17) to generate the deadline of each job, the majority of
the jobs can be handled in less than 10 hours, which is a
reasonable assumption. Fig. 6 (b) shows the distribution of the
job CPU rate requirement. As the trace is scaled to eliminate
any confidential content, we cannot retrieve the actual CPU
rates but a relative scale in the range of [0, 1]. In order to
reduce the computational complexity, we arbitrarily form 42
groups of ten users and perform evaluation on each group.
Fig. 6 (c) shows the number of job requests against time for
five randomly selected groups. We can observe that the curves
exhibit distinct patterns with different levels of workloads.
The diversity of the workloads from different groups of users
makes our evaluation result more robust.
The job records of the trace data contain three attributes
relevant to our simulation, job submission time, job workload,
and job resource requirement2 . For an arbitrary job Ji , the
2 We use the CPU rate requirement as the resource requirement of the jobs
for simplicity.
 2014 IEEE 22nd International Symposium of Quality of Service (IWQoS)

1.4 1.4 1.4

=1 =1 =1
=5 =5 =5
Competitive Ratio

Competitive Ratio

Competitive Ratio
1.3 = 10 1.3 = 10 1.3 = 10
= 20 = 20 = 20

1.2 1.2 1.2

1.1 1.1 1.1

1 1 1
1000 2000 3000 4000 5000 1000 2000 3000 4000 5000 1000 2000 3000 4000 5000
Total Number of VM Requests Total Number of VM Requests Total Number of VM Requests

(a) (b) (c)

Fig. 5: Competitive ratio of ROSA with a linear pricing function plus positive fixed cost. (a) H = 1, K = 0.1 (b) H = 1, K =
0.5 (c) H = 1, K = 1
4
0.25 x 10
0.3 3

Number of Job Submissions

Group 41
Group 33
0.2 Group 37
0.2 2 Group 42
Percentage

Percentage

0.15 Group 0

0.1
0.1 1
0.05

0 0 0
0 1 2 3 0 0.1 0.2 0.3 0 200 400 600
Job Length (Hour) Job CPU Rate Requirement Time (Hour)
(a) (b) (c)
Fig. 6: Analysis of Google Cluster Data. (a) Distribution of job length. (b) Distribution of job requirement. (c) Number of job
submissions
4 4 4
ROSA : FIRSTFIT ROSA : FIRSTFIT ROSA : FIRSTFIT
RANDOM : FIRSTFIT RANDOM : FIRSTFIT RANDOM : FIRSTFIT
3 3 3
Cost Ratio

Cost Ratio

Cost Ratio

2 2 2

1 1 1

0 0 0
0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25

(a) = 0.25 (b) = 0.5 (c) = 0.75

Fig. 7: Performance comparison for the online scheduling algorithms ROSA, Firstfit, and Random with continuous concave
cost function f (x) = x
job submission time, tai , indicates when the job is available
to the scheduler. The CPU rate requirement, ui , indicates the
0, x=0
maximum processing speed and the maximum instantaneous


x + 1, 0x<1

resource consumption of Ji . We obtain the job workload, wi , f (x) = x+29 (18)
15 , 1 x < 16
by multiplying the average CPU rate with the processing time

x+284
100 , x > 16

of Ji . Similar to the previous experiment, we set the total
elapsed time, T , as 1000 and we generate the deadline of We compare the performance of ROSA with two other
Ji using the method introduced in (17) with as the control conventional online scheduling algorithms, a nave random
parameter. The cost of a scheduling decision is computed using scheduler, referred to as Random, and the first-fit scheduler,
(2) with two types of cost functions: (a) f (x) = x , 0 < < referred to as Firstfit. The Random algorithm randomly selects
1, and (b) a piecewise linear function defined by (18), which a time interval of length no shorter than w a
ui between ti and ti
i d

complies with the concavity property. to process Ji . The Firstfit algorithm processes the incoming
jobs at their submission time immediately with a finishing time

