Research on Intelligentization of Cloud Computing Programs Based on Self-awareness

Cloud computing technology itself has an architecture, which summarizes the main features of different solutions. According to the related technologies and services provided by cloud computing, the architecture of cloud computing technology is divided into four layers: physical resource layer, resource pool layer, management middleware layer and SOA (service-oriented architecture) construction layer. However, each solution in the architecture of cloud computing may only realize some of its functions, and some relatively minor functions have not been summarized, so intelligent cloud computing is proposed. The architecture of intelligent cloud computing consists of six parts: belief, ability, desire, planning, execution and behavior control mechanism. Among them, belief represents the description of recognized entities and their relationships by Ego, entities are acquaintance sets and objects sets, and relationships are class sets. Belief also represents descriptive knowledge; Ability represents the operation that Ego can perform, that is, process knowledge; The intelligentization of cloud computing programs, as a capability of Ego, provides an intelligent program service to the real world. Desire is a set of states that Ego wants to achieve, and it is a description of the desired state that Ego produces after perceiving and understanding information. Planning is an Ego's plan to solve problems by combining strategic knowledge (special process knowledge) in order to fulfill its current wishes. Execution means the execution of the planning scheme; The behavior control mechanism is the “controller” of Ego, which is responsible for the coordination of various modules and the intelligent operation of Ego. The architecture of intelligent cloud computing is shown in Figure 2.

Figure 2.

All the subjects in the cloud computing cluster, that is, the Agent in the cloud computing field, abstract all the subjects known by Ego. Intelligent cloud cluster usually adopts a master-slave structure, with Ego as the manager of the cluster, while acquaintances in the cluster are ordinary members. The difference between them lies in the different execution rights and tasks within the cluster. AORBCO model divides Agents into Ego and Acq_Agent. There is an equal relationship between acquaintances recognized by Ego, that is, acquaintances in Ego's cognition are at the same level, and there is also an unequal relationship between superiors and subordinates, that is, acquaintances in Ego's cognition are at different levels. In the cluster of intelligent cloud computing, Ego is the global manager and the system management and task allocation of the whole cluster. All the subjects in the cloud computing cluster, that is, the Agent in the cloud computing field, abstract all the subjects known by Ego. Intelligent cloud cluster usually adopts a master-slave structure, with Ego as the manager of the cluster, while acquaintances in the cluster are ordinary members. The difference between them lies in the different execution rights and tasks within the cluster. AORBCO model divides Agents into Ego and Acq_Agent. There is an equal relationship between acquaintances recognized by Ego, that is, acquaintances in Ego's cognition are at the same level, and there is also an unequal relationship between superiors and subordinates, that is, acquaintances in Ego's cognition are at different levels. In the cluster of intelligent cloud computing, Ego is the global manager and the system management and task allocation of the whole cluster.

The detailed task processing process of intelligent cloud computing cluster is as follows: the client submits a task requirement to intelligent cloud cluster. After the task is submitted, if the requirements of the task are solved by desire decomposition by Ego, the sub-wishes decomposed by Ego for different wishes will be different and the sub-wishes will match different abilities in the Ego competence library; If the requirements of tasks need to be solved by the desire decomposition of Ego and acquaintances, Ego is the manager of the whole cluster, and acquaintances are responsible for data processing of tasks. Because acquaintances have different abilities, each acquaintance has different degrees of decomposition of wishes, and the degree of decomposition granularity of wishes depends on acquaintances' computing ability. According to the intimacy with acquaintances and Ego's cognition of acquaintances' ability, Ego assigns specific tasks to acquaintances in the cluster after wish decomposition, selects some acquaintances to generate application planners for each specific task, submits task planning applications to Ego, manages each specific task and reports the task implementation to Ego. Other acquaintances who perform specific computing tasks will call MapReduce capabilities through the capability library to perform computing tasks. According to the implementation of each small task and their own state information, they are provided to acquaintances with application planners. The architecture of intelligent cloud computing cluster is shown in Figure 3.

Figure 3.

