A Novel Application of Option Pricing to Distributed Resources Management

David Allenotor Ruppa Thulasiram Parimala Thulasiraman

Dept. of Computer Science Dept. of Computer Science Dept. of Computer Science
University of Manitoba University of Manitoba University of Manitoba
Winnipeg, R3T 2N2, Canada Winnipeg, R3T 2N2, Canada Winnipeg, R3T 2N2, Canada
[email protected] tulsi @ cs. umanitoba. ca thulasir@ cs. umanitoba. ca

Abstract where computational needs exceed their available capacity

from traditional sources. Currently, there is at least over
In this paper, we address a novel application offinancial 80% resource utilization in production environments [2]
option pricing theory to the management of distributed (which continually increases). The government funding has
computing resources. To achieve the set objective, first, we lead to two major concerns: First, there is little or no
highlight the importance of finance models for the given efforts towards pricing the grid resources. Instead, research
problem and explain how option theory fits well to price the efforts have focused mostly on issues such as security [3],
distributed grid compute resources. Second, we design and middleware and grid infrastructure [1], and Grid resource
develop a pricing model and generate pricing results based management [4] rather than a scheme to price the resources.
on the trace data drawnfrom two real grids: one commercial However, if the grid users are not charged, for quantifying
grid Auvergrid and one experimental platform grid LCG. We the services provided in terms of actual money, it is neces-
evaluate our proposed model using various grid compute sary to price these resources for their use. It will justify the
resources (such as memory, storage, software, and compute government efforts and will serve as a means to report to
cycles) as individual commodities. By carrying out several the government agency on the extent of the funds received.
experiments, a justification of the pricing model is obtained The second concern is the effective management of the
by comparing real behavior to a simulated system based on distributed resources due to surge in the number of users
the spot price for the resources. We further enhanced our that sign up to use the resources daily.
model to achieve a desirable balance between Quality of Secondly, the grid resources have a particular charac-
Service (QoS) and profitability from the perspectives of the teristic feature in their existence; they exist, for example,
users and resource operators respectively. as compute cycles (refereed here to as grid compute com-
modities (gcc) and other non-storable products). Despite the
1. Introduction constraints imposed by their diverse ownership (mentioned
above), the non-storable characteristics of the gccs impose
The grid is defined [1] as a hardware and software infras- further constraints in offering a guaranteed QoS in the grid
tructure that provides dependable, consistent, pervasive, and environments. This is because the overall QoS obtained
inexpensive access to high-end computational capabilities to from the use of a grid resource is a global function of the
its subscribers. The major objective that a grid computing individual local Service Level Agreements (SLA). It is hard
paradigm strive to achieve is the provision of shared compute to satisfy a global SLA.
resources at high resource availability [1] to users with high Thirdly, there is a multitude of technological dynamics
Quality of Service (QoS). However, several factors have to catch up with. For example, new and faster algorithms
militated against this objective. or improved devices which may mean a little change, may
Firstly, a grid is a capital-intensive venture which makes however, cause a huge change in the grid resources user
it too expensive to be owned by an individual. Hence behavior (a drop in the projected resources utilization or a
resource composition characterized ownership from disjoint cancelation of initially booked service). Since most of the
domains and access policies vary as diverse as the owner applications that run on the grid are pre-budgeted, a user
organizations. To circumvent the disadvantage in the cost of may (often times) change patronage attitude with the emer-
ownership and provide access to grid resources, government gence of a better technology. Therefore, the underlying grid
agencies fund most of the grid infrastructures (used for architecture should incorporate the dynamics that satisfies
research purposes only) where resources are provided free- such changes and challenges.
of-charge. As a result of the free-of-charge service offered, To price the grid resources, we address three factors
applications of the computational grid has cut across areas (free-of-charge utilization, non-storable, and sharp changes
of Science, Engineering, Business, Psychology, and others in technology) first by treating the gee as real assets and

