A CP-Net based Qualitative Composition Approach for an IaaS Provider
Sheik Mohammad Mostakim Fattah, Athman Bouguettaya, Sajib Mistry
AA CP-Net based Qualitative CompositionApproach for an IaaS Provider
Sheik Mohammad Mostakim Fattah, Athman Bouguettaya, and Sajib Mistry
School of Information Technologies, University of Sydney, Sydney, Australia [email protected],{athman.bouguettaya,sajib.mistry}@sydney.edu.au
Abstract.
We propose a novel CP-Net based composition approach toqualitatively select an optimal set of consumers for an IaaS provider.The IaaS provider’s and consumers’ qualitative preferences are capturedusing CP-Nets. We propose a CP-Net composability model using thesemantic congruence property of a qualitative composition. A greedy-based and a heuristic-based consumer selection approaches are proposedthat effectively reduce the search space of candidate consumers in thecomposition. Experimental results prove the feasibility of the proposedcomposition approach.
Keywords:
Cloud Service Composition · IaaS Composition · Quali-tative Preference Composition · CP-Net composability model · MonteCarlo Simulation
Infrastructure-as-a-Service (IaaS) model is a cloud service delivery model wherecomputational resources are usually delivered as Virtual Machines (VMs) tocloud consumers [1,14]. The functional properties of an IaaS service are usuallyCPU, storage, and memory [6]. The Non-functional properties (e.g., availability,throughput, and price) are usually attached with VMs or IaaS services as Qualityof Services (QoS). IaaS services are generally configured based on the functionaland non-functional requirements of consumers. For example, Amazon EC2 IaaSprovider has different types of VMs (e.g., CPU-intensive, Memory-intensive, andNetwork-intensive) that are targeted for different types of consumers (e.g., indi-viduals, small enterprises, and large organizations).The long-term IaaS composition is a topical research issue [12]. The com-position from an IaaS provider’s perspective is defined as the selection of a setof optimal consumer requests [11]. An effective IaaS composition achieves theeconomic expectations, i.e., revenue and profit maximization of the provider.The IaaS composition ensures the optimal utilization of available computingresources for an IaaS provider. The selection of optimal consumer requests is es-sential to achieve the IaaS composition. For example, selecting service requestsfrom a group of small enterprises may be more profitable than the single servicerequest from a large organization due to the scale of economy. a r X i v : . [ c s . A I] F e b S. Fattah et al.
We focus on the qualitative IaaS composition, i.e., selecting the optimal con-sumer requests according to the qualitative preferences of the provider. Quali-tative preference models are effective tools for the selection where there existsuncertainties or incomplete information. The service requirements of future con-sumers are uncertain and probabilistic in nature [2,10,13]. A provider’s prefer-ence may change with the requirements of the consumers. The dynamic businessenvironment may also trigger a change in the provider’s qualitative preferences.For example, the provider may observe a very high demand for Network-intensiveservices in the Christmas or holiday period. The provider may prefer to composeNetwork-intensive services than CPU-intensive services to increase its revenue.IaaS consumers’ requirements can be represented in a natural and intuitivemanner using qualitative approaches [19]. Qualitative models provide the nec-essary tool to select appropriate providers where quantitative models are notapplicable. A consumer requires to explicitly indicate the exact values of thefunctional and non-functional properties of a service in a quantitative model.It may not be possible to find providers that can meet the exact requirementsof the consumers. For example, a consumer may require 10 units of CPU and20 units of memory at 20 dollars/month at the level of 100% availability. Suchrequirements may not be exactly fulfilled by any service provider. In contrast,qualitative preferences can be expressed by comparison. For instance, a contentprovider (IaaS consumer) prefers a “high” network bandwidth to a “low” net-work bandwidth. The content provider may also specify conditional preferences.For example, if the price of network bandwidth is very low, a “high” networkbandwidth is preferred over a “low” network bandwidth. These qualitative pref-erences are used to select suitable providers for consumers.
The IaaS composition problem is modeled in both quantitative and quali-tative approaches [14,12,20,21,8]. The quantitative approaches do not considerthe qualitative preferences of the provider. The composition of requests is trans-formed into an optimization problem in quantitative approaches. The proposedapproaches (e.g., metaheuristic optimization and integer programming) are notapplicable in the qualitative IaaS composition. A heuristic based sequential opti-mization approach is proposed in the qualitative IaaS composition [11]. This ap-proach considers quantitative requirements of the consumers and matches themwith the qualitative preferences of the provider. To the best of our knowledge,existing composition approaches are not applicable where both the provider andconsumer have qualitative preferences.
