JP2000298593A - System and method for predicting performance of multitask system and recording medium stored with program for the method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately predict a performance index for the parallel degree of parallel computers in a multitask environment. SOLUTION: When the specification of a parallel degree or the like of parallel computers to be a platform is inputted to an input part 11 in the performance prediction system, a model generation part 12 generates a model on the basis of the specification. A prediction execution part 13 calculates performance index prediction values such as the degree of improvement and more detailed performance indexes such as throughput, a response and a resource using rate from the generated model. A prediction result output part 14 visually expresses the values outputted from the execution part 13.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a system and method for predicting the performance of a multitask system and a recording medium on which a method program is recorded, and more particularly to a system and a method for predicting the performance of a multitask system using a parallel computer as a platform. The present invention relates to a recording medium on which a method program is recorded.

[0002]

2. Description of the Related Art An AFIPS Conference Pr issued in 1967 is a method for predicting a degree of improvement (an index indicating an improvement in processing performance with respect to a degree of parallelism (the maximum number of tasks and threads to be simultaneously executed)) in a parallel computer.
“Validity of the single-” published on pages 483 to 485 of the receiveds.
processor approach to ach
ieving largescale computing
ng "in a paper by G. Amdahl. In this conventional method, a single task environment is assumed, and a method of predicting the degree of improvement in processing performance with respect to the degree of parallelism is assumed based on the assumption. It is shown.

[0003] Also, CACM V issued in 1988
ol. 31 (5), pages 532-533 "
Reevaluating Amdahl's La
w ", a paper by J. Gustafson, 19
The 227th parallel processing symposium published in 1996
A paper by Furuichi et al. Entitled "Configuration of a Behavior Prediction Model for Performance Evaluation of Highly Parallel Computers" on page 234 also shows a prediction method of the degree of improvement based on the assumption of a single task environment. .

[0004] Further, a prediction method when a multitask environment is assumed is disclosed in Commun. AC
“Including Queueing E,” published in Article 231 of M 39 (12).
R. Nelson, entitled "Facts in Amdahl's Law". In this paper, parameters that take into account processor competition between tasks (or threads) in a multitasking environment are incorporated into the prediction formula.

[0005] By the way, a multiprocessor-equipped S
MP (symmetrical multiproce
With the relatively low cost of ssor (symmetric multi-processor) machines, parallel computers that have been used for large-scale scientific calculations are becoming more familiar, such as being used for computer system platforms. It is becoming.

In system development, one of the problems when employing a parallel computer as a platform is how much the degree of parallelism is set in the parallel computer. The degree of parallelism often coincides with the number of processors in the computer. Generally, increasing the degree of parallelism improves performance, but this leads to an increase in the number of processors,
The cost for the platform will increase.

[0007] In order to develop a system with high cost performance, it is necessary to set an appropriate degree of parallelism, and to design the system so as to reduce the degree of parallelism as much as possible (within a range satisfying the required processing performance). Preferably it is possible. For that purpose, when a system configuration is given, it refers to an index of processing performance with respect to the degree of parallelism (specifically, speed improvement rate, efficiency, throughput (processing capacity per unit time), response time, resource usage rate, and the like. .) Must be predicted. In the case of predicting such an index, conventionally, it has been mainly assumed to be used in a single task environment. As described above, the prediction method of the speed improvement rate and efficiency under the single task environment (for the definition of these indices, see item 228 of the above-mentioned article by Furuichi et al.) Is described in G. This is shown in papers such as those by Amdah. In addition, a method of predicting the degree of improvement assuming a multitask environment is also described in R. This is shown in the Nelson paper.

FIG. 8 is a block diagram showing an example of a computer system in a multitask environment using a parallel computer as a platform. Referring to FIG. 8, a parallel computer 101 has a plurality of processors, in this example, three processors 102.
To 104. On the other hand, in this example, there are four tasks 105 to 108 as a plurality of tasks to be executed. In this computer system, a plurality of tasks 105 to 108 are simultaneously executed on a computer (parallel computer) 101 on which a plurality of processors 102 to 104 are mounted.

