JP2009245362A - Automatic distribution system and method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an automatic distribution system or the like that automatically determines the arrangement optimum for functional modules on the basis of analysis information of a program, even if not determining the arrangement of all functional modules. <P>SOLUTION: The automatic distribution system includes a program analysis unit, graph creation unit, minimum cut calculation unit, and distribution system generation unit. The program analysis unit analyzes the program including the plurality of functional modules, and acquires a reference relationship between the functional modules and calling frequencies between the functional modules in program execution. The graph creation unit creates a weighted graph on the basis of the reference relationship and the number of times of calling. The minimum cut calculation unit calculates the minimum cut of the weighted graph. The distribution system generation unit automatically distributes a plurality of modules on the basis of the minimum cut. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to the field of distributed systems, and more particularly to an automatic distributed system and an automatic distributed method that automatically perform optimal arrangement of functional modules on a distributed system based on information obtained by analyzing a program.

  Distributed systems are commonly used in many software applications. A typical distributed system consists of a plurality of hosts that communicate via a network. Each functional module is running on multiple hosts. This functional module is a program, and the developer separates the functions of the distributed system into a plurality of hosts, and performs design and development. When a distributed system is operating, each function module that operates on multiple hosts realizes that function on each host, and calls a function module that operates on another host via the network. Operates as a system.

  FIG. 1 is an explanatory diagram of a distributed system. In FIG. 1, two hosts 10 and 12 are connected by a network 14. In the host 10, two functional modules A1, B1 are operating, and in the host 12, three functional modules C1, D1, E1 are operating. The function module A1 of the host 10 can call the function module D1 of the host 12, and the function module D1 of the host 12 can call the function module A1 of the host 10. Further, the function module B1 of the host 10 can call the function module C1 of the host 12, and the function module C1 of the host 12 can call the function module B1 of the host 10.

  However, it is generally difficult to develop a distributed system. In particular, the development of the part that calls the functional module over the network may be performed using support by a tool, etc., and some of the tools include an automatic distributed system that does not allow the developer to consider calling over the network in the first place. Exists (Addistant, J-Orchestra; Japanese Patent Laid-Open No. 11-085519 (see Patent Document 1)).

  In particular, an automatic decentralized system that does not allow developers to take into account calls over the network will be described in detail with reference to FIG. First, a developer develops a non-distributed system on the assumption that the distributed system is operated on a single host 20. Next, the developer determines on which host the function modules A2 to E2 are to be placed. This arrangement setting is performed by creating a table 28 that stores combinations of function module names and host names.

  Based on the input of the non-distributed system consisting of the five functional modules A2 to E2 and the table 28, the automatic distributed system 26 converts the non-distributed system operating on the single host 20 into a distributed system. In the example of FIG. 2, based on the table 28, the two functional modules A 2 and B 2 are arranged on the host [Host 1] 22, and the three functional modules C 2, D 2 and E 2 are arranged on the host [Host 2] 24. is doing. In this application, such a conversion system is called an automatic decentralized system.

  The problem with the automatic decentralization system shown in FIG. 2 is that the developer has to determine the arrangement of all functional modules A2 to E2. The reason is that the automatic distribution system of FIG. 2 does not take into account the automatic determination of the optimal arrangement of functional modules.

  In addition, related technology of the distributed system is described in Japanese Patent Laid-Open No. 03-182960 (see Patent Document 2). In the invention of the “parallel computer system” described in this publicly known document, the parallel computer system is expressed using a graph. The nodes in this graph represent processors. Further, the edge of this graph represents a communication path between processors.

  Japanese Patent Laying-Open No. 06-075786 (see Patent Document 3) describes the invention of “task scheduling method”. In this known document, a communication cost is obtained in order to determine the priority order of task schedules (Patent Document 3 FIGS. 5 and 6). This communication cost is calculated by [overhead time of communication start-up] + [data transfer amount] × [data transfer time per unit data amount].

Japanese Patent Laid-Open No. 11-085519 Japanese Patent Laid-Open No. 03-182960 Japanese Patent Application Laid-Open No. 06-075786

  An object of the present invention is to provide an automatic decentralization system and an automatic decentralization method that automatically determine an optimal arrangement of functional modules based on program analysis information without determining the arrangement of all functional modules. There is.