We propose a Conditional Preference Network (CP-Net) based qualitativecomposition approach for an IaaS provider . We represent qualitative conditionalpreferences using CP-Nets. The CP-Net is a very effective tool to represent andreason with qualitative conditional preferences under ceteris paribus (“every-thing else being equal”) semantics. A CP-Net creates a directed graph whereeach node is an attribute of a service preference. The edge between nodes de-fines the priority among service preferences. The rank of service preferences isgenerated by traversing the graph. We assume that the CP-Nets of the providerand the consumers are provided for simplicity.
Our target is to select the optimal
CP-Net based Qualitative Composition Approach for an IaaS Provider 3 composition of consumers’ CP-Nets that has the highest similarity measure withthe CP-Net of the provider.
We propose a CP-Net based qualitative composition approach for an IaaSprovider. First, we propose a novel CP-Net composability model to composeCP-Nets of multiple consumers using the semantic congruence property of aqualitative composition. Next, we propose a similarity checking mechanism be-tween CP-Nets using the coefficient of correlations . It directs us to apply thebrute-force approach where all possible composition of consumers’ CP-Nets isconsidered to select the optimal composition. The brute-force approach is nota practical solution for composing a large set of consumer’s CP-Nets due to itsexponential runtime. we propose a heuristic-based algorithm and a greedy algo-rithm that reduces search space for compositions. The key contribution of ourresearch is summarized below: – A CP-Net composability model for the qualitative composition of IaaS con-sumers using the semantic congruence property. – A qualitative similarity measure approach using the correlation coefficient. – A heuristic-based and a greedy-based consumer selection approaches to re-duce the search space using semantic similarities between the provider’s andconsumer’s qualitative preferences.
Let us assume an IaaS provider offers VM services based on a fixed set of com-putational resources for a specific period of time. Its resource capacity is up to100 virtual CPU units and 100 memory units. For simplicity, we consider “price”as the only QoS in a VM and we omit “network bandwidth (NB)” functionalityfrom a VM. We also assume that both the consumers’ qualitative requirementsand the providers’ qualitative preferences on CPU, memory, and price are fol-lowing the same semantic levels in Figure 1(a). Three levels of semantics, i.e.,high, moderate, and low for each attribute are specified in the semantic table inFigure 1(a). The IaaS provider builds different types of strategies of service pro-visions. As there exist uncertainties on future consumers’ requirements, it buildsthe strategies using qualitative models. We assume that the provider receivesqualitative requests from three consumers. The target is to select an optimal setof consumer requests that matches with its preferred ways to service provisions.We consider a “CPU-intensive” and a “memory-intensive” service provisionsstrategies for the provider. As the CP-Net provides an effective way to representconditional qualitative preferences, we represent the “CPU-intensive” strategyas CP1 and the “memory-intensive” strategy as CP2 in Figure 1(b). The “CPU-intensive” strategy is to offer CPU intensive services at relatively moderate pricesto attract consumers with CPU intensive requests. Therefore, the CPU is themost important attribute in the dependency graph of CP1 followed by memoryand price. The arc from “CPU” to “memory” implies that the provision of “mem-ory” levels in a VM depends on the selection of “CPU” levels. The preference of