97S-1-4244-3750-4/09/S25.00 @2009 IEEE

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 19,2010 at 07:20:32 UTC from IEEE Xplore. Restrictions apply.
then price them using real option thoery. The gee-s in- alternatives, waiting or abandoning an option contract. We
clude CPU cycles, memory, network bandwidths, computing capture these alternatives using fuzzy logic and express the
power, disks, processor times, and various visualization choices as a fuzzy number. A Fuzzy number is expressed
tools, software and specialized instrumentation. Since gee-s as a membership function that exist in the range [0, 1] and
are transient, we assume their availability in three class of we apply it for the development of QoS guarantee for the
users. These include users whose requirement to use the grid benefits of the grid users and operators. We map all possible
resources is "immediate", users whose usage requirement is ftexibilities using membership function.
"non-immediate", and users whose usage requirements could The rest of this paper is organized as follows. Section 2
accommodate possible delays, "delayed". The variations in reviews related work. Section 3 provides option pricing
time of usage accounts for uncertainty and we measure theory. Section 4 describes the pricing architect. Sections 5
this uncertainty as a fuzzy quantity or fuzzy membership describes the simulation environment. Section 6 describes
function, where gee : a ---+ [0,···,1]. We define the the results and discussions of our experiments. Section 7
associated uncertainty in terms of fuzzy domain [0,··· ,1] concludes the paper.
(also called membership function) [5] and apply option (see
next subsection for definition of an option) for choice of 2. Related Work
time of exercising the right to use the grid resources.
In the current study, we use traces from one commercial Besides the focus on security related issues [3] and
grid Auvergrid and one experimental platform grid LCG We middleware and grid infrastructure [1], most of the current
evaluate our proposed model and provide a justification by research in grid resource pricing focus on market based
comparing real grid behavior to simulation results obtained economy approaches (for example [8], [9], [10], [11]) and
using some base spot prices for the commodities. In par- contingent bids in auctions [12]. These research efforts
ticular, we strive to provide service guarantees measured highlight a common goal that involves the use of economic
as QoS and profitability from the perspectives of the users principles to decide a fair share of grid resources that
and grid resources providers respectively. The objective is to involves resources redistribution and scheduling. Currently,
strike and keep a balance between a given service, expected there is no charge for using grid resources. However, a
profitability, and satisfaction for using grid resources. In [6], trend is developing because of large interest in grid for
we presented some preliminary results using our model public computing. Therefore, a sudden explosion of grid
while we relaxed the grid resources' QoS assurance con- use is expected in near future. Iosup et al. [2] obtain traces
straints. We provide motivation for financial option theory of grid resources utilization. Their results show possibility
and associated terminologies as follows. for a future increase in resources use. They suggested that
A financial option is defined (see, for example [7]) as the the grid resources use will reach a peak value soon and
right to buy or to sell an underlying asset that is traded in an this could lead to one of the grid problems. To address
exchange for an agreed-on sum. The right to buy or sell an the problem of sudden explosion of the use of computing
option may expire if the right is not exercised on or before a resource, Amazon has introduced a Simple Storage Service
specific period and the option buyer loses the premium paid (S3) [13]. S3 offers a pay-as-you-go online storage, and
at the beginning of the contract. The exercise price (strike provides an alternative to in-house mass storage. Mayur et
price) mentioned in an option contract is the stated price at al. [14] noted availability and data access performance, and
which the asset can be bought or sold at a future date. A they evaluated the feasibility, performance, and costs of S3-
call option grants the holder the right (but not obligation) to supported storage. They also noted that S3's current security
buy the underlying asset at the specified strike price. On the architecture is inadequate for scientific collaborations by
other hand, a put option grants the holder the right to sell access control, support for delegation and auditing (pricing),
the underlying asset at the specified strike price. American and built-in trusts (which were a set of assumptions).
option can be exercised any time during the life of the option Several other research efforts that explores the possi-
contract; a European option can only be exercised at expiry. bilities of bringing resources pricing into the grid infras-
Since options are instruments derived from some underlying tructure include Sang et al. [15] and Tan and Gurd [16].
assets, they also called derivative securities. They are risky Researchers under this forums, have followed two distinct
securities because the price of their underlying asset at any approaches that is, either to extend the existing standardized
future time may not be predicted with certainty. This means grid middleware or to present some novel work with focus on
the option holder has no assurance the option will bring grid economy referencing resources share and management.
profit before expiry. To hold a real option means to have a However, resources management is not only about schedul-
certain possibility for a given time to either choose for, or ing large and compute-intensive applications (or resources),
against investment decision. A real option provides a choice or some form of advanced reservations [17], [18], [19]. It
from a set of alternatives. For this study, these alternatives also involves the manner of putting compute resources to
include the ftexibilities of exercising, deferring, finding other work for the benefit of the user and owner [20]; that is,