An example of this type of technology is disclosed in Japanese Patent Laid-Open No. 9-23 / 1990.
JP-A-7203, JP-A-62-182864, JP-A-59-174957 and JP-A-10-069
No. 469.

[0010]

A first problem is that it is not possible to correctly predict the degree of performance improvement or performance index with respect to the degree of parallelism of a parallel computer in a multitask environment or a multithread environment. The reason is that in the prediction performed on the assumption of the single task environment, only the wrong prediction can be performed because the resource competition such as the processor competition by a plurality of tasks is not considered. Also,
R. This is because the accuracy of the method shown in the Nelson's paper is questionable because the theory underlying the prediction formula is not shown.

The second problem is that it is necessary to consider and determine the specifications of the parallel computer suitable for the cost, more specifically, the determination of the degree of parallelism and the processing speed of the processor. The reason is that the performance index cannot be properly predicted as shown in the first problem.

The third problem is that the structure of a program executed on a parallel computer has a bottleneck.
eck; inhibition factor) cannot be identified at the location. The reason is that semaphores and critical sections (c
The control is performed using a vertical section, for example, and this often becomes a bottleneck. However, this is because conventional methods do not properly consider the structure of these programs in a multitask environment. Note that "semaphore" refers to a signal used to synchronize tasks when multiple tasks are running simultaneously, and "critical section" refers to an environment where multiple tasks are executed simultaneously.
While a task uses a critical (important) resource, it means that another task waits for the resource to be used until the resource is released, that is, a period during which the task is kept waiting.

A fourth problem is that in order to satisfy the required performance, it is necessary to take into account only the configuration of programs to be executed at the same time and the setting of the system load such as the number of tasks and the number of threads, based only on experience. The reason is that the conventional method does not show a model for properly predicting the performance index of the system in a multitask environment or a multithread environment and an analysis method thereof.

The fifth problem is that the system has a throughput and a response time.
e; response time), resource (resource)
That is, it is impossible to predict a more detailed performance index such as a usage rate. The reason is that in the conventional method, in order to facilitate the analysis, the index to be calculated is limited to a specified index such as a speed improvement rate or efficiency.

It is an object of the present invention to provide a performance prediction system and a prediction method for a multitask system capable of solving the above-mentioned problems, and a recording medium on which a method program is recorded.

[0016]

According to a first aspect of the present invention, there is provided a performance prediction system for a multitasking system using a parallel computer as a platform. Modeling means for modeling resource contention using a queuing network model, and performance index predicting means for analyzing the modeled resource contention using queuing theory to predict a performance index of the multitask system And characterized in that:

A second invention according to the present invention is a method for estimating the performance of a multitask system using a parallel computer as a platform, and the method uses a queuing network model for resource contention in a multitask environment. The method includes a first step of modeling and a second step of analyzing contention of the modeled resources by using queuing theory to predict a performance index of the multitask system.

According to a third aspect of the present invention, there is provided a recording medium recording a program for predicting the performance of a multitasking system using a parallel computer as a platform, wherein the recording medium has resource competition in a multitasking environment. A first step of modeling using a queuing network model;
A second step of analyzing the modeled resource competition using queuing theory and predicting the performance index of the multitask system.

According to the first to third aspects of the present invention, first, it is possible to correctly predict the degree of improvement in performance and the performance index with respect to the degree of parallelism of a parallel computer in a multitask environment or a multithread environment. Next, it is possible to appropriately determine the degree of parallelism appropriate for the second cost, more specifically, the determination of the number of processors. Third, a bottleneck in the program structure can be specified. Fourth, it is possible to appropriately determine how to set the load on the system such as the configuration of programs to be executed simultaneously and the number of tasks to satisfy the required performance. Fifth, it is possible to predict detailed performance indicators such as throughput, response time, and resource usage in the system.