S. Fattah et al.
CPU(C) c1: low (1 to 45 units)c3: High (71 to 100 units)c2: moderate (46 to 70 units)
Price (P) p1: low ($100 to $499)p3: High (more than $100)p2: moderate ($500 to $999)
Memory (M) m1: low (1 to 45 units)m3: High (71 to 100 units)m2: moderate (46 to 70 units) (a)
CMP c3 ≻ c2c3: m2 ≻ m3c2: m3 ≻ m2m3: p3 ≻ p2m2: p2 ≻ p3 MCP m3 ≻ m2 ≻ m1m3: c2 ≻ c3m2: c3 ≻ c2c3: p3 ≻ p2 ≻ p1c2: p2 ≻ p3 ≻ p1CP1 CP2CPU Intensive Memory Intensive (b) Consumer B(Memory Intensive)PMC p1 ≻ p2p1: m2 ≻ m3p2: m3 ≻ m2m3: c3 ≻ c2m2: c2 ≻ c3CP3 MCP m3 ≻ m2 ≻ m1m3: c2 ≻ c3 ≻ c1m2: c3 ≻ c2 ≻ c2c3: p3 ≻ p2 ≻ p1c2: p2 ≻ p3 ≻ p1c2: p1 ≻ p3CP4Consumer A(Price Sensitive) c3 ≻ c2 ≻ c1c3: m2 ≻ m3 ≻ m1c2: m3 ≻ m2 ≻ m1c1: m1 ≻ m2 ≻ m3m3: p3 ≻ p2 ≻ p1m2: p1 ≻ p2 ≻ p3m1: p2 ≻ p3 ≻ p1CP5CMP Consumer C(CPU Intensive) (c) Fig. 1: (a) Semantic Table for Service Attributes (b) CP-Nets of an IaaS Provider(c) CP-Nets of consumersCPU provisions is expressed as c (cid:31) c p (cid:31) p CP-Net based Qualitative Composition Approach for an IaaS Provider 5 similar way. The consumer B has “memory-intensive” qualitative requirements.Therefore, memory is the root attribute in CP4. The choice CPU depends onthe choice of memory and the choice of price depends on the choice of CPU inCP4. The consumer C defines “CPU-intensive” preferences in a similar way.The IaaS provider can select the optimal composition of consumer’s requests,i.e., CP-Nets from three consumers in 2 − N − − A CP-Net is a graphical model to formally represent and reason about qualitativepreference relations. A CP-Net consists of a directed dependency graph andconditional preference tables (CPTs). The dependency graph is defined over aset of functional and non-functional attributes V = { X , . . . ., X n } . A child nodein a dependency graph depends on a set of direct parent nodes P a ( X i ). Thechild node is connected by an arc from P a ( X i ) to X i in the dependency graph.Parent attributes affect the user’s preferences over the value of X i . Each node X i in the dependency graph has P a ( X i ) except for the root nodes.The CPT of each variable X i is defined over the finite, discrete domain D ( X i )and semantic domains S ( X i ). Each value x n in D ( X i ) is mapped into a semanticvalue in S ( X i ) using a semantic mapping table, SemT able ( X n , x n ). Figure 1(a)is a semantic table that maps 71-100 units of CPU as a “high” CPU value. Weonly focus on the attributes that are compatible with additive operations. Hence,we define s i + s j = s k for S ( X i ). For example, c c c X i is specified by (cid:31) for agiven value of the paraent attribute P a ( X i ). A user explicitly defines its prefer-ences over the semantic values of X i for each complete outcome on P a ( X i ). Thepreferences take the form of total or partial order over S ( X i ). For example, theattributes of CP1 are C , M , and P with semantic domains containing x i if X S. Fattah et al.
CMP c3 ≻ c2m3: p3 ≻ p2m2: p2 ≻ p3 c2, m2, p3c2, m2, p2 c3, m3, p2c3, m3, p3c3, m2, p3c3, m2, p2c2, m3, p3c2, m3, p2c3: m2 ≻ m3c2: m3 ≻ m2 Fig. 2: Induced Graphis the name of the feature. The preferences statements are as follows: c (cid:31) c c m (cid:31) m c m ∧ m m p (cid:31) p m p (cid:31) p
2. The statement x (cid:31) x represents the unconditional preference for X = x over X = x .A preference outcome is a combination of values of all attributes of a CP-Net.For example, { c , m , p } and { c , m , p } are two preference outcomes for CP1denoted by o and o . According to the value of attribute P , it can be shownthat o (cid:31) o or o dominates o . The dominance relationship of two preferenceoutcomes is defined as a pre-order between them. Figure 2 depicts the inducedgraph [5] of a CP-Net with all preference outcomes. Let us consider two CP-Nets CP A and CP B . We define the composability oftwo CP-Nets as composable ( CP A , CP B ) = { true, f alse } to find CP C where CP C = compose ( CP A , CP B ). Two CP-Nets are composable if their compositionis semantically congruent. A composition is called semantically congruent if therelative importance order of preference attributes for each consumer is preservedwithout any ambiguity.
The relative importance order is represented as ( a, b )which means a is preferred over b . Let us consider a consumer who has therelative importance order of attributes as ( CP U, memory ). Another consumerhas the relative importance order of attributes as ( memory, price ). If the im-portance order of their composition is (
CP U, memory, price ), the compositionis considered as semantically congruent.
Definition 1.