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 19,2010 at 07:20:32 UTC from IEEE Xplore. Restrictions apply.
"profitability". In a similar study Chunlin and Layuan [21] solution of a partial differential equation which captures
presented an optimization-based resources pricing algorithm. the price movements continuously; (ii) application of a
They focus on resources sharing in a computational grid discrete time and state binomial model of the underlying
while they keep increasing the grid providers' effective gain. asset price that captures the price movement discretely [25].
In a recent study [6], we focused on balancing the grid profits The binomial lattice is the most commonly used method.
as seen from the perspectives of the grid resources provider. However, in our simulation, we use the trinomial model [7],
In another related study, Sulistio et aI., [22] evaluate the [26] to solve the real option pricing problem to enhance the
effectiveness of grid revenue management using resource accuracy. This is a discrete time approach to calculate the
reservations as a focus. They show that by charging cus- discounted expectations in a trinomial-tree. See [27] for a
tomers with differentiated prices will increase the effective description of a parallel implementation of binomial lattice
total revenue for the resource in question. They also showed model.
that their scheme guarantees a fare share of the resources
applications with highest computing priorities. The focus of 3.1. Model Assumptions and Theory
the study given in [15], [16], [22] is on resource sharing
and resource scheduling with references to market economy. We made the following assumptions to aid the develop-
Mutz et al. [14] have some interesting schemes that points ment of our model.
to our current research. Mutz et al. critically reexamined Assumption 1: We set some base prices for the gee-so
a batched queue-environment. They considered a simple These assumed prices are the prices that reflect the current
form of batched-queue of jobs ji for i = 1,2,·, n awaiting real sale prices but discounted almost as close to 100%. For
resources; where ji receives service before ji+l. The owners example, if a 1GB of Random Access Memory (RAM) cost
parameters determine the granted resources. Their basis for $70, we set a price of $0.035 a week for 0.5MB memory.
modeling the payment function depends on the users behav- Assumption 2: Since the resources exists in nonstorable
ior which impose some undesirable externality constraints (nonstable) states, we value them as real assets. This as-
on the jobs on queue. With specific reference to the job sumption qualifies them to fit into the general stream of
value Vi (currency based), and the delay in total turnaround investment included in the real option valuation approach.
time d expressed as a tolerance factor. Mutz et al. modeled a This assumption also justifies resources availability. Since
job priority model using efficient design mechanism in [23]. we assumed the resources are nonstable, a high volatility
They based their proposal on a compensation function that ((J) affects the resources availability. This is responsible for
schedules jobs for a time ti. Their objective realizes a a shorter use time of grid resources compared with the life of
compensation function d paid by job at tn-l that wishes option in financial valuation methods. A holder of the option
to access computation at tn. The compensation may be to use the grid resources has an obligation-free chance of
disbursed as incentives (say more gcc) to the waiting jobs. exercising the right. The obligation-free status enables us to
Recently, however, Bhargava and Sundaresan in [12] model apply existing finance option valuation theory to model our
a computing utility and examine the possibility to extend pricing scheme. Consider an asset whose price is initially So
the pay-as-you-go pricing using the auction system. Their and an option on the asset whose current price is f. Suppose
proposal is limited by some assumptions. They assumed the the option has a lifetime of T. It can either move up from So
participants that operate under risk neutral and with a no- to a new level Sou with a payoff value of fu or move down
show (for a bidder that failed to take part in the auction) from So to a new level, Sod and with a payoff value of fd
should get a refund (under the partial commitment mecha- where u > 1 and d < 1. This lead to a one-step binomial.
nism). In a recent study [6], we focused on balancing the grid We define a job on the grid as a service that need one or
profits as seen from the perspectives of the grid resources more of the gee-s from start to finish.
provider. In our current model, we introduce concepts for
such a purpose (i) option with dividend paying underlying 3.2. Price Variant Factor
asset and (ii) a penalty function - the price variant factor
(pvD (Section 3 provides details). The work that we present An important functionality of our model is the price
in this paper is the performance evaluation using real data variance factor (PI). The PI is a fuzzy number, a multiplier
to test our previously proposed model [6] on grid resources and based on the fuzziness (or uncertainty in changes in
pricing. technology) given as 0 ::; PI::; 1. The value depends on
changes in technological developments such as new and
3. Grid Commodity Pricing faster algorithms, faster and cheaper processors, and changes
in access rights and policies. The certainty in predicting the
A large number of extant literature exist to price financial effects caused by these is hard using crisp schemes. As a
options. Some of these approaches include (i) application of result, we capture the resultant changes using fuzzy logic
the Black-Scholes (BS) model [24]. The BS model requires and treat PI as a fuzzy number. For a use time of (t ut ), we