[0020]

DESCRIPTION OF THE PREFERRED EMBODIMENTS First, an outline of the present invention will be described. The performance prediction system of the multitasking system according to the present invention has a parallel degree of a parallel computer as a platform,
Parameters that represent specifications such as the processing speed of the processor, the types of programs that are processed simultaneously, their number,
A parameter indicating how threading is performed in each program, a parameter indicating how a semaphore and a critical section are adopted in each program are input parameters, and based on these parameters, Multiprocessor (server number) multiple servers (serv
er), and those servers process tasks or threads by processor sharing, and use semaphores and critical sections as tokens.
It is characterized by having a model generation unit that models the load on the system as a customer.

Since the queuing theory can be applied to the model generated by the model generation unit, an equation for explicitly predicting the degree of improvement of the processing performance and the performance index is obtained. The effect of being easily obtained is obtained. Furthermore, by using queuing theory as an analysis means, it is characterized in that required input parameters can be greatly reduced. As input parameters, detailed data such as the operation of the system becomes unnecessary, and an effect that preparation for prediction becomes easy is obtained.

As an analysis means, a simulation may be used instead of the queuing theory. In this case, the degree of improvement of the processing performance and the performance index can be predicted numerically. In the modeling, the resources of the platform may be modeled more finely. In this case, it is not generally possible to predict the degree of improvement in processing performance or performance index as an explicit expression using queuing theory, but it is necessary to accurately predict numerically by using an approximate calculation method or simulation. Is possible.

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. FIG. 1 is a configuration diagram of a first embodiment of a performance prediction system for a multitask system according to the present invention. Referring to FIG. 1, the performance prediction system includes a specification such as a parallelism of a parallel computer as a platform and a processing speed of a processor, types and numbers of programs to be simultaneously processed, and how threading is performed in each program. A system data input unit 11 for inputting data indicating whether or not a semaphore or a critical section is adopted;
A model generation unit 12 that generates a model by inputting information from the system data input unit 11 and calculates a performance index prediction value such as a throughput, a response, or a resource usage rate, which is a degree of improvement or a finer performance index, from the generated model And a prediction result output unit 14 that visually represents the value output from the prediction execution unit 13.

FIG. 2 is a flowchart showing the procedure of the operation of the performance prediction system. Next, the operation of the performance prediction system will be described with reference to FIG. First, the specifications of a parallel computer serving as a platform are input to the system data input unit 11 (S1). The specification here refers to the degree of parallelism of the parallel computer and the like. Next, data relating to the program is input to the system data input unit 11 (S2). The data about the program
(1) data indicating types of programs to be executed simultaneously and their numbers, (2) data indicating what tasks and threads each program is composed of,
(3) Data indicating how each program uses a semaphore or a critical section, and (4) Data indicating average use frequency and average use time of resources on the platform of each task and thread. A model is generated by the model generation unit 12 based on the input data (S3).

In the model generator 12 in FIG.
(1) The structure of the program, such as semaphores and critical sections, is modeled using tokens. (2) Programs and threads in the execution state are modeled as customers (in the queue model). (3) A parallel computer is modeled as a number of servers (as referred to in the queuing model) that are served by processor sharing and equal in parallelism, thereby creating a queuing network model.

The queuing network model is analyzed by the prediction execution unit 13 (S4). It is known in queuing theory that changes in the state of this model can be followed if the processing speed of each task or thread is known.
This processing speed can be calculated by performing simple modeling as described above. After calculating these processing speeds, the prediction execution unit 13 uses them to calculate the steady state distribution of the model state, the model throughput,
A performance prediction value such as a response time is calculated using queuing theory. Then, the calculated value is output to the prediction result output unit 1.
4 is received and displayed in a format that is easy for the user to understand using not only the numerical value but also a graph if specified (S
5).