59
 2014 IEEE 22nd International Symposium of Quality of Service (IWQoS)
4
ROSA : FIRSTFIT
RANDOM : FIRSTFIT
3 competitive ratio. Simulation results on a Google data trace
Cost Ratio
2 outperforms other conventional scheduling algorithms. The
1 strategies of cloud service providers, brokers, and end cus-
0 discount. It opens a door for many interesting problems along
0 5 10 15 20
Fig. 8: Performance assessment for the online algorithms
ROSA, Firstfit and Random with a piecewise linear cost
function
as early as possible.
We use such three online algorithms to schedule each of
the 42 groups of job traces. We run the simulations with five
different values of , 1, 5, 10, 15 and 20, respectively. We vary
the cost function f (x) = x by setting to three different
values, 0.25, 0.5 and 0.75. As the Firstfit algorithm produces
the same scheduling decision for different values of , we use
it as a baseline algorithm. We report the cost ratio of ROSA
over Firstfit, denoted by ROSA : F irstf it, and the cost ratio
of Random over Firstfit, denoted by Random : F irstf it. The
result is shown in Fig. 7. As shown in Fig. 7, the mean cost
ratios between ROSA and Firstfit over 42 grouped user traces
are smaller than 1 under all parameter settings. Especially
for the cases when = 0.25, the mean ratios are 0.8967,
0.7595, 0.6605, 0.6094, and 0.5597 for = 1, 5, 10, 15, and
20, respectively. Practically, the customers are able to get a
10% cost saving if the interval between the submission time
and deadline of each job is two times of the duration for
processing the job at its maximal speed.
Fig. 8 shows the cost ratio assessment for the three online
algorithms with a piecewise linear cost function defined by
(18). We can observe that the cost ratio between ROSA and
Firstfit is 0.8834, 0.7368, 0.6321, 0.5812 and 0.5311 when
= 1, 5, 10, 15 and 20, respectively. Under such piecewise
linear but approximately concave cost function, the customers
are able to obtain similar benefit as that under the cost function
1 pages 8390, 2010.
of f (x) = x 4 .
VI. C ONCLUSIONS
Cloud is an emerging computing market where cloud
providers, brokers, and users share, mediate, and consume
computing resource. With the evolution of cloud computing,
Pay-as-you-go pricing model has been diversified with the
flavours of volume discount, which stimulate the users adop-
tion to cloud computing. In this paper, we study how a broker
can schedule the jobs of users to leverage the pricing model
with volume discount so that the maximum cost saving can be
achieved for its customers. We have analysed the properties
that an optimal solution should have and we conjecture that
the scheduling problem is NP-hard for general concave pricing
functions. We have developed an offline optimal scheduling
10
algorithm for a special concave pricing function. We have
developed an online scheduling algorithm and derived its
have shown that the proposed online scheduling algorithm
work is the initial step toward studying the behaviours and
tomers when offering or facing a pricing model with volume
25
this line. For example, how a cloud service provider could
determine its pricing schemes (with volume discount) given
the rational customer behaviour of cost saving along with other
competitors to increase its revenue. To enjoy volume discount,
the customers are encouraged to provide loose deadlines which
may degrade user experience. Tight deadlines leave a small
window for cost saving. Further research is required on the
methods of obtaining better trade-off decisions.
R EFERENCES
[1] Alibaba cloud computing. http://www.aliyun.com/.
[2] Amazon elastic compute cloud (amazon ec2). http://aws.amazon.com/
cn/ec2/.
[3] Cloud services brokerage (csb) market. http://www.prweb.com/releases/
cloud-services/brokerage-market/prweb11222307.htm.
[4] Gogrid. www.gogrid.com/.
[5] Google cluster data. https://code.google.com/p/googleclusterdata/.
[6] Huawei cloud service. http://www.hwclouds.com/.
[7] Microsoft windows azure. http://www.windowsazure.com/.
[8] Rackspace. http://www.rackspace.com/.
[9] Readyspace. http://www.readyspace.com/.
[10] L. Andrew, A. Wierman, and A. Tang. Optimal speed scaling under
arbitrary power functions. ACM SIGMETRICS Performance Evaluation
Review, 37(2):3941, 2009.
[11] N. Bansal, H. Chan, and K. Pruhs. Speed scaling with an arbitrary
power function. In Proceedings of the twentieth Annual ACM-SIAM
Symposium on Discrete Algorithms (SODA), pages 693701, 2009.
[12] P. Charalampous. Increasing the adoption rates of cloud comput-
ing. http://www.academia.edu/3400195/Increasing the adoption rates
of cloud computing.
[13] N. Gohring. Confirmed: Cloud infrastructure pricing is
absurd. http://www.itworld.com/cloud-computing/387149/
confirmed-cloud-iaas-pricing-absurd.
[14] G. Guisewite and P. Pardalos. Algorithms for the single-source uncapac-
itated minimum concave-cost network flow problem. Journal of Global
Optimization, 1(3):245265, 1991.
[15] T. Henzinger, A. Singh, V. Singh, T. Wies, and D. Zufferey. Flexprice:
Flexible provisioning of resources in a cloud environment. In Cloud
Computing (CLOUD), 2010 IEEE 3rd International Conference on,
[16] R. Van den Bossche, K. Vanmechelen, and J. Broeckhove. Cost-efficient
scheduling heuristics for deadline constrained workloads on hybrid
clouds. In Cloud Computing Technology and Science (CloudCom), 2011
IEEE Third International Conference on, pages 320327, 2011.
[17] W. Wang, D. Niu, B. Li, and B. Liang. Dynamic cloud resource
reservation via cloud brokerage. In ICDCS, 2013.
[18] C. Weinhardt, A. Anandasivam, B. Blau, N. Borissov, T. Meinl,
W. Michalk, and J. Stosser. Cloud computing a classification, business
models, and research directions. Business & Information Systems
Engineering, 1(5):391399, 2009.
[19] A. Yao. Probabilistic computations: Toward a unified measure of
complexity. In Proceedings of IEEE Symp. Foundations of Computer
Science (FOCS), pages 222227, 1977.
[20] F. Yao, A. Demers, and S. Shenker. A scheduling model for reduced
cpu energy. In Proceedings of IEEE Symp. Foundations of Computer
Science (FOCS), pages 374382, 1995.
[21] G. Zhai, L. Tian, Y. Zhou, and J. Shi. Load diversity based optimal
processing resource allocation for super base stations in centralized radio
access networks. Science China Information Sciences, 57(4):112, 2014.
60