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 19,2010 at 07:20:32 UTC from IEEE Xplore. Restrictions apply.
express a fuzzy value of PI as a fuzzy membership function and (3) yields the transitional probabilities;
that is, M(PI). For example, the grid resources may become
Pu = 0.5 * ((a2~t + v2~t2)/~x2 + (v~t)/~x) (4)
under used if users find better and faster ways to solve
their computing problems. Therefore, to increase the grid Pm = 1 - (( a 2~t + v 2~t2) / ~x2) (5)
resources usage with more capacity for computations under
same technology, we set the value of PI( ut) to 0.1 and with Pd = 0.5 * ((a2~t + v2~t2)/~x2 - (v~t)/~x) (6)
new technology, the PI = 1.0. Our model therefore, adjusts The I-step trinomial process could be repeated several times
the price in the use of grid resources by (PI (ut) ) -1 (for the to form an n-step trinomial tree. For number of time steps
grid operator) while providing quality of service set at the (horizontal level) n = 4, the number of leaves (height)
Service Level Agreement (SLA) of the contract. in such a tree is given by 2n + 1. We index a node by
referencing a pair (i, j) where i points at the level (row
3.3. Discretized Real Option index) and j shows the distance from the top (column index).
Time t is referenced from the level index by i : t = i~t.
We apply the trinomial-tree model [28] to price mainly Node (i,j) is thus connected to node (i + l,j) (upward
American-style and European-style options on a single move), to node (i + 1, j + 1) (steady move), and to node
underlying asset. Options pricing under the Black-Scholes (i + l,j + 2) (downward move). The option price and the
model [24] requires the solution of the partial differential asset price at node (i, j) are given by C[i, j] = Ci,j and
equation and satisfied by the option price. To get option S[i, j] = Si,j respectively. The number of up and down
prices, we build a discrete time and state binomial model of moves required to reach (i, j) from (0,0) estimates the asset
the asset price and then apply discounted expectations [29]. price and is given by
Suppose S is current asset price and r is the riskless
and continuously compounded interest rate, the risk-neutral (7)
Black-Scholes model of an asset price paying the contin- The options at maturity (that is, when T = n~t for Euro-
uous dividend yield of 8 for each year [7] is given by pean style options; T ::; n~t for American style options) are
dS = (r - 8)Sdt - aSdz. For convenience, let x = lnS, determined by the payoff. So for a call option (the intent
this equation can be written as dx = vdt + adz, where to buy an asset at a previously determined strike price), the
v = r - 8 - a 2 /2. Consider a trinomial model of asset price payoff Cn,j = M ax(O, Sn,j - K) and for a put option (the
in a small interval 8t, we set the asset price changes by intent to sell) is given by Cn,j = M ax(O, K - Sn,j). The
8x. Suppose this change remain the same or changes by 8x, value K represents the strike price at maturity T = n~t
with likelihood of an up movement Pu, chance of steady for a European-style option, and the strike price before,
move (without a change) Pm, and chance of a downward or on maturity for an American-style option. To calculate
movement Pd. option prices, we apply the discounted expectations under
The drift (because of known reasons) and volatility (a, the risk neutral assumption. For an American put option (for
because of unknown reasons) parameters of the asset price example), for i < n:
can be obtained in the simplified discrete process using
(PU C i+l,j + Pm C i+l,j+l
rLlt
8x, Pu, Pm, and Pd. In a trinomial lattice the price step Ci,j = max(e-
(8)
(with a choice) is given by 8x = aV38i. By equating +PdCi+l,j+2), K - Si-j).
the mean and variance over the interval 8t and imposing For a European call option (exercised on maturity only),
the unitary sum of the likelihoods, we obtain a relationship
between the parameters of the continuous time and trinomial Ci,j = e-
rLlt
(PU C i+l,j + Pm Ci+l,j+l + PdCi+l,j+2). (9)
(a discretization of the geometric Brownian motion (GBM)),
While option price starts at Co,o, we apply the expression for
that is,
Cn,j with Equations (7), and (8) or (9) to get the option price
E[8x] = Pu(8x) + Pm(O) + Pd( -8x) = v8t (1) at every time step and node of the trinomial-tree. We now
model grid resources based on the transient availability of
where E[8x] is the expectation as mentioned before. From the grid compute cycles, the availability of compute cycles,
Equation (1), and the value of volatility of prices associated with the
E[8x 2] = Pu(8x 2) +Pm(O) +Pd(8x 2) = a 28t+v 28t 2 (2) compute cycles. Given maturity date t, expectation of the
risk-neutral value (E), the future price F(t) of a contract on
where the unitary sum of probabilities represented as grid resources could be expressed as (see for example [7]):

Pu + Pm + Pd = 1 (3) F(t) = E[S(t)] = S(O)eJ~ jL(r)dr (10)

Pu, Pm, and Pd are probabilities of the price going up, down Consider a trinomial model of asset price in a small interval
or remaining same respectively. Solving Equations (1), (2), ~t, the asset price increases by ~x, remain the same