For example, the following system is taken as an example. FIG. 3 is a timing chart showing an example of the execution timing of the program. It is assumed that one type of program is executed at the same time, and the number is always K (K is an integer of 2 or more). As shown in FIG. 3, each program is composed of a task 1 and a task 2.
The execution of the two tasks 1 and 2 is repeated. Task i (i is 1 or 2) is controlled by h (i) semaphores. That is, the number of programs that can execute the task i at the same time is h (i) in the entire system. or,
Regardless of the degree of parallelism of the parallel computer, it is assumed that the processing of task 1 uses a processor only on average μ (1) seconds / time, and the processing of task 2 uses a processor only on μ (2) seconds / time.
The degree of parallelism is n.

In such a system, as shown in the schematic explanatory diagram of the queuing network model in FIG. 4, (1) K customers 21 circulate in the network, and (2) customers 21 are of type i ( i secures the token of 1 or 2), receives the processing of task i at the service station 22, releases the token when the processing is completed, and (3) allocates the token to the FIFO.
(4) The customer 21 performs task 1 and task 2 alternately, (5) there are h (i) tokens of type i, and (6) the service station 22 shares the processor. It is modeled in the model generation unit 12 as a queuing network model that is composed of n (n is a positive integer) servers 23 that service customers 21 by a ring.

More specifically, “K customer 2
"1" refers to each of the programs 1 to K, and the customer 21 alternately repeats two types of processes, a process 1 performed after acquiring a type 1 token and a process 2 performed after acquiring a type 2 token. However, since the number of tokens of each type is limited, the customer 21 waits for a free space in the buffer 24 until a token can be obtained. Then, the customer 21 who has obtained the token proceeds to the service station 22,
The service is received from the n servers 23 under the discipline of processor sharing. And the service station 22
The customer 21 who has completed the processing 1 moves to obtain a type 2 token, and the customer 21 who has completed the processing 2 moves to obtain a type 1 token.

The queuing network model is analyzed by the prediction execution unit 13. If the number of customers executing task i is represented by X (i), the state of the model can be represented by X = (X (1), X (2)) (1). The speed of the change of the state X can be obtained as the processing speed of the task i (dependent on the state X). L (i) = min (h (i), X
(I)), in this case, the processing speed of each task in the state X is the processing speed τ (1 | X) = n * μ (1) L (1) / max ( n, L (1) + L (2)) (2) Processing speed τ (2 | X) = n * μ (2) L (2) / max (n, L (1) + L (2) of processing 2 ) (3) can be obtained. It is easy to calculate the steady-state distribution which becomes the state X if the speed of the state change of the model can be calculated, the throughput of the model, and the response time by using the queuing theory. For example, the throughput λ in this example is

[0031]

(Equation 1)

The calculation can be explicitly performed as follows. Other performance indicators, such as speedup, efficiency, response, and resource utilization, can also be explicitly predicted by applying queuing theory. The prediction execution unit 13 outputs a calculation result of the performance index of the model as a predicted value. The output is received by the prediction result output unit 14, and the numerical values are displayed in a format that is easy for the user to understand using a graph or the like.

Next, effects of the first embodiment will be described. Since the present invention models an actual system in a multitask environment or a multithread environment as a queuing network model capable of appropriately and simply performing a theoretical analysis, the present invention provides a parallel computer system in a multitask environment or a multithread environment. The performance index can be easily and accurately predicted. By appropriate modeling, various indices can be predicted as compared with the conventional method.

In the above embodiment, as mentioned above, a simulation may be used as the analysis means in the prediction execution unit 13 instead of the queue theory.
In the modeling, the resources of the platform may be modeled more finely as a server referred to in a queue. In this case, even if the queuing theory is used, it is not possible to obtain the degree of improvement in processing performance or the performance prediction index as an explicit expression, but it is necessary to similarly calculate numerically with high accuracy by using an approximate calculation method or simulation. Is possible.

An example in which the resources of the platform are more finely modeled will be described. FIG. 5 is a schematic explanatory diagram of a queuing network model in which the resources of the platform are further reduced. Note that the same components as those in the schematic explanatory diagram of FIG. 4 are denoted by the same reference numerals, and description thereof will be omitted. The systems to be predicted are the same as those described above, but the hard disks 32 and 34 exist as resources of the parallel computer serving as the platform, and the performance index is predicted by taking into account the competition between tasks and threads of these resources. Do.