Semantic Congruence of a Qualitative Composition. A composi-tion is called semantically congruent when the importance order of preferenceattributes for each consumer is preserved without any ambiguity.
The semantic congruence of a composition can be efficiently represented us-ing the directed acyclic graph of CP-Nets. We use semantic congruence propertyto define the composability of two CP-Nets. A CP-Net contains a directed de-pendency graph (DDG) and conditional preference table for each node in thegraph. We define the composability of the DDG and the CPT to define thecomposability of CP-Nets. Two DDG are considered to be composable if theircombined DDG does not contain any cycle. Figure 3(a) shows two DDGs fromtwo different CP-Nets (CP1 and CP2). The CPU is the root of CP1. The memorydepends on the CPU and the price depends on the memory in CP1. The DDG
CP-Net based Qualitative Composition Approach for an IaaS Provider 7
CMP MCP CMPCP1 CP2 Merged (a)
CMP CPM
Different ParentsDifferent attributes (b)
Fig. 3: (a) Dependency Graph Composability (b) CPT Composabilityof CP2 has the order of memory, CPU, and price. To merge them in a singleDDG (i.e., CP12) (Figure 3(a)), we create a new DDG where all the attributes(i.e., CPU, memory, and price) are added from both CP-Nets. The next step isto create edges between the attributes. First, we take a pair of attributes (e.g.,CPU and price) from the new DDG. If the same pair of attributes has an edgein either DDG (i.e., CP1 or CP2), we add an edge to the new DDG. We run thisprocess for each pair of nodes until we cover all edges from both DDGs. If theresulted DDG contains any cycle, then DDGs are not composable. Two CP-Netsare composable if their dependency graphs and CPTs are composable.
Definition 2.
CP-Net Dependency Graph Composability. Two dependency graphsare composable if their combined dependency graph does not contain any cycle.
Definition 3.
CPT Composability. Two CPTs from two different CP-Nets arecomposable if their preference attributes are same and their values depend on thesame set of parent nodes.
Two CPTs are composable if they have two properties. First, both CPTs’ at-tribute nodes should be the same. In Figure 3(b), “CPU” nodes from both DDGscan be valid candidates to be composable. A “CPU” node from one CP-Net anda “memory” node from another CP-Net are not composable. Second, the parentnodes of both nodes should have the same attribute. If two nodes have the sameset of parent nodes, then preference statements of both nodes will depend onthe same set of attributes. Therefore, the preference statements will be com-posable. For example, “CPU” nodes of both CP-Nets are unconditional nodes(i.e., no parent) in Figure 3(b). However, “price” nodes have different parents(i.e., “memory” and “CPU”). The preference statements will have conditionalattributes. Therefore, the CPTs of the "price" nodes are not composable.
We assume that the IaaS provider expresses its qualitative preferences using theCP-Net, CP A . The provider requires to find an optimal set of consumer CP-Netsfrom { CP , CP , CP } that matches with CP A . We apply the composabilitymodel described in the above section and find the composed CP-Net CP B for a S. Fattah et al. set of composable consumers. If CP A completely matches with CP B , we say itis the optimal composition. To compare the composed CP-Net with provider’sCP-Net we need to find a similarity measurement algorithm. The similaritymeasurement between two CP-Nets can be performed in two ways [17,18]. Oneway is to generate the induced graph of two CP-Nets and compute the numberof common edges between two CP-Nets. This similarity can be computed by: Sim ( CP A : CP B ) = |{ e : e ∈ In ( CP A ) ∧ e ∈ In ( CP B ) }||{ e : e ∈ In ( CP A ) ∨ e ∈ In ( CP B ) }| − |{ e : e ∈ In ( CP A ) ∧ e ∈ In ( CP B ) }| (1)where In ( CP A ) and In ( CP B ) denotes the induced graph for CP A and CP B . Theedge between two attributes is denoted by e . The equation 1 computes the ratiobetween the number of common and total edges between two induced graphs.This method is computationally expensive and not applicable in real time [19].Another way is to compare the CPTs between two CP-Nets using the depen-dency graphs. This method is only applicable when two CP-Nets share the samedependency graph [18]. In that case, the similarity between the provider’s andconsumers’ CP-Net is calculated by the following equation: Sim ( CP A : CP B ) = P Xi (cid:16) | CP T A ( X i ) ∩ CP T B ( X i ) |× Q Xj / ∈ Pa ( Xi ) | SemT able ( X i ) | (cid:17)P Xi (cid:16) | CP T A ( X i ) ∪ CP T B ( X i ) |× Q Xj / ∈ Pa ( Xi ) | SemT able ( X j ) | (cid:17) (2)where CP T A and CP T B are the conditional preference table of CP A and CP B . P a ( X i ) denotes the parent attributes of X i and SemT able ( X i ) represents all val-ues that can be assigned into X i . We assume the composed CP-Net of consumersand the provider’s CP-Net have the same dependency graph. The number of composable consumers grows exponentially with the increase inthe number of consumers (2 n ). Finding all possible combinations of consumersis inapplicable as it may require a very large time depending on the numberof consumers [12]. Our target is to reduce the search space for the consumerselection. We propose a greedy-based and a heuristic based consumer selectionalgorithm using the similarity between a provider’s and a consumer’s CP-Nets. We choose the first consumer who has the highest relative similarity with provider’sCP-Net in the greedy selection approach. We iteratively choose the next con-sumers to achieve the maximum similarity with the provider’s CP-Net. Thefollowing steps are performed in the greedy based approach:1. Select a consumer CP-Net that is has the maximum coefficient of correlationwith the provider’s CP-Net.