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 19,2010 at 07:20:32 UTC from IEEE Xplore. Restrictions apply.
or decreases by ~x, with probabilities; probability of up fuzziness of the parameters tn, tut, and QoS, we express
movement Pu, probability of steady move (staying at the them in terms of their fuzzy membership functions. That is,
middle) Pm, and probability of a downward movement Pd. M(t n ), M(t ut ), and M(QoS) respectively. If T is a fuzzy set,
To price the multiresources system, we suppose a real the membership function is defined (see for example [5]) as
option depends on some other variables such as the expected T = (t, M(t )), MT (t) E [0, 1].
growth rate geeJ-l and the volatility respectively geeC,. Then To price one of the grid resources, we consider Auvergrid
if we let dg eei / geei = geeJ-ldt + geeadzi, for any number memory because of its relative higher availability compared
of derivatives of gee such as (gee 1 ,gee2' ... ,geen) with to other grids in our study. Therefore, the 89% availability
prices p (PI,P2,··· ,Pn) respectively, we have dInS = of memory in Auvergrid is considered normal. Using this
dpi/Pi = Mi dt + aidz , where the variables gCCi = relationship, in Figure 1, we have a 89% normalized memory
{the set of resources}. Applying the price variant factor availability in Auvergrid. Hence, we use memory availability
PI for pricing options, we have: index set at 0.89 for Auverrid in our simulation. This index
(for example, the memory availability) will provide a user
dIn S = [gee(t) - PI In S]dt + [stochastic term] (11)
with an initial idea to select a particular grid, or certain gcc
where the stochastic term is adz. The value of its member- from certain grid depending on the resource requirement
ship function (high for PI> 0) control the strength of the which the user knows best. To obtain a balance between
PI. So for a multiasset problem, we have: service and cost for using the grid resources, we express
the generated indices i as a membership function Mi (Pn)
dinSi = [geei(t) - PI InSi]dt + ai dzili=I,2,. .. ,n (12)
of the prices p. In our simulation, we calibrate prices as
The value of gee(t) is determined so F(t) = E[S(t)]; that is, PI, P2, . .. ,Pn and the corresponding membership function
the expected value of S is equal to the future price. A user Mi(Pn) using fuzzy values in a range of [0,1]. Equation (13)
may need compute cycles (bandwidth) in 3, 6, and 9 months provides the fuzzy membership function of the range of
from today and therefore decides to pay some amount, $s to prices.
hold a position for the expected increase. We show this using
1 for x = Pc for 100% avialability
a 3-step trinomial. If the spot price for bandwidth is $ST X-PI
for PI ::; x ::; Pc
bit per second (bps) and the projected 3, 6, and 9 months Pc-PI
Pn- X
future prices are $SI, $S2, and $S3 respectively. In this case, Pn-Pc
for Pc ::; x ::; Pn
the two uncertainties are the quantity of bandwidth that will o tt
otherwise that is, if x [PI, Pn]
(13)
be available and the price per bit. However, we can get an
estimate for the stochastic process for bandwidth prices by
substituting some assumed values of PI and a (for example, 4. Pricing Architecture
PI = 10%, a = 20%) in Equation (11) and get the value
of S from Equation (12). Suppose Vl,j represents the option In our model architecture, we normalize base prices for
values at l for l = 0, 1, ... ,n - 1 level and j node for the grid resources using SLA and QoS as constraints for
j = 1, 2, ... ,(2l + 1) (for a trinomial lattice only); that is, individual (local) grids. We also consider economic and
VI,I represents the option value at levelland at Pu. market behaviors for resources conflict in the grid. For a
detailed discussion on the model architecture see [6].
3.4. Fuzzy Logic Framework We setup the base prices for gcc-s based on real and
current market valueadode. For example, a 1GB of RAM
To fuzzify the utility of gcc, we express the quality of costs about $70. For a minimal 2 years for the 100%
the gee availability as a function of the time when gcc is return on investment for a grid operator, we will charge
needed and the time the resources become available for use $47.95 x 10- 6 per day per MB. Similarly, suppose it costs
as gee = f (t ut , t n ), where t n is the life of the contract and $50.00 for 100GB hard disk then we fix a base price of
is given as 0 ::; t n ::; 1, and t ut is the actual utilization time. $34.25 x 10- 8 per day per MB as a charge for the grid
A best scenario is when t n = t ut Le., when the resources storage. Similarly, we set a base price of $34.25 x 10- 6
are available when gcc is needed or t n = 0 (no wait time). If per day per MHz of CPU cycles for a 1.00 GHz dual-
t n = 0, gee use is "now" otherwise, t n = 1 and usage is in core processor. These base prices that we choose are as
the future until the end of the contract period (say 6 months). low as possible because of repeated use of the gcc-s by
Users often request and use gee for computation and expects many consumers (users). This means the resource provider
a best scenario where service provided meets expectations could still get the return on their investments despite the
or when t ut - t n ~ 0 for a high QoS. In this instance, it small charges and small charges for the consumers mean
is hard to guarantee provision of the gee on-demand and they will be satisfied to a large extent for the return on their
satisfy the users' QoS without additional gee to satisfy the premium payments. However, our base prices does not take
conditions named in the SLAs document. To capture the into account the overhead cost due to other infrastructures