Task 1 uses the hard disk 32,
Task 2 uses the hard disk 34. In this case, as shown in FIG. 5, a service station 31 corresponding to the hard disk 32 and a service station 33 corresponding to the hard disk 34 are added to the queuing network model shown in FIG. You only need to model it. In this case, the service station 31 and the service station 33 are both composed of one server and one buffer, and the service discipline is FIFO. However, it is necessary to appropriately perform resource modeling in a form that reflects its characteristics. is there. The analysis method for prediction is the same as described above.

Next, a second embodiment will be described. The basic configuration is the same as that of the first embodiment, but in order to specify an appropriate platform specification in the system, a parameter in a range in which the specification of the platform (specifically, the degree of parallelism, etc.) is to be predicted is specified. To the system data input unit 11. The model generation unit 12 performs modeling using the specifications of the platform as parameters. The prediction executing unit 13 changes the parameter within the given range, and calculates a predicted value by the same method as in the first embodiment. The prediction result output unit 14 receives the calculation result, and displays a prediction value for the parameter in a format that is easy for the user to understand. From this result, the user determines which parameter is appropriate. That is, the specifications of the parallel computer which is an appropriate platform are determined.

For example, consider the case where the throughput in the case where the parallel degree n is in the range of 1 to 16 is predicted for the system used in the first embodiment, and an appropriate parallel degree is calculated. The user creates a model in the same procedure as in the embodiment of the first embodiment, but inputs an input designating that the degree of parallelism is a parameter of 1 to n.
Do for The model generation unit 12 sets the degree of parallelism from 1 to n
Models are generated as the parameters up to and the prediction execution unit 13 calculates the predicted values of the performance indexes while changing the parameter values. Then, the data is displayed in a form that is easy for the user to understand. In this case, the throughput with respect to the parallelism n may be displayed as a graph as shown in FIG. Referring to FIG. 6, it can be seen that if the required performance is 4 per second, the degree of parallelism must be 5 or more.

Next, effects of the second embodiment will be described. According to the present invention, the specifications of the platform are used as parameters, and predicted values for the parameters can be viewed. For this reason, objective data (performance index) required to determine the specifications of the platform required to satisfy the performance specifications required by the system can be obtained as predicted values.

Next, a third embodiment will be described. Its basic configuration is the same as that of the first embodiment, but it relates to the program structure such as the number and configuration of semaphores and critical sections in a program, and the average use time and average use frequency of resources on each platform of tasks and threads. The data is provided to the system data input unit 11 as a parameter in a range in which prediction is to be performed.

The model generation unit 12 performs modeling using program data as parameters. Prediction execution unit 13
Calculates the predicted value in the same manner as in the first embodiment by changing the parameter within the given range. The prediction result output unit 14 receives the calculation result, and displays a prediction value for the parameter in a format that is easy for the user to understand. From this result, the user determines which parameter is appropriate. That is, an appropriate program structure is determined.

For example, for the system used in the first embodiment, the number of semaphores of task 1 is set to 1 (h (1)
= 1), the number of semaphores in task 2 is used as a parameter, and the throughput when the number of semaphores is changed from 1 to h is predicted to calculate the appropriate number of semaphores. The user creates a model in the same procedure as in the embodiment in the first embodiment, but inputs an input specifying that the number h (2) of semaphores is a parameter from 1 to h.
Perform for 1 The model generation unit 12 calculates the number of semaphores h
Models are generated using (2) as parameters from 1 to h, and the prediction execution unit 13 calculates performance prediction values of these models while changing parameter values. Then, the prediction result output unit 14 performs display in a form that is easy for the user to understand. Also in this case, it is preferable to display not only the numerical value of the throughput for h (2) but also graph data.