CP-Net based Qualitative Composition Approach for an IaaS Provider 9
2. Create a new CP-Net based on the difference between the provider’s CP-Netand the selected consumer’s CP-Net.3. Find and select a consumer CP-Net who has the maximum correlation coef-ficient with the new CP-Net.4. Create a new CP-Net based on the difference between the consumer CP-Netand the new CP-Net.5. Perform step 3 and 4 until the difference is zero or minimum.
The greedy approach may not always provide the accurate results as it considersonly consumers with maximum correlation with the provider’s CP-Net. We pro-posed a heuristic approach where we find those consumers who have relativelysimilar CP-Nets with the provider’s CP-Net. Relatively similar CP-Nets aremore likely to form a composition that will match the CP-Nets of the provider.Two CP-Nets are relatively similar if (1) they have the same dependency graph(2) Nodes with similar attributes have similar preferences statements in theirCPTs. CP1 and CP5 have the same dependency graph in Figure 1. The relativesimilarity between two preferences statements is measured based on their rela-tive ordering. For example, consider two preferences statements c (cid:31) c (cid:31) c c (cid:31) c (cid:31) c
9. Although values of the attributes are different, patternsof both statements are same. We consider c (cid:31) c (cid:31) c c (cid:31) c (cid:31) c c ∧ m p (cid:31) p (cid:31) p c ∧ m
10 : p (cid:31) p (cid:31) p
8. The condition of thefirst statement c ∧ m c ∧ m
10. A similar statement can be found in the CPT of the same attribute ofthe provider’s CP-Net for each statement of the CPT of an attribute from theconsumer’s CP-Net. We perform the following steps to find the relative similaritybetween a consumer’s and the provider’s CP-Net:1. Compare the dependency graph of the provider’s and the consumer’s CP-Net. If the dependency graphs are not the same, the CP-Nets are not similar.2. If the dependency graphs are same, find an unconditional node from provider’sCP-Net for each unconditional node of the consumer CP-Net.3. Compute similarity between the unconditional nodes selected in step 2.4. Store the similarity measurement in a global variable.5. Find similar conditional nodes for each attribute from provider’s and con-sumer’s CP-Nets.6. For each preference statement in a CPT of an attribute of the consumer’sCP-Net, find a similar preference statement in the CPT of the same attributeof provider’s CP-Net. The attributes and the conditions of both statementsshould be also relatively similar.7. The similarity between the conditional nodes is computed. Update the totalsimilarity measurement.