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 19,2010 at 07:20:32 UTC from IEEE Xplore. Restrictions apply.
that are not gcc-s. Such infrastructures include operational trinomial tree we use strike price (K = $0.70), resources
expenses that are incurred - power, building infrastructures, price (8 = $0.80), expiration time (T = 0.5 in years),
and air-conditioning expenses, are that the gee-s are fully interest rate (r = 0.06), volatility (0' = 0.2), and the number
utilized. of time steps (Nj = 2N + 1). We extend our study by
varying the volatility 0' in steps of 0.0,0.1,··· ,0.7 and
5. Simulation N = 4,8,16,24. For a 6 month contract, for example,
N = 3 would mean a 2 month step size and N = 12 would
In our study, we consider two real grids; one commercial mean a 2 week step size. We need not go for small step sizes
grid (Auvergrid) and one experimental platform grid (LCG). like those in stock prices in finance market. We take two
We selected one of the grids that has a high normalized approaches. First, we present real grid resources use from
resources usage (that is, supports resource usage with mini- the two real grids chosen based on the traces available from
mal delays as seen by users) resources usage pattern as our them and analyze them. Second, we evaluate our simulation
reference grid. For example, memory usage in LCG was results about the observations made in the trace analysis.
compared to memory usage in Auvergrid. Figure 1(b) shows For the analysis, we study the grid behavior as shown
in the memory usage, consumed CPU time, average pro-
LeG: Used Memory Vs. No. of Jobs AuverGrid: Used Memory Vs. No. of Jobs cessor use, and number of processors to number of jobs
and evaluate their characteristics. We collected the traces
2000 ----___.--+-----11-
N~~~~obs1500+----~
_ _---+--f---+-
for the period of January 01,2007 till December 31,2007
l000 . . . . . ._ r ; H ~ _ - + - t - - - . . - (without any date assigned). In Figure 1(a) we consider the
relationship between number of jobs in LCG and Figure l(b)
considers the memory use for AuverGrid.
(a) (b) The trace characteristics shown in Figure 1(a) and (b) de-
scribes the memory in relation to the number of jobs in LCG
Figure 1. Used Memory Vs. Number of Jobs: (a) LeG and AuverGrid. The AuverGrid supports a higher number
(b) AuverGrid of jobs and maintains a constant 20% memory availability
than LCG. From Figure 1 we can deduce that AuverGrid
that Auvergrid memory supports a larger number of jobs (up has uniform CPU distribution which supports up to 50%
to 2500 X 103 ) than LCG (260 x 10 3 ) in Figure l(a). That is, memory distribution among running jobs (Figure l(b)). In
Auvergrid has a higher memory availability relative to the other words, AuverGrid outperforms the LCG (Figure l(a)).
other grids considered. Therefore, availability of memory in In our experiments, we simulate the grid compute com-
Auvergrid is considered normal (since it supports a number modities (gee) and watch users' request. For a call option,
of jobs 89% more). Using this relationship, in Figure 1, we we simulate the effects of time of use of one of the gee-
use memory availability index at 0.89 for Auvergrid and 0.11 s such as memory (RAM), hard disk (HD), and CPU. In
for LCG. Likewise we have indices generated for each of the other words, we study effect of time of exercise of the
gee-s on various grids. This index will provide a user with an option. We start with memory (one of the gee-s) using the
initial idea to decide on a particular grid, or certain gee from following parameters: 8 = $6.84x 10- 7 , T = 0.5, r = 0.06,
certain grid depending on the resource requirement which N = 4, 8, 16, 24, 0' = 0.2, and N j = 2N + 1; we vary K for
the user knows best. To obtain a balance between service and both advantageous (in-the-money option) disadvantageous
profit for using the grid resources, we express the generated (out-of-the-money) direction for the users. These values
indices i as a membership function J-Li (Pn) of the prices p. reflect the market value of this raw infrastructure, in general.
In our simulation, we calibrate prices as PI, P2, . .. ,Pn and We are not certain about the RAM available in the example
the indices membership function J-Li (Pn) using fuzzy values grids. However, one can easily map parametric values to
in the range [0.0, 1.0]; where a fuzzy value of 0.0 is an index match with the infrastructure available in the grids. This is
for a grid that supports poor resources usage and an index true for other gee-s such as CPU and hard disk discussed
of 1.0 high resource usage. later. We get option values from our experiment and study
the variation in several step sizes. We analyze the effects
6. Results and Discussions of the variations (uncertainty) that exists between the total
period of the option contract and the time of exercise on
For analysis, we collected traces from LCG and Auver- option value.
grid. The resources usage pattern we collected from these Figure 2 (a) shows an in-the-money option value for RAM
grids include number of processors, memory, CPU time, while Figure 2 (b) shows out-of-the-money call. Over the
run time, and wait time. First, we analyze these traces. To number of step sizes, the option value reaches a steady
determine gee-s prices, we run the trinomial lattice using the state. In Auvergrid, the number of jobs running increases
following model parameters: For example, for a one-step steadily throughout the year. In the case of LCG, it is

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 19,2010 at 07:20:32 UTC from IEEE Xplore. Restrictions apply.
 Option Value for RAM Option Value for RAM the PI = 1.0. Fuzzified boundary value of PI is set up
as PI(ut) = [0.1,1.0] to simplify fuzzification. Our model,
= .=
1.5 0.0039 - r - - - - - - - -
1.532 Option

=
1.53 - - - Value 0.0038

OPtion~'~~
(x10-7$) 0.0037 therefore, adjusts the price in the use of grid resources by
Value($~:52
1.522
_
-
_
-
_
-
_
-
0.0036
0.0035
(p I (ut)) -1 (for the grid operator) while providing quality
"5;' .... - - ... - - 0.0034 ~_~--II!I~ _ _L . -
service to the user. For example, applying PI reverses an
16 24
4 8 16

Number of Step
24
Number of Step
unprofitable late exercise of an out-of-the-money option
'--