The implementation mechanism in the intelligent cloud computing architecture is modeled around Ego. The model divides the process of Ego understanding the world or solving problems into five steps: perception, understanding, planning, execution and learning. In the AORBCO model, the research on the intelligence of cloud computing programs is to establish contact with Ego's beliefs, abilities and wishes after Ego perceives the problem demand of the external world, so as to generate a plan for generating MapReduce programs for solving problems and realize the generation of cloud computing programs. Execute an operation that represents the action generated for the plan; Learning refers to the experience summary of Ego, including unintentional learning and intentional learning. Unintentional learning refers to the habitual change of ability proficiency when Ego knows the world. Intentional learning means that when Ego encounters new problems, it learns new abilities from acquaintances or solves new problems through its own planning. Based on the architecture of intelligent cloud computing, the implementation mechanism of intelligent cloud computing is shown in Figure 4.

Figure 4.

Secondly, in order to realize the intelligentization of cloud computing program, it is necessary to divide the business logic of big data processing into different types, so as to carry out targeted program intelligentization for each type. It is also necessary to determine the type of task requirements they belong to before a particular MapReduce program is built. For the document requirements with significant text feature information, then according to the keyword matching text feature class, the input document requirements are automatically determined. Match the MapReduce program based on the document requirements. The cosine similarity can be used to calculate the text content and the prior knowledge in the Ego knowledge base for similarity calculation, so as to judge the actual demand of the document and provide intelligent matching MapReduce program code according to the demand.

TextRank algorithm is to convert a document into a directed weighted word graph model, which divides the text into basic units, namely words. Each basic unit is regarded as a node, and the edge between each node is determined by the cooccurrence relationship between word nodes, while the importance of nodes is determined by the number of pointing adjacent nodes. Construct TextRank keyword graph G = (V, E), where V is node set and E is edge set between nodes. TextRank algorithm is calculated as follows. (1) WS(Vi)=(1−d)+d×∑Vj∈in(Vi)wji∑Vk∈out(Vj)wjkWS(Vj) WS\left( {{V_i}} \right) = \left( {1 - d} \right) + d \times \mathop \sum \limits_{{V_j} \in in\left( {{V_i}} \right)} {{{w_{ji}}} \over {\sum\limits_{{V_k} \in {\rm{out}}\left( {{V_j}} \right)} {{w_{jk}}} }}WS\left( {{V_j}} \right)

Which ln(v_i) represents the node set pointing to the node v_i; out(v_j) represents the set of nodes pointed by v_j, W_ij represents the weight from node W_j to W_i-side. d is the damping coefficient, which represents the probability of transferring from any designated node in the graph to other nodes, and its value range is [0,1]. If the damping coefficient is too large, the number of iterations will increase sharply and the ranking of the algorithm will be unstable; If the damping coefficient is too small, the iterative process has no obvious effect, which ensures that the weight can be stably transferred to convergence. Finally, the weight of each word is calculated and sorted, and generally it is 0.85 [14].

TextRank algorithm extracts keywords as key entity information, and judges the similarity of the class to which the entity belongs. Sim A and B represent the similarity of two entities A and B, where A is the key entity extracted from the text, B is the entity in the Ego knowledge base, and R represents the prediction result. When R is +1, it indicates that the two entities are related, and the relationship can be judged. Each dimension of the vector represents the degree to which the entity belongs to the I-th class. By calculating the cosine similarity between two vectors, the similarity of the class is obtained. Cosine similarity is a measure of the similarity of two vectors in the direction. The calculation process of cosine similarity is as follows: (2) Sim(A,B)=A⋅B‖A‖×‖B‖=∑ni=1(Ai×Bi)∑ni=1Ai2×∑ni=1Bi2 Sim\left( {{\rm{A}},{\rm{B}}} \right) = {{A \cdot B} \over {\left\| A \right\| \times \left\| B \right\|}} = {{\sum\limits_n^{i = 1} {\left( {{A_i} \times {B_i}} \right)} } \over {\sqrt {\sum\limits_n^{i = 1} {A_i^2} } \times \sqrt {\sum\limits_n^{i = 1} {B_i^2} } }}

In the formula, a and b are two input vectors respectively, and the value range of is [−1,1].

The formula of the sign(x) function is shown as follows: (3) R=sign(Sim(A,B)) R = {\rm{sign}}\left( {{\rm{Sim}}\left( {{\rm{A}},{\rm{B}}} \right)} \right) (4) sign(x)={−1,x<0+1,x≥0 sign\left( x \right) = \left\{ {\matrix{ { - 1,x < 0} \hfill \cr { + 1,x \ge 0} \hfill \cr } } \right.

If you can't find this type of document requirements in the knowledge base of Ego, you can enhance the semantic recognition of requirements by user input and natural language processing. In natural language, the key information in the sentence is extracted to the triple information in natural language, and Ego uses the same method to calculate and judge the customer's requirements, thus giving a MapReduce program for this requirement.

The matching of MapReduce module features is closely related to the capability alignment in AORBCO model, and the formal definition of capability Function in AORBCO model is shown in the formula: (5) F(t)={name,parameter,precondition,postcondition,body} F\left( t \right) = \left\{ {name,parameter,precondition,postcondition,body} \right\}

Where name represents the name of the ability and parameter represents the parameters of the ability, including input parameters and output parameters (return value). Preconditions and postconditions are constraints on parameters, where preconditions represent preconditions and postconditions represent postconditions. Body is the ability body, which represents a series of operations from the initial state to the target state of ability execution. The constraints of preconditions are the types and numbers of parameters and the judgment of the class membership degree of Ego's requirements for text documents, and these constraints are expressed by logical expressions. The constraints of postconditions are to judge the parameter types and numbers of the corresponding capabilities of Ego, and these constraints are expressed by logical expressions. According to the number and type of parameters and the description of preconditions and postconditions, the logical expressions of parameters are merged through the conjunction and disjunction relationship. If the corresponding capability description file can be matched in the capability library of Ego, then the corresponding MapReduce program can be matched according to the module feature template of the capability. If the corresponding capability is not matched from the capability library, similar MapReduce programs can be recommended through fuzzy matching.

Based on the data processing tasks applicable to MapReduce program, the corresponding Map and Reduce function operations and parameters are summarized, and the corresponding MapReduce programming module features are formed. The Map function takes KEYIN1 and VALUEIN1 as input parameter types, and the Context class object completes the writing of output content, then the reduce function indicates the types of keys and values with KEYOUT2 and VALUEOUT2 respectively, and the types of input keys and values are the same as the output types of the Map function. Figure 5 shows the module feature diagram of MapReduce programming model.

Figure 5.

Finally, the feature selection of MapReduce function and Reduce function is matched by rule based method. Production rules should be the embodiment of strategic knowledge application in intelligent model theory. In intelligent MapReduce, users' demands are received through the perception module in the behavior control mechanism of AORBCO model. The selection according to the characteristics of Map function is understanding, and the planning and execution of Map, Shuffle and Reduce processes are planning and execution. The MapReduce program in the intelligent cloud determines the data processing type based on the text content and matches corresponding rules. After matching different rules, different Map functions are selected for processing. The Map function processing is handed over to the Reduce function for processing and the final processing result is returned. According to input and output type information and code template, the behavior control mechanism in AORBCO model is used to generate the corresponding code. Code templates include Job configuration templates, Mapper templates, Reduce templates, and Key/Value templates related to the MapReduce execution platform. The model ultimately generates MapReduce code based on user-provided input.

On the premise that Ego owns various MapReduce programs, the priori knowledge and behavior control mechanism of AORBCO model are used to analyze the data processing types of text contents, determine the class of contents, and then find the MapReduce program code that can be solved from Ego's ability through MapReduce module feature matching. This experiment verifies that the average score is obtained by inputting the text document, and Ego senses the content through prior knowledge. The background perception input and results are shown in the figure respectively. Since this experiment mainly verifies the feasibility of MapReduce program generation in intelligent cloud on AORBCO model platform, it does not involve communication between Ego and acquaintances, asking acquaintances for assistance, etc., and Ego is used to solve problems independently by default.

Figure 6.

In the MapReduce program generated by AORBCO model development platform, the benchmark program WordCount is selected as the experiment case. The experimental data set is a random text data generated by Hadoop RandomTextWriter. In this experiment, the task completion time of word frequency statistics was analyzed by comparing MapReduce program and Java program under different data set document sizes. As shown in Figure 7, experimental tests show that Java programs perform better with smaller input document data sizes. As the size of the input document data increases, the MapReduce program performs better than the Java program, reflecting the advantages of MapReduce program distributed computing.

Figure 7.