  One automatic distributed system of the present invention includes a program analysis unit, a graph creation unit, a minimum cut calculation unit, and a distributed system generation unit. The program analysis unit analyzes a program including a plurality of functional modules, and obtains the reference relationship between the functional modules and the number of calls between the functional modules during program execution. The graph creation unit creates a weighted graph based on the reference relationship and the number of calls. The minimum cut calculation unit calculates the minimum cut of the weighted graph. The distribution system generation unit automatically distributes a plurality of modules based on the minimum cut.

  Another automatic decentralization method of the present invention comprises obtaining, creating, calculating, and decentralizing. In the acquisition, a program including a plurality of functional modules is analyzed, and the reference relationship between the functional modules and the number of calls between the functional modules at the time of program execution are acquired. In creating, a weighted graph is created based on the reference relationship and the number of calls. In the calculation, the minimum cut of the weighted graph is calculated. In the distribution, a plurality of modules are automatically distributed based on the minimum cut.

  According to the present invention, it is not necessary for the developer to determine an optimal arrangement for all the functional modules. The reason is that the present invention automatically determines the optimal arrangement of functional modules.

  One of the best modes for carrying out the present invention will be described in detail with reference to the drawings. With reference to FIG. 3, the configuration and operation of an automatic decentralization system according to an embodiment of the present invention will be described. The automatic decentralization system in the present embodiment includes a program analysis unit, a graph creation unit, a minimum cut calculation unit, and a distributed program generation unit.

  As shown in FIG. 3, in the automatic decentralization system according to the present embodiment, first, a program analysis unit analyzes a program (S30). The program analysis unit has a function of analyzing a non-distributed program assuming that a distributed system is operated on a single host. The program analysis unit obtains the reference relationship between the function modules and the number of times each function module is called during program execution. The analysis target can be a source code, a binary code, or an intermediate code.

  Next, the automatic decentralization system according to the present embodiment creates a weighted graph by the graph creation unit (S32). The graph creation unit has a function of generating a weighted graph based on the information obtained by the program analysis unit. Each node of the weighted graph represents one of a plurality of functional modules included in the input program, and each edge represents a reference relationship between the functional modules. Further, each cost of the weighted graph is defined by the following equation.

[Cost] = α [Number of calls to each function module] x β [Amount of argument data when calling]

  Here, the number of calls to each functional module is one piece of information obtained by the program analysis unit, and can be obtained as a result of executing the program. The data amount of the argument at the time of calling is the total amount of data to be passed to the callee function module when the function module is called. Α and β are constants that can be set by the user of the automatic decentralized system.

  In the automatic decentralization system in the present embodiment, the minimum cut calculation unit subsequently calculates the minimum cut (S34). The minimum cut calculation unit has a function of calculating the minimum cut of the graph created by the graph creation unit. However, the minimum cut calculation unit can calculate the minimum cut according to the restriction when the restriction of the arrangement setting is given. For example, a functional module designated by the user not to be placed on the same host is placed on a different host. At this time, those designated functional modules do not exist in the same subgraph.

  In the automatic decentralization system in the present embodiment, finally, the distributed system generation unit performs conversion into a distributed program (S36). The distributed system generation unit converts the program into a distributed system based on the subgraph calculated by the minimum cut calculation unit, and outputs the converted system.

  Assume that a developer has created a non-distributed program that runs on a single host. In order to facilitate understanding, it is assumed that this program is composed of a plurality of functional modules A2 to E2 as in the case of FIG. The automatic distributed system according to the first embodiment automatically generates a program for a distributed system with a program including a plurality of functional modules A2 to E2 as an input.

  When the automatic decentralization system receives a program as an input, first, the program analysis unit analyzes the program. By this analysis, the program analysis unit obtains the reference relationship between the function modules A2 to E2 and the number of calls of each function module at the time of execution. Next, the graph creation unit generates a weighted graph as shown in FIG. 4 based on the analysis information. In FIG. 4, the nodes of the graph represent functional modules A2 to E2 on the program, and the edges of the graph represent the calling relationship of the functional modules A2 to E2. The weight of the weighted graph is a cost defined by the following equation, and is calculated as c1 to c4.

[Cost] = Σα [Number of calls to each function module] × β [Amount of argument data when calling]

  Here, the number of calls to each of the function modules A2 to E2 is execution time information obtained by the program analysis unit, and the data amount of the argument at the time of the call is determined by the callee function module when the function module is called. This is the total amount of data to pass. Α and β are constants that can be set by the user.

  For example, as shown in FIG. 5, two function modules A2 and B2 call each other's function modules B2 and A2, and the function module A2 has three functions f0, f1 and f2, and the function module B2 Suppose that it has four functions f3, f4, f5, and f6. In this case, the cost between the functional module A2 and the functional module B2 is calculated as in Equation 1.

  In Equation 1, “count” is a function for counting the number of times of calling the function fi. For example, count (f0) means the number of times that the function module B2 calls the function f0 of the function module A2. Also, size is the total amount of data that is passed when calling the function fi. For example, size (f0) means the total amount of data that is passed when the function module B2 calls the function f0 of the function module A2.

  Subsequently, the minimum cut calculation unit calculates the minimum cut of the weighted graph of FIG. Finally, the distributed program generation unit converts the input program into a distributed program based on the calculated minimum cut and outputs it. For example, a distributed program having the same arrangement as that shown in FIG. 2 is output.

  The automatic decentralization system can be given a restriction of arrangement setting. FIG. 6 is a diagram illustrating a table storing arrangement setting constraints. In FIG. 6, a table 60 stores combinations of functional module names and host names. Referring to the contents of the table 60, the arrangement setting constraint indicates that the functional module A2 must be arranged in the host [HOST1], and the functional module E2 must be arranged in the host [HOST2]. Yes.

  When the automatic decentralization system inputs a layout setting constraint, the minimum cut calculation unit performs a calculation according to the layout setting constraint when calculating the minimum cut. For example, when the table of FIG. 6 is input as the restriction of the arrangement setting, the automatic decentralization system can recognize the function module A2 and the function module E2. Are subject to the restriction that cuts that should be placed in the same subgraph must not be selected.

FIG. 1 is an explanatory diagram of a basic distributed system. FIG. 2 is an explanatory diagram of an automatic decentralization system in the related art. FIG. 3 is a flowchart for explaining the operation of the automatic decentralization system according to this embodiment. FIG. 4 is an explanatory diagram of a program modeled by a weighted graph. FIG. 5 is an explanatory diagram of a cost calculation method. FIG. 6 is a diagram illustrating a table storing arrangement setting constraints.

Explanation of symbols

10, 12, 20, 22, 24 Host 14 Network 26 Automatic decentralized system 28, 60 Table

Claims (10)

A program analysis means for analyzing a program including a plurality of functional modules and obtaining a reference relationship between the functional modules and the number of calls between the functional modules during program execution;
Graph creating means for creating a weighted graph based on the reference relationship and the number of calls;
A minimum cut calculating means for calculating a minimum cut of the weighted graph;
An automatic distribution system comprising distribution system generation means for automatically distributing the plurality of modules based on the minimum cut. The graph creating means includes
The automatic decentralized system according to claim 1, wherein the weighted graph is created using a node of the weighted graph as a functional module and an edge of the weighted graph as a reference relationship between the functional modules. The graph creating means includes
The automatic decentralized system according to claim 2, wherein the weighted graph is calculated based on a cost proportional to the number of calls, and the weighted graph is created. The graph creating means includes
The automatic decentralized system according to claim 3, wherein the weight is calculated based on a cost that is also proportional to an amount of data of an argument when the function module is called, and the weighted graph is created. The graph creating means includes
When α and β are constants, the cost is
[Cost] = α [Number of calls between functional modules] x β [Amount of argument data at the time of call]
The automatic decentralization system according to claim 4, wherein the value is defined by the mathematical formula. The minimum cut calculation means is:
The automatic decentralization system according to claim 5, wherein when a restriction on arrangement setting is given, the minimum cut is calculated according to the restriction. The minimum cut calculation means is:
The automatic decentralization system according to claim 6, wherein data associating one of the plurality of function modules with a host type that executes the function module is given as the arrangement setting constraint. Analyzing a program including a plurality of functional modules, obtaining a reference relationship between the functional modules and the number of calls between the functional modules during program execution;
Creating a weighted graph based on the reference relationship and the number of calls;
Calculating a minimum cut of the weighted graph;
Automatically distributing the plurality of modules based on the minimum cut. The creation is
The automatic distribution method according to claim 8, further comprising: creating the weighted graph by using a node of the weighted graph as a functional module and using an edge of the weighted graph as a reference relationship between the functional modules. The creation is
The automatic calculation according to claim 9, further comprising: calculating a weight of the weighted graph based on a cost proportional to the number of calls and a data amount of an argument at the time of the call, and creating the weighted graph. Decentralization method.

Images