Algorithm 1:
Similarity Checking between two CP-Nets
Input : CP A , CP B , SemanticT able
Output:
Similarity
Sim ( CP A , CP B ) Integer commonEdges ← Integer allEdges ← CP T A ← find all CPT in CP A CP T B ← find all CPT in CP B foreach X i attribute in CP A do visitedP ref erences ← ∅ foreach P A in CP T A [ X i ] do boolean f lag ← f alse foreach P B in CP T B [ X i ] do if P A has similar pattern P B then visitedP ref erences ← P B f lag ← true commonEdges ← commonEdges + Q X j / ∈ P ( X i ) | SemT able ( X j ) | allEdges ← allEdges + Q X j / ∈ P ( X i ) | SemT able ( X j ) | end if ! f lag then allEdges ← allEdges + Q X j / ∈ P ( X i ) | SemT able ( X j ) | end end end foreach P B in CP T B [ X i ] do if P B / ∈ visitedP ref erences then allEdges ← allEdges + Q X j / ∈ P ( X i ) | SemT able ( X j ) | end end end return Sim ( CP A , CP B ) ← commonEdges/allEdges We propose Algorithm 1 to find relatively similar consumers based on theprovider’s CP-Nets. Algorithm 1 calculates the coefficient of correlation betweentwo CP-Nets using equation 2. The algorithm takes two CP-Nets with same de-pendency graphs. Two variables are defined to calculate the number of commonedges and all edges between the CP-Nets ( commonEdges and allEdges ). Ac-cording to our assumption, both CP-Nets have the same number of attributes.For each attribute, we perform a check if the conditional preferences from bothCP-Nets have a similar pattern. When a preference has a similar pattern in bothCP-Nets, we update the number of common edges ( commonEdges ) and all edges( allEdges ). The preference is added in visitedP ref erences . However, if there isno preference from CP A is found, the algorithm updates only the number of all CP-Net based Qualitative Composition Approach for an IaaS Provider 11
Table 1: Experiment Variables
Variable Names Values
Simulation Run 100Number of Consumers 2 to 23Coefficient of Correlation 0.15, 0.20, 0.25Homogeneous Domain Size 20 edges ( allEdges ). allEdges is updated with every iteration. The relative similar-ity is calculated by the ratio of the number of common edges ( commonEdges )and the number of all edges ( allEdges ). We have conducted a set of experiments to evaluate the efficiency and the feasi-bility of the proposed heuristic based composition approach. The heuristic andthe greedy approaches are compared with the brute force approach in term ofaccuracy and time. We conducted the experiment on computers with Intel Corei7 (3.60GHz and 8GB RAM) using Java and Matlab.
It is difficult to find the real-world preferences of IaaS consumers. We have gener-ated 20 CP-Nets to represent consumers’ preferences. We have also generated asemantic table for consumers which is a subset of provider’s semantic table. Theprovider has the entire view of its resource capacity. As the simulation has beenperformed based on randomly generated CP-Nets, the result varies depending onthe type of the CP-Nets. We run the experiment based on Monte Carlo [3] sim-ulation method for a conclusive result. We have run the simulation several timesfor each approach and taken the average accuracy and time for the different sizeof consumers. Table 1 shows the simulation variables and their correspondingvalues that we have used in the experiment to perform the performance analysis.
We generate all combination of the consumers for the brute force approach. Foreach combination, we compose them using the composability model. The setof composable CP-Nets are composed and compared with the provider’s CP-Net using equation 1. A composed CP-Net that has maximum similarity withthe provider’s CP-Net is selected. As the brute force approach considers allconsumers, it achieves maximum similarity up to 90% with provider’s CP-Net.
We have applied the proposed heuristic approach to select and compose rela-tively similar consumers according to the provider’s CP-Net. The accuracy ofthe heuristic approach is calculated with respect to the brute force approach.
Accruacy F r e qu e n c y (a) Accruacy F r e qu e n c y (b) Accruacy F r e qu e n c y (c) Accruacy F r e qu e n c y (d) Number of Consumers A cc u r ac y Heuristic ApproachGreedy Approach (e)
Number of Consumers -2 T i m e ( m ili s ec ) Brute force approachHeuristic ApproachGreedy Approach (f)
Fig. 4: Accuracy of the proposed heuristic approach with Coeffecient (a) 0.15 (b)0.20 (c) 0.25 (d) Accuracy of the greedy approach (e) Accuracy of the heuristicand greedy approach (f) Execution time in log scaleWe also compose the consumers based on the greedy selection approach. Theaccuracy of the greedy approach is calculated in a similar manner.The brute force approach normally provides more accurate result then theheuristic approach as the brute force approach considers all possible combinationof consumers. For the heuristic approach, we select a consumer in the compo-sition only if its CP-Net is more than 20% relatively similar to the provider’sCP-Net. We have run the experiments several times to find the optimal thresh-old. For a specific provider and a service, this threshold should be set manuallybefore composition. Figure 4(a), (b), and (c) show the histogram of the accu-racy of the heuristic approach where the correlation coefficients are 0.15, 0.20,
CP-Net based Qualitative Composition Approach for an IaaS Provider 13 and 0.25. The brute force approach provides the optimal results. Compared tothe brute force result, the heuristic approach generates almost 60% accurateresult on average when the coefficient is 0.2. Figure 4(d) shows the accuracyresult of the greedy based approach. The result shows the histogram of the out-comes. Here, outcomes have below 50% accuracy most of the time. Figure 4 (e)shows the average accuracy of the proposed heuristic approach and the greedyapproach. The greedy approach provides very good accuracy if the number ofconsumers is low. The accuracy of the greedy approach becomes low with theincrease of consumers, as it starts to discard more consumers. The heuristic ap-proach provides better accuracy with the increase in the number of consumersbecause it finds more similar consumers according to the provider’s CP-Net.
Figure 4(f) depicts the time comparison between the brute force, heuristic, andgreedy approaches in log scale. The figure shows that with the increase in thenumber of consumers, time for composition in the brute force approach growsexponentially ( i.e., linearly in log scale). The heuristic approach does not showexponential behavior. For a particular IaaS provider, the composition time in-creases very slowly compared to the brute force approach. The greedy approachshows a very interesting result. It composes consumers with a constant time fora particular service. The accuracy of the greedy approach is unreliable.
Qualitative user-preferences are represented by graphical models, especially inmulti-objective decision-making domain [7]. IaaS consumers can express theirpreferences more directly, intuitively using qualitative representations. A condi-tional preference network (CP-Net) provides a natural and an efficient way torepresent consumers’ preferences in qualitative manner [5]. CP-Nets are widelyused to represent user’s preferences to select and compose services [18]. A webservice selection mechanism is proposed to incorporate incomplete or inconsis-tent user preferences from historical preferences [18]. A CP-Net based similaritymeasurement approach is proposed to find users with similar preferences [19].Several variations of CP-Nets are also proposed to enhance the expressiveness ofusers. A UCP-Net is a graphical representation of utility functions that combinesgeneralize additive models and CP-Nets [4]. The TCP-Net is another variation ofCP-Net where relative importance between attributes can be captured throughweighted edges [15]. The WCP-Net is another weighted CP-Net that is proposedto capture user’s preference more precisely to select web services. A deterministictemporal CP-Net is used to express IaaS provider’s long-term business strategies[11]. A probabilistic CP-Net is proposed to capture the IaaS provider’s businessstrategies in a probabilistic manner [12].Several existing studies propose different methods to compose multiple CP-Nets. A multi-agent CP-Net or mCP-Net is proposed as an extension of CP-Net [16]. Preferences from multiple users are aggregated into a single CP-Net. Theproposed method aggregates preferences according to a voting analogy wherepreferences are selected based on the preference of majority user. A Majority-rule-based preference aggregation method is proposed based on a hypercube-wise composition to optimize the composition process. An aggregation methodis proposed to capture multi-valued CP-Nets based on majoritarian aggregationrule [9]. The proposed method can aggregate CP-Nets even if they are cyclic.Most existing works on the aggregation of CP-Nets consider the compositionproblem as a multi-agent voting system. These approaches are not applicable inthe case of resource allocation based on multi-agent preferences. We composeCP-Nets based on the composability model and the resource constraints in ourproposed composition approach. The composed CP-Nets capture the preferencesof multi-users instead of just considering the common preferences.
We propose a CP-Net based composition approach for an IaaS provider. Theproposed approach allows the IaaS provider and consumers to express theirqualitative conditional preferences using CP-Nets. We propose a composabilitymodel for IaaS consumers using the semantic congruence property of a quali-tative composition. Finding the optimal composition may be difficult when thenumber of consumers is large. A greedy-based and a heuristic-based selection ap-proaches are proposed to reduce the search space of candidate consumers. Bothapproaches utilize correlation coefficients between CP-Nets to find consumerswho have similar preferences with the provider. Experimental results show thatthe proposed heuristic-based approach is applicable to the runtime and the per-formance is acceptable. One key limitation of the proposed approach is that itconsiders only the deterministic model of CP-Nets. In the future, we want toexplore the composability model of probabilistic CP-Nets in the context of thelong-term IaaS composition.
10 Acknowledgement
This research was made possible by NPRP 7-481-1-088 grant from the QatarNational Research Fund (a member of The Qatar Foundation). The statementsmade herein are solely the responsibility of the authors.