Next, the effects of the third embodiment will be described. According to the present invention, data relating to a program structure such as the number of semaphores in a program is used as a parameter, and a predicted value for the parameter can be viewed. Therefore, it is possible to obtain objective data necessary for determining how to configure the program structure to satisfy the required performance specifications of the system.

Next, a fourth embodiment will be described. Its basic configuration is the same as that of the first embodiment, but the system data input unit is used as a parameter in a range in which data relating to the load of the system such as the number of programs or tasks to be executed at the same time, the number of threads, etc. Give to 11. The model generation unit 12 performs modeling using the load of such a system as a parameter. The prediction executing unit 13 changes the parameter within the given range, and calculates a predicted value by the same method as in the first embodiment.
The prediction result output unit 14 receives the calculation result, and displays a prediction value for the parameter in a format that is easy for the user to understand. From this result, the user determines which parameter is appropriate. That is, an appropriate load is determined.

For example, the throughput when the number of programs is changed from 1 to K with respect to the system used in the first embodiment is estimated, and the appropriate number of simultaneously executing programs is calculated. think of. The user generates a model in the same procedure as in the example of the first embodiment, but inputs an input to the system data input unit 11 specifying the number of programs as a parameter from 1 to K.
The model generation unit 12 generates models using the number k of programs as a parameter from 1 to K, and the prediction execution unit 13 calculates the performance prediction values of these models while changing the parameter values. Then, the prediction result output unit 14 receives the calculation result, and displays it in a form that is easy for the user to understand.

Next, effects of the fourth embodiment will be described. According to the present invention, a load in the system such as the number of programs in a program is used as a parameter, and a predicted value for the parameter can be viewed. Therefore, it is possible to obtain objective data necessary for determining how to load the system in order to satisfy the required performance specifications of the system.

Next, a fifth embodiment will be described. The fifth embodiment relates to a recording medium on which a performance prediction method program is recorded. FIG. 7 is a configuration diagram of a recording medium and a recording medium driving device. Referring to FIG. 7, the recording medium driving device includes a CPU (central processing unit) 41, an input unit 42, a storage unit 43, and a performance prediction system 44. Drive.

The performance prediction system 44 comprises a system data input unit 11, a model generation unit 12, a prediction execution unit 13, and a prediction result output unit 14 shown in FIG.
The recording medium 45 stores a performance prediction method program shown in the flowchart of FIG. 2 in advance.

Next, the operation of the driving device will be described. First, when a load (LOAD) instruction for a performance prediction method program is input to the CPU 41 via the input unit 42, the CPU 41 reads the performance prediction method program from the recording medium 45, and writes the read program in the storage unit 43. Put in. Next, when a run (RUN) instruction of the performance prediction method program is input to the CPU 41 via the input unit 42, the CPU 41 reads the performance prediction method program from the storage unit 43, and uses the read program to execute the performance prediction system 44. Control. Since the contents of the control have been described above, the description is omitted.

[0050]

According to a first aspect of the present invention, there is provided a performance prediction system for a multitasking system using a parallel computer as a platform. The system predicts resource competition in a multitasking environment by using a queuing network model. First, since it includes a modeling means for modeling using the above, and a performance index prediction means for analyzing the competition of the modeled resources by using queuing theory and predicting the performance index of the multitask system, In the multitask environment or the multithread environment, it is possible to correctly predict the performance improvement degree and the performance index with respect to the parallelism of the parallel computer. Next, it is possible to appropriately determine the degree of parallelism appropriate for the second cost, more specifically, the determination of the number of processors. Third, a bottleneck in the program structure can be specified. Fourth, it is possible to appropriately determine how to set the load on the system such as the configuration of programs to be executed simultaneously and the number of tasks to satisfy the required performance. Fifth, it is possible to predict detailed performance indicators such as throughput, response time, and resource usage in the system.