(a)
L-

(b)
-------'
value to an early exercise of in-the-money option value in a
10% adjustments. Figure 3 (b) shows a corresponding out-
Figure 2. Option Value for RAM: (a) In-the-Money (b) of-the-money option value for CPU.
Out-of-the-Money
7. Conclusions and Future Work
likely that some jobs not in the original batch (queue) were In this paper, we have studied, analyzed, and made
completed before their initially estimated completion time. comparison for grid resources utilizing using traces from two
This means that some jobs may unnecessarily wait in the real grids: LCG and Auvergrid. We focus on the resources
original queue. These original jobs should be compensated usage patterns to simplify our design and to develop a grid
by way of dividend (incentive) paying assets. (Though, we resources pricing model. Our results from the trace analysis
do not show results to this effect, it is easy to incorporate show that some grids could provide resources to the user
the dividend opportunity in the option value computation for at a high value at one time and unable to support the
a specific gee following a projected higher use). We leave same application at other times. In other words, resources
this as a future. availability varies while a measure of their certainty is hard
Similarly, we obtain from our simulation the option values to guarantee.
for both in-the-money and out-of-the-money for CPU using Then we use the same grid resources usage patterns
the limits S = $34.25 and K = $64.48 and $80.48 (all obtained from the traces of real grids to develop a novel
values scaled at (x 10- 6 )) and simulated for a varying time pricing model as a real option problem. Our two important
step of 4,8, 16,24. Figure 3 (a) shows the in-the-money contributions are: (i) option value computation for grid
option value for CPU. The option values for in-the-money resources usage and to select the best point of exercise of
for other gees under our current study include RAM in the option to utilize any of the grid resources. This helps
Figure 2 (a). They show an increasing option value which the user as well as the grid resources provider to optimize
resources for profitability; in other words, we achieve an
Option Value for CPU Option Value for CPU equilibrium condition; (ii) our study also incorporate a price
varying function PI which controls the price of the resources
69~
6.8
Option 6.7
-
_
-
_ Option
Value
'5~
~.:
:~ ==
2'2 ._ ._ _
and ensure the grid users gets the maximum at best prices
(~~:n~ 6.6 - - - -
(x16'$) 2:1 ._:_:__ and the resources provider also make reasonable revenue
6.5 - - - -
6.4 . 2
16
. ..
24
at the current base price settings. At the same time grid
4 8 16 24

Number of Time Step Number of Time Step operators do not unduly over-commit grid resources whether
(a) (b)
the system is in-the-money or out-of-the-money conditions
from the user perspective.
Figure 3. Option Value for CPU: (a) In-the-Money (b) Our future work will focus on the larger problem of
Out-of-the-Money. pricing grid resources for applications that use diverse re-
sources across varied grids simultaneously. This would need
increases with the number of time steps. This behavior shows a more complex optimization of the solution space of the
that at any given time, a consumers' cost for using the grid resources usage as well as finding out the best node
grid resources is the base cost and the extra cost which (time) to exercise the option (utilize the resources).
depends on the time of use of the gee. However for an
equilibrium service-profit, we impose a price modulation - Acknowledgment
price variant factor called PI (see Section 3.2). The value of
the PI depends on changes in the technology or architecture We thank the officials of the LCG and Auvergrid for
of the grid infrastructure. These variations are unknown granting us permission to use their web resources. The
before exercising the options to use grid resources. Therefore authors acknowledge the partial financial support from the
deciding the exact price of gee in real life is uncertain University of Manitoba Research Grant Program (URGP)
and hard to predict. Hence, to increase gee use (ut) with and the second author acknowledges partial finance support
more computing facilities and with same technology, we from the Natural Sciences and Engineering Research Coun-
set the value of PI(ut) to 0.1 and with new technology, cil of Canada (NSERC).

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 19,2010 at 07:20:32 UTC from IEEE Xplore. Restrictions apply.
References [15] W. Kang, H. H. Huang, and A. Grimshaw, "A highly available
job execution service in computational service market," in
[1] I. Foster, C. Kesselman, and S. Tuecke, "The Anatomy of The The 8th IEEE/ACM IntI Conference on Grid Computing.
Grid: Enabling Scalable Virtual Organizations," Int. Journal September 1921, Austin, Texas, USA., September 2007.
of Supercomputer Applications, vol. 15, no. 3, 2001.
[16] Z. Tan and J. R. Gurd, "Market-based grid resource alloca-
[2] A. Iosup, C. Dumitrescuand, D. Epema, H. Li, and L. Wolters, tion using a stable continuous double auction," in The 8th
"How are Real Grids Used? The Analysis of Four Grid Traces IEEE/ACM IntI Conference on Grid Computing. September
and its Implications," in Proc. 7th IEEE/ACM IntI Conference 1921, Austin, Texas, USA., September 2007.
on Grid Computing, Barcelona, Spain, 2006 September 28-
[17] I. Foster, C. Kesselman, C. Lee, R. Lindell, K. Nahrstedt, and
29,2006.
A. Roy, "A distributed resource management architecture that
[3] I. Foster, C. Kesselman, G. Tsudik, and S. Tuecke, "A security supports advance reservations and co-allocation," in Proc. of
Architecture for Computational Grids," in ACM Conference the IntI Workshop on Quality of Service. London, 1999.
on Computer and Communications Security, 1998, pp. 83-92.
[18] W. Smith, I. Foster, and V. Taylor, "Scheduling with advanced
[4] K. M. Sim, "Grid Commerce, Market-Driven G-Negotiation, reservations," in Proc. of the IntI parallel and Distributed
and Grid Resource Management," Systems, Man, and Cyber- Processing Symposium, Cancum Mexico, May 1-05, 2000.
netics, vol. 36, no. 6, 2006.
[19] A. Sulistio, W. Schiffmann, and R. Buyya, "Advanced
[5] G. Bojadziew and M. Bojadziew, Fuzzy Logic for Bus., reservation-based scheduling of task graphs on clusters," in
Finance, and Management Modeling. World Scientific Press, Proc. of the 13th IntI Conference on High Performance
2nd Ed., 1997. Computing (HiPC), Bangalore, India, Dec. 18-21, 2006.

[6] D. Allenotor and R. Thulasiram, "Grid resources pricing: A [20] D. Allenotor and R. K. Thulasiram, "G-FRoM: Grid Re-
novel financial option-based quality of service-profit quasi- sources Pricing A Fuzzy Real Option Model," in Proc. 3rd
static equilibrium model," in The 9th IEEE/ACM IntI Confer- Int' Conference on e-Science and Grid Computing (eScience
ence on Grid Computing (Grid 2008). September 29-0ctober 2007) Bangalore, India, December 10-13, 2007.
01, Tsukuba, Japan., September 2008.
[21] L. Chunlin and L. Layuan, "Pricing and resource allocation
[7] J. Hull, Options, Futures, and Other Derivatives. Prentice in computational grid with utility functions," in International
Hall, Sixth Edition, 2006. Conference on Information Technology: Coding and Comput-
ing (ITCC'05)., vol. II, September 2005.
[8] R. Buyya, D. Abramson, and J. Giddy, "A Case for Economy
Grid Architecture for Service Oriented Grid Computing," in [22] A. Sulistio, W. Schiffmann, and R. Buyya, "Using rev-
IPDPS 2001: Proc. of the 10th Heterogeneous Computing enue management to determine pricing of revervations," in
Workshop &#226 HCW 2001 (Workshop 1), 2001. Proc. 3rd IntI Conference on e-Science and Grid Computing
(eScience 2007) Banga10re, India, December 10-13, 2007,
[9] R. Buyya, D. Abramson, and S. Venugopal, "The Grid 2007.
Economy," in IEEE Journal, 2005.
[23] V. Krishna and M. Perry, "Efficient mechanism Design, 1998
[10] C. S. Yeo and R. Buyya, "Integrated Risk Analysis for a [online]. Available http://ratio.huji.ac.il/dp/dpI33.pdf,'' 2007.
Commercial Computing Service," in Proc. of the 21st IEEE
IntI Parallel and Distributed Processing Symposium (IPDPS [24] F. Black and Scholes, "The Pricing of Options and Corporate
2007, IEEE CS Press, LosAlamitos, CA, USA),2007. Liabilities," Journal of Political Economy, vol. 81, no. 3,
1973.
[11] R. Wolski, J. Plank, J. Brevik, and T. Bryan, "Analyzing
Market-Based Resource Allocation Strategies for the Compu- [25] J. C. Cox, S. Ross, and M. Rubinstein, "Option Pricing: A
tational Grid," Int. Journal of High Performance Computing Simplified Approach," Journal of Financial Economics, no. 7,
Applications, vol. 15, no. 3, pp. 285-281, 2001. 1979.

[12] H. K. Bhargava and S. Sundaresan, "Contingent Bids in Auc- [26] E. Schwartz, "The Valuation of Warrants: Implementing a
tions: Availability, Commitment and Pricing of Computing as New Approach," Journal ofFinancial Economics, no. 4, 2003.
Utility," in HICSS 2004: Proc. of the 37th Annual Hawaii IntI
Conference on System Sciences (HICSS'04) - Track 8, 2004. [27] R. K. Thulasiram, L. Litov, H. Nojumi, C. T. Downing,
and G. R. Gao., "Multithreaded Algorithms for Pricing a
[13] M. Palankar, A. Onibokun, A. Iamnitchi, and M. Ripeanu, Class of Complex Options," in Proc. (CD-ROM) of the IntI
"Amazon s3 for science grids: a viable solution?" in Poster: Parallel and Distributed Processing Symposium (IPDPS), San
4th USENIX Symposium on Networked Systems Design and Francisco, CA, 2001.
Implementation (NSDI'07), April 2007.
[28] P. P. Boyle, "Option Valuing Using a Three Jump Process,"
[14] A. Mutz, R. Wolski, and J. Brevik, "Eliciting Honest Value IntI Options Journal, vol. 3, no. 2, 1986.
Information in a Batch-Queue Environment," in The 8th
IEEE/ACM IntI Conference on Grid Computing. September [29] P. Wilmott, Derivatives: The Theory and Practice ofFinancial
1921, Austin, Texas, USA., September 2007. Engineering. John Wiley and Son Ltd, 1998.

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 19,2010 at 07:20:32 UTC from IEEE Xplore. Restrictions apply.