References
1. Armbrust, M., Fox, A., Griffith, R., Joseph, A.D., Katz, R., Konwinski, A., Lee,G., Patterson, D., Rabkin, A., Stoica, I., et al.: A view of cloud computing. Com-munications of the ACM (4), 50–58 (2010)2. Balke, W.T., Wagner, M.: Towards personalized selection of web services. In:WWW (Alternate Paper Tracks). pp. 20–24 (2003) CP-Net based Qualitative Composition Approach for an IaaS Provider 153. Binder, K., Heermann, D., Roelofs, L., Mallinckrodt, A.J., McKay, S.: Monte carlosimulation in statistical physics. Computers in Physics (2), 156–157 (1993)4. Boutilier, C., Bacchus, F., Brafman, R.I.: Ucp-networks: A directed graphical rep-resentation of conditional utilities. In: Proceedings of the Seventeenth conferenceon Uncertainty in artificial intelligence. pp. 56–64. Morgan Kaufmann PublishersInc. (2001)5. Boutilier, C., Brafman, R.I., Domshlak, C., Hoos, H.H., Poole, D.: Cp-nets: A toolfor representing and reasoning with conditional ceteris paribus preference state-ments. J. Artif. Intell. Res.(JAIR) , 135–191 (2004)6. Chaisiri, S., Lee, B.S., Niyato, D.: Optimization of resource provisioning cost incloud computing. IEEE Transactions on Services Computing (2), 164–177 (2012)7. Lang, J., Xia, L.: Sequential composition of voting rules in multi-issue domains.Mathematical social sciences (3), 304–324 (2009)8. Li, L., Liu, D., Bouguettaya, A.: Semantic based aspect-oriented programmingfor context-aware web service composition. Information Systems (3), 551–564(2011)9. Li, M., Vo, Q.B., Kowalczyk, R.: Aggregating multi-valued cp-nets: a csp-basedapproach. Journal of Heuristics (1), 107–140 (2015)10. Liu, X., Bouguettaya, A., Wu, J., Zhou, L.: Ev-lcs: A system for the evolutionof long-term composed services. IEEE Transactions on Services Computing (1),102–115 (2013)11. Mistry, S., Bouguettaya, A., Dong, H., Erradi, A.: Qualitative economic modelfor long-term iaas composition. In: International Conference on Service-OrientedComputing. pp. 317–332. Springer (2016)12. Mistry, S., Bouguettaya, A., Dong, H., Erradi, A.: Probabilistic qualitative pref-erence matching in long-term iaas composition. In: International Conference onService-Oriented Computing. pp. 256–271. Springer (2017)13. Mistry, S., Bouguettaya, A., Dong, H., Qin, A.K.: Predicting dynamic requestsbehavior in long-term iaas service composition. In: ICWS. pp. 49–56. IEEE (2015)14. Mistry, S., Bouguettaya, A., Dong, H., Qin, A.: Metaheuristic optimization forlong-term iaas service composition. IEEE Transactions on Services Computing(2017)15. Mukhtar, H., Belaïd, D., Bernard, G.: A quantitative model for user preferencesbased on qualitative specifications. In: ICPS. pp. 179–188. ACM (2009)16. Rossi, F., Venable, K.B., Walsh, T.: mcp nets: representing and reasoning withpreferences of multiple agents. In: AAAI. vol. 4, pp. 729–734 (2004)17. Wang, H., Shao, S., Zhou, X., Wan, C., Bouguettaya, A.: Web service selectionwith incomplete or inconsistent user preferences. In: International Conference onService-Oriented Computing, pp. 83–98. Springer (2009)18. Wang, H., Tao, Y., Yu, Q., Lin, X., Hong, T.: Incorporating both qualitative andquantitative preferences for service recommendation. Journal of Parallel and Dis-tributed Computing (2017)19. Wang, H., Wang, H., Guo, G., Tang, Y., Zhang, J.: Measuring similarity of userswith qualitative preferences for service selection. Knowledge and Information Sys-tems (2), 561–594 (2017)20. Ye, Z., Bouguettaya, A., Zhou, X.: Qos-aware cloud service composition based oneconomic models. In: International Conference on Service-Oriented Computing.pp. 111–126. Springer (2012)21. Ye, Z., Mistry, S., Bouguettaya, A., Dong, H.: Long-term qos-aware cloud servicecomposition using multivariate time series analysis. IEEE Transactions on ServicesComputing9