According to a second aspect of the present invention, there is provided a method for predicting the performance of a multitask system using a parallel computer as a platform. And a second step of analyzing the competition of the modeled resources using queuing theory and predicting a performance index of the multitask system. Has the effect of

According to a third aspect of the present invention, there is provided a recording medium recording a program for predicting the performance of a multitasking system using a parallel computer as a platform, the recording medium comprising a resource in a multitasking environment. And a second step of analyzing the competition of the modeled resources using queuing theory and predicting the performance index of the multitask system. Since the program including the program is recorded, the same effect as that of the first invention can be obtained.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a first embodiment of a performance prediction system of a multitask system according to the present invention.

FIG. 2 is a flowchart illustrating a procedure of an operation of the performance prediction system.

FIG. 3 is a timing chart showing an example of a program execution timing.

FIG. 4 is a schematic explanatory diagram of a queuing network model.

FIG. 5 is a schematic explanatory diagram of a queuing network model in which the resources of the platform are further refined.

FIG. 6 is a plot of throughput versus parallelism.

FIG. 7 is a configuration diagram of a recording medium and a recording medium driving device.

FIG. 8 is a configuration diagram illustrating an example of a computer system in a multitask environment using a parallel computer as a platform.

[Explanation of symbols]

 Reference Signs List 11 system data input unit 12 model generation unit 13 prediction execution unit 14 prediction result output unit 45 recording medium

Claims (13)

[Claims] 1. A performance prediction system for a multitasking system having a parallel computer as a platform, comprising: a modeling means for modeling contention of resources in a multitasking environment using a queuing network model; And a performance index predicting means for analyzing the contention of the resources using queuing theory and predicting the performance index of the multitask system. 2. The method according to claim 1, wherein the modeling unit includes a data input unit to which the specifications of the parallel computer are input, and a model generation unit that generates a model using the data input to the data input unit. The performance prediction system according to claim 1, wherein 3. The performance index predicting unit analyzes a competition of resources modeled by the modeling unit and predicts a performance index, and a prediction result output by the prediction executing unit. The performance prediction system according to claim 1, further comprising an output unit. 4. The apparatus according to claim 1, wherein the modeling unit performs modeling using a platform specification as a parameter, and the performance index prediction unit predicts a performance index for the parameter. Performance prediction system. 5. The method according to claim 1, wherein the modeling unit performs modeling using a structure of a program to be executed including a semaphore and a critical section as a parameter, and the performance index prediction unit predicts a performance index for the parameter. The performance prediction system according to any one of claims 1 to 3. 6. The modeling means performs modeling using a system load such as the number of programs or tasks executed simultaneously and the number of threads as a parameter, and the performance index prediction means predicts a performance index for the parameter. The performance prediction system according to any one of claims 1 to 3, wherein 7. The performance index predicting means calculates the throughput of the queuing network model as a performance prediction value by following a change in the state of the queuing network model from the processing speed of each task or thread. The performance prediction system according to claim 1. 8. A method for predicting the performance of a multitasking system using a parallel computer as a platform, comprising: a first step of modeling resource contention in a multitasking environment using a queuing network model; Analyzing the contention of the resources using queuing theory and predicting the performance index of the multitasking system. 9. The first step includes an eleventh step of inputting the specifications of the parallel computer to a data input unit and a twelfth step of generating a model using the data input to the data input unit. 9. The performance prediction method according to claim 8, wherein: 10. The second step is a twenty-first step of analyzing contention of resources modeled in the first step and predicting a performance index, and a twenty-second step of outputting a prediction result in the twenty-first step. 10. The performance prediction method according to claim 8, further comprising: 11. A recording medium recording a program for predicting the performance of a multitask system using a parallel computer as a platform, wherein a first step of modeling contention of resources in a multitask environment using a queuing network model. And a second step of analyzing the modeled resource conflict using queuing theory and predicting the performance index of the multitasking system. 12. The first step is an eleventh step of inputting specifications of the parallel computer to a data input unit;
A twelfth step of generating a model using the data input to the data input unit. 13. The second step is a twenty-first step of analyzing contention of resources modeled in the first step and predicting a performance index, and a twenty-second step of outputting a prediction result in the twenty-first step. 13. The recording medium according to claim 11, comprising: