JP2008165412A - Performance-computing device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To adequately compute a performance value of a device even if the performance value varies due to a measurement error or a difference between processing parameters. <P>SOLUTION: A sensitivity analysis part 2 determines totaling sections based on measurement data and generates totaling section data. An evaluation value computation part 4 computes evaluation function values based on performance parameters and the totaling section data. The sensitivity analysis part 2 computes the sensitivity of the evaluation function values to the performance parameters. The sensitivity analysis part 2 computes the vector length of each sensitivity vector. The sensitivity analysis part 2 determines that the parameters are to be corrected if the vector length S is a sensitivity threshold value or larger. A parameter updating part 5 updates the performance parameters. The sensitivity and the evaluation function values are re-computed according to the updated performance parameters. The sensitivity analysis part 2 determines that correcting the parameters is to be finished if the vector length is less than the sensitivity threshold value and outputs the performance parameters. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a performance calculation device for a computer system such as an Internet system or an intranet system.

Conventionally, in a computer system, the performance tends to gradually deteriorate as the amount of processing increases due to operator learning and the increase in the number of users, the amount of data increases, and the cache miss rate increases accordingly. Therefore, as disclosed in, for example, Patent Document 1 and Patent Document 2, an apparatus for monitoring the performance of the system is used.
JP 2002-99448 A JP 2002-268922 A

  In order to prevent system performance problems in advance, it is necessary to monitor the margin and processing amount of the system, catch the signs of performance degradation, and respond in advance.

  For each type of processing, if the resource usage amount when the processing is executed once can be estimated, the resource usage rate can be estimated according to the rate at which each processing amount is executed at the peak processing amount. it can. By periodically estimating the resource utilization rate, measures such as resource enhancement can be implemented before performance degradation, and performance degradation can be prevented in advance.

  By the way, in a computer system in operation, it is difficult to stop the operation and measure the resource usage when one process is executed. Therefore, it is necessary to estimate the resource usage when one process is executed using data that can be measured during operation. In an operating system, a plurality of types of processes are executed simultaneously, and the number of processes executed for each type and the resource utilization rate can be measured within a certain period of time.

  If the resource usage when one process is executed is always a constant value, the resource usage for each process can be calculated from data that can be measured during operation by solving simultaneous equations. However, since the resource usage fluctuates due to a measurement error or a difference in parameters of processing, there is a problem that cannot be solved by simultaneous equations.

  Accordingly, an object of the present invention is to provide a performance calculation device that can calculate an appropriate performance value even if the performance value fluctuates due to a measurement error or a difference in processing parameters.

  That is, the performance calculation device according to the present invention includes measurement data input means for inputting measurement data of the operation of the target device, and performance parameter input means for inputting a performance parameter that is a parameter of the performance model. The evaluation function value is calculated based on the performance parameter input by the performance parameter input means, the sensitivity calculation means for calculating the sensitivity of the evaluation function value with respect to the performance parameter, and the evaluation function value changes according to a predetermined condition. As described above, the performance parameter is updated according to the calculated sensitivity, and the updated performance parameter is output when a predetermined condition is satisfied.

  According to the present invention, even if the performance value fluctuates due to a measurement error or a difference in processing parameters, the performance value can be appropriately calculated.

Embodiments of the present invention will be described below with reference to the drawings.
(First embodiment)
First, a first embodiment of the present invention will be described.
FIG. 1 is a block diagram showing an example of the configuration of a performance calculation device according to an embodiment of the present invention.

  The performance calculation device according to the first embodiment of the present invention has a function of identifying a performance parameter. The performance calculation device includes a measurement data input unit 1 as a measurement data input unit, a sensitivity analysis unit 2 as a sensitivity calculation unit, a parameter input / output unit 3 as a performance parameter input unit and an output unit, and an evaluation value calculation unit. A value calculation unit 4, a parameter update unit 5 that is update means, and a storage device 6 are provided, and each is connected via a bus 7.

  The measurement data input unit 1 receives performance data such as the resource usage rate and the processing amount of the CPU and disk of the computer mounted on the target system from the outside, and transmits this data to the sensitivity analysis unit 2 as measurement data.

The sensitivity analysis unit 2 receives the measurement data from the measurement data input unit 1, determines a total section based on the measurement data, and generates total section data.
Further, the sensitivity analysis unit 2 receives the sensitivity threshold value Sa, the update coefficient ε, and the performance parameter from the parameter input / output unit 3, and transmits the received performance parameter and aggregated interval data to the evaluation value calculation unit 4. The evaluation value is received from 4.

  Further, the sensitivity analysis unit 2 calculates the sensitivity of the evaluation value with respect to the performance parameter, transmits the performance parameter, the sensitivity, and the update coefficient ε to the parameter update unit 5, receives the updated performance parameter from the parameter update unit 5, The performance parameter is transmitted to the parameter input / output unit 3.

The total section is a time section obtained by dividing the time zone in which the measurement data is measured, and is referred to as the total section because the processing amount totaled for each time section becomes the input value of the performance model.
The total section data is the average value of the resource utilization rate measured in the total section and the amount of processing generated for each type.

The sensitivity threshold value Sa is a threshold value used for determining whether or not to correct the performance parameter. Specifically, the sensitivity vector length of the evaluation function value for each performance parameter is equal to or greater than the sensitivity threshold value Sa. If so, the correction is repeated, and the correction ends when the vector length is less than the sensitivity threshold value Sa.
The update coefficient ε is a coefficient used for updating the performance parameter, and a value obtained by multiplying the update coefficient ε by the sensitivity is added to the performance parameter and updated.

  The parameter input / output unit 3 receives a performance parameter from the outside, transmits the parameter to the sensitivity analysis unit 2, and receives the performance parameter from the sensitivity analysis unit 2, transmits the performance parameter to the outside, and receives a sensitivity threshold value from the outside. When Sa and the update coefficient ε are received, the sensitivity threshold value Sa and the update coefficient ε are transmitted to the sensitivity analysis unit 2.

When the evaluation value calculation unit 4 receives the performance parameter and the total interval data from the sensitivity analysis unit 2, the evaluation value calculation unit 4 calculates an evaluation value and transmits the calculated evaluation value to the sensitivity analysis unit 2.
When the parameter update unit 5 receives the performance parameter, the sensitivity, and the update coefficient ε from the sensitivity analysis unit 2, the parameter update unit 5 adds a value obtained by multiplying the performance parameter by the sensitivity and the update coefficient ε, and transmits the result to the sensitivity analysis unit 2.
The storage device 6 is a storage device configured by hardware such as a hard disk drive or a nonvolatile memory device, and stores an execution program for each unit.

Next, the operation of the performance calculation device configured as described above will be described. FIG. 2 is a flowchart showing an example of the processing operation of the performance calculation device according to the first embodiment of the present invention.
First, the measurement data input unit 1 receives measurement data from the outside (step S1), and transmits this measurement data to the sensitivity analysis unit 2.
The parameter input / output unit 3 receives performance parameters from the outside (step S2), and transmits the performance parameters to the sensitivity analysis unit 2.

The sensitivity analysis unit 2 determines a total section based on the measurement data received from the measurement data input unit 1, and generates total section data (step S3).
The parameter input / output unit 3 receives the sensitivity threshold value Sa and the update coefficient ε from the outside (step S4), and transmits these pieces of information to the sensitivity analysis unit 2.

  The sensitivity analysis unit 2 transmits the performance parameter and the total interval data to the evaluation value calculation unit 4.

The evaluation value calculation unit 4 calculates the evaluation function value E based on the performance parameter and the total interval data. The evaluation value calculation unit 4 transmits the calculated evaluation function value E to the sensitivity analysis unit 2. The sensitivity analysis unit 2 determines the initial value of i of all performance parameters ui (i = 1, 2,..., M). Is set to 1 (step S5), and the sensitivity Si of the evaluation function value E with respect to the performance parameter ui is calculated as a derivative of the evaluation function value E (step S6). M is the number of parameters.
If the value of i is not M (NO in step S7), the sensitivity analysis unit 2 updates the value of i by 1 (step S7), and then calculates the sensitivity Si again (steps S8 → S6). ).

  When the value of i is M (YES in step S7), the sensitivity analysis unit 2 determines that the sensitivity Si has been calculated for all parameters, and the vector length S of the sensitivity vector is expressed by the following equation (1). Therefore, it calculates (step S9).

  If the vector length S of the sensitivity vector is equal to or greater than the sensitivity threshold value Sa (YES in step S10), the sensitivity analysis unit 2 determines that the parameter should be corrected, and the performance parameter ui (i = 1, 2,..., M), sensitivity Si (i = 1, 2,..., M) and update coefficient ε are transmitted to the parameter update unit 5.

  The parameter updating unit 5 sets the initial value of i to 1 (step S11), and the performance parameter ui (i = 1) based on the sensitivity Si (i = 1, 2,..., M) and the update coefficient ε. , 2,..., M) are updated (step S12).

If the value of i is not M (NO in step S13), the parameter updating unit 5 updates the value of i by 1 (step S14) and then updates the performance parameter ui again (step S13 → S12).
If the value of i is M (YES in step S13), the parameter update unit 5 determines that the necessary update of the performance parameter ui has been completed, and returns to the process of step S5.

If the vector length S of the sensitivity vector is less than the sensitivity threshold value Sa (NO in step S10), the sensitivity analysis unit 2 determines that the parameter correction should be terminated, and the performance parameter ui (i = 1, 2,... , M) is transmitted to the parameter input / output unit 3 (step S15).
Further, instead of determining whether or not the correction is necessary based on the sensitivity threshold value Sa, the sensitivity analysis unit 2 may set an upper limit value for the number of corrections and determine that the correction is terminated when the number reaches the number of corrections.

  As described above, the performance calculation device according to the first embodiment of the present invention uses the performance data including the processing amount and the resource usage rate measured in the target system in operation, and the evaluation value is obtained by sensitivity analysis. By repeatedly correcting the performance parameter value so as to be optimized, the performance parameter can be identified.

(Second Embodiment)
Next, a second embodiment of the present invention will be described. Note that, in the configuration of the performance calculation device according to each of the following embodiments, the same partial description as that shown in FIG. 1 is omitted.
In addition to the functions described in the first embodiment, the sensitivity analysis unit 2 increases the performance parameter by a minute amount and transmits the performance parameter to the evaluation value calculation unit 4, and the evaluation value related to the increased performance parameter from the evaluation value calculation unit 4 And a function for calculating the sensitivity from the difference between this evaluation value and the evaluation value before the performance parameter is changed.

Next, processing of this performance calculation device will be described. FIG. 3 is a flowchart showing an example of the processing operation of the performance calculation device according to the second embodiment of the present invention.
This performance calculation device performs the processing from steps S1 to S4 described in the first embodiment.

  The sensitivity analysis unit 2 transmits the performance parameter and the total interval data to the evaluation value calculation unit 4.

The evaluation value calculation unit 4 calculates the evaluation function value E based on the performance parameter and the total interval data, and transmits the calculated evaluation function value E to the sensitivity analysis unit 2 (step S21).

The sensitivity analysis unit 2 increases the parameter value of one performance parameter ui by a minute amount Δu (step S22), and transmits the performance parameter and the total interval data to the evaluation value calculation unit 4.
The evaluation value calculation unit 4 calculates the evaluation function value Ei related to the increased performance parameter based on the performance parameter value increased by the minute amount Δu and the total interval data (step S23). The evaluation value calculation unit 4 transmits the calculated evaluation function value Ei to the sensitivity analysis unit 2.
The sensitivity analysis unit 2 calculates the sensitivity Si using the evaluation function values before and after increasing the parameter value of the performance parameter by the minute amount Δu (step S24).

Next, the sensitivity analysis unit 2 calculates the original value by subtracting the minute amount Δu from the value of the performance parameter in which the minute amount Δu is increased (step S25).
If the value of i is not M (NO in step S7), the sensitivity analysis unit 2 updates the value of i by 1 (step S8) and then returns to the process of step S22 (step S8 → S22). ).

  When it is determined “YES” in the process of step S7, the performance calculation apparatus proceeds to the process of step S9 described in the first embodiment. Subsequent processing is the same as in the first embodiment.

  As described above, the performance calculation device according to the second embodiment of the present invention recalculates the evaluation function value by changing the performance parameter by a minute amount even when the derivative of the evaluation function value with respect to the performance parameter cannot be calculated. Then, the performance parameter can be identified by obtaining the sensitivity Si from the difference between the evaluation function values before and after the minute change of the performance parameter, and repeating the correction of the performance parameter value so that the evaluation function value becomes optimal.

(Third embodiment)
Next, a third embodiment of the present invention will be described. FIG. 4 is a block diagram showing an example of the configuration of the performance calculation device according to the third embodiment of the present invention.
The performance calculation device according to the present embodiment has a performance parameter identification function and a performance monitoring function. In addition to the configuration described in the first embodiment, this apparatus further includes a performance prediction unit 11 that is an estimation unit and a performance prediction unit, and a notification unit 12 that is a notification unit, and each is connected to the bus 7.

In the present embodiment, the measurement data input unit 1 further has a function of transmitting measurement data to the performance prediction unit 11.
Further, when a performance parameter is input from the outside, the parameter input / output unit 3 transmits the performance parameter to the sensitivity analysis unit 2 and the performance prediction unit 11, and receives the performance parameter from the sensitivity analysis unit 2, and then receives the received performance parameter. Is output to the performance prediction unit 11 and the outside.

The performance prediction unit 11 receives the measurement data from the measurement data input unit 1, estimates the peak processing amount based on the measurement data, receives the performance parameter from the parameter input / output unit 3, and receives the received performance parameter. Based on the peak processing amount, the system performance at the time of peak processing is predicted, and this information is transmitted to the notification unit 12.
When the notification unit 12 receives information on the predicted system performance from the performance prediction unit 11, the notification unit 12 notifies the administrator by means of screen display or mail transmission.

Next, processing of the performance calculation device configured as described above will be described. FIG. 5 is a flowchart showing an example of the processing operation of the performance calculation device according to the third embodiment of the present invention.
The performance calculation device according to this embodiment performs the processing from step S1 to step S15 described in the second embodiment.

  When the performance parameter is output by the process of step S15, the performance prediction unit 11 receives the performance parameter from the parameter input / output unit 3. The performance prediction unit 11 estimates the peak processing amount based on the received measurement data. Based on the received performance parameter and the calculated peak processing amount, the performance prediction unit 11 predicts the system performance at the peak processing amount (step S31), and transmits this information to the notification unit 12.

When the notification unit 12 receives information on the peak system performance from the performance prediction unit 11, the notification unit 12 notifies the administrator by means of screen display or mail transmission (step S32).
Thereafter, the performance calculation device periodically repeats the processing from steps S1 to S32 shown in FIG. 5 to monitor the performance of the target system.

(Fourth embodiment)
Next, a fourth embodiment of the present invention will be described. FIG. 6 is a block diagram showing an example of the configuration of the performance calculation device according to the fourth embodiment of the present invention.
The performance calculation device according to the present embodiment has a performance parameter identification function and a performance diagnosis function. In addition to the configuration described in the first embodiment, the performance determination unit 21 that is a performance determination unit and a notification that is a notification unit Each of the units 12 is further connected to the bus 7.
When the parameter input / output unit 3 receives a performance parameter from the sensitivity analysis unit 2, the parameter input / output unit 3 further has a function of transmitting the received performance parameter to the performance determination unit 21.

Whether the performance determination unit 21 receives the performance parameter from the parameter input / output unit 3, compares the received performance parameter with the performance parameter received in the past, and transmits a notification signal to the notification unit 12 based on the result. It determines whether or not, and transmits a signal to the notification unit 12.
When the notification unit 12 receives a signal from the performance determination unit 21, the notification unit 12 notifies the administrator by means of screen display or mail transmission.

Next, processing of the performance calculation device configured as described above will be described. FIG. 7 is a flowchart showing an example of the processing operation of the performance calculation device according to the fourth embodiment of the present invention.
The performance calculation device according to this embodiment performs the processing from step S1 to step S15 described in the second embodiment.

  When the performance parameter is output by the process of step S15, the performance determination unit 21 receives the performance parameter from the parameter input / output unit 3, and compares the received performance parameter with the performance parameter received in the past. . The performance determination unit 21 determines whether or not to transmit a notification signal to the notification unit 12 based on the comparison result (step S41).

  Specifically, for example, if the absolute value of the change in the performance parameter is larger than a preset threshold value, the performance determination unit 21 transmits the process name as a signal to the notification unit 12, and if the absolute value is less than the threshold value The signal is not transmitted to the notification unit 12.

When the notification unit 12 receives the notification signal from the performance determination unit 21, the notification unit 12 notifies the administrator of the process name by means of screen display or mail transmission (step S42).
Thereafter, this performance calculation device periodically repeats the processing from steps S1 to S42 shown in FIG. 7 to perform performance diagnosis of the target system.

  Examples of each embodiment will be described below.

  Example 1 will be described first. Example 1 is an example of the first embodiment described above, and describes a performance model parameter identification method.

  The target system described in this embodiment includes a server group such as a Web server, an application server, and a database server, and a client. The following information is measured as measurement data from these servers and clients and output to the measurement data input unit 1.

  The measurement data is resource utilization rate and processing amount information. The resource usage rate is a usage rate of resources such as a CPU and a disk of the server machine, and is measured at regular time intervals such as 10 seconds. The server group can be measured and output by a program tool such as a system monitor program performance monitor (perfmon.exe) of Microsoft Corporation.

  As for the processing amount information, a usage history is obtained by a log function of a Web server or an application server. There are the following two types of usage history. Each of these indicates the amount of processing given to the system at regular intervals.

  The first usage history is a record of the type of processing started in response to a request from the client and the occurrence time, like an access log. In the second usage history, the number of occurrences for each process in the section is recorded at regular intervals.

As described in the first embodiment, when the measurement data is input, the measurement data input unit 1 transmits the measurement data to the sensitivity analysis unit 2. When the sensitivity analysis unit 2 receives the measurement data, the sensitivity analysis unit 2 generates aggregate section data. The aggregate section data is data on the resource usage amount and the processing amount in the aggregate section described below. Hereinafter, a method for generating the total interval data will be described.
The sensitivity analysis unit 2 first determines the total section in the generation of the total section data, and then calculates the resource usage amount and the processing amount in each total section.

  FIG. 8 is a diagram illustrating a first cut-out example of the aggregation section when the number of processes is relatively small. FIG. 9 is a diagram showing the calculation result of the resource usage amount in each section in a table format, and FIG. 10 is a diagram showing the calculation result as a time series graph. FIG. 11 is a diagram illustrating a second cut-out example of the total section.

As shown in FIG. 8, if the total section is obtained simply in accordance with the resource utilization rate measurement cycle, the process may straddle the boundary of the total section and cause an error.
Therefore, first, the sensitivity analysis unit 2 obtains a measurement section in which the resource usage rate is 0, that is, the noise level, among the sections illustrated in FIGS. 9 and 10.

  Secondly, as shown in FIG. 11, the sensitivity analysis unit 2 uses a section sandwiched between measurement sections with a resource utilization rate of 0 as a total section, and obtains the number of each process in each total section from the processing amount information.

  Thirdly, the sensitivity analysis unit 2 obtains the resource usage amount in each aggregation section from the resource usage rate information.

  FIG. 12 shows an example of the result of the number of processes and the resource utilization rate for each aggregation section obtained according to these procedures. Here, the resource is described as one resource, but there are actually a plurality of resources.

In addition, when the number of occurrences and the resource usage rate for each process in the section are recorded at regular intervals, the period of the usage history and the common multiple of the cycle of the resource usage rate may be used as the period of the totaling section.
The sensitivity analysis unit 2 identifies the performance parameter based on the total interval data, and transmits the identified performance parameter to the parameter input / output unit 3.

  The performance model 101 included in the sensitivity analysis unit 2 will be described. The performance model 101 is a model that predicts a resource usage amount when the processing amount is processed from an assumed value of the processing amount, and includes a processing resource usage model 102.

  In the performance model of the present embodiment, the processing resource usage is expressed as a fixed value. FIG. 13 is a diagram illustrating an example of the performance model of the present embodiment. When the number of processes for each process is input, the performance model 101 performs calculation using the process resource usage model 102 and outputs the resource usage. Although FIG. 13 shows an example of calculating three types of resources based on three types of processing, the number of processing types and the number of resource types may be different.

  The output calculation formula of the performance model is shown by the following formula (2).

  In each embodiment, it is assumed that the following expressions (3), (4), and (5) are satisfied.

x _ji is the usage amount of the resource j by the process i and is a performance parameter. n _ik is the number of processes i of the aggregate section data k, and y _jk is an estimated value of the usage amount of the resource j of the aggregate section data k. In Expression (2), T is the number of processing types, U is the number of resources, and N is the number of total sections.

Next, the correction of the evaluation function and the performance parameter of this embodiment will be described.
Here, the usage amount x _ji of the resource j by the process i, which is a performance parameter, is expressed by the performance parameter u _{(j−1) T + i} as shown in the following formula (6).

  In this embodiment, the evaluation function value E is defined as a value obtained by dividing the output value of the performance model and the error square sum of the resource usage of the aggregated section data by 2, as shown in the following equation (7).

  In each embodiment, it is assumed that the following equation (8) holds.

R _jk is a measured value of the usage amount of the resource j of the aggregate section data k.
The parameter update unit 5 corrects the performance parameter so that the evaluation function approaches zero. Here, as shown in FIG. 14, a method of correcting the performance parameter in the direction of steepest descent in the manner of the hill-climbing method will be described.

  First, the sensitivity of the evaluation function with respect to the performance parameter is calculated. The formula for calculating sensitivity is shown in the following formula (9).

The reverse direction of the vectors (s ₁ , s ₂ ,..., S _TU ) in which the sensitivities are arranged is the steepest descent direction.
The sensitivity analysis unit 2 changes the performance parameter by a minute amount in the steepest descent direction according to the following equations (10) and (11).
u _l = u _i −εs _l (l = 1, 2,..., TU) (10)

ε is an update coefficient and is a small positive value.
The performance calculation device can correct the performance parameter so that the evaluation function approaches 0 by repeating such sensitivity calculation and correction of the performance parameter.

Next, experimental results of parameter identification are shown.
Here, an experiment was conducted on one resource. There are three types of processing: Trans1, Trans2, and Trans3. FIG. 15 shows 40 total section data virtually created.
FIG. 16 shows the average value of the resource usage of each process that is the basis of the 40 total section data.

Next, FIG. 17 shows an initial value of the resource usage that is a performance parameter. Here, the number of corrections was fixed to 100 and parameter correction was performed. The results are shown in the graph.
FIG. 18 is a graph showing transition of evaluation function values, where the horizontal axis represents the number of corrections and the vertical axis represents the evaluation function value.

FIG. 19 is a graph showing the transition of the average value parameter of the process Trans1. FIG. 20 is a graph showing the transition of the average value parameter of the process Trans2. FIG. 21 is a graph showing the transition of the average value parameter of the process Trans3.
FIG. 22 is a table showing the performance parameter values of the experimental results. As a result, it can be seen that the corrected value approaches the value shown in FIG.

  Next, Example 2 will be described. Example 2 is an example of the above-described second embodiment, and describes a performance model parameter identification method.

  In the second embodiment, the resource usage model of the performance model expresses the usage amount as a probability distribution, and identifies a parameter using the occurrence probability of the resource usage amount of the aggregated section data as an evaluation function.

The target system and total section data described in this embodiment are the same as those in the first embodiment.
The performance model 111 of the present embodiment will be described.
The resource usage amount of the process may be a fixed value, but in the case of search processing, for example, the usage amount varies depending on the search condition. In this embodiment, in order to express such a phenomenon, the resource usage model of the performance model handles the usage as having a certain probability distribution.

FIG. 23 is a diagram illustrating an example of a performance model. When the processing number (n ₁ , n ₂ , n ₃ ) of each process is input, the performance model 111 performs a calculation using the resource usage model 112 of the process and outputs a probability distribution of the resource usage. This calculation formula is shown in the following formulas (12) and (13).

  In each embodiment, it is assumed that the following expressions (14), (15), and (16) hold.

μ _ji is an average value of a normal distribution representing the resource usage of processing. σ _ji is a standard deviation value of a normal distribution representing the resource usage of processing. p _ji (x) is a probability density function of a normal distribution of the usage amount of the resource j by the processing i, and represents the probability that the usage amount of the resource j by the processing i is x.

Further, p ^j (R) in the equation (12) is a probability density function of the usage amount of the resource j, and represents the probability that the usage amount of the resource J is R. This equation is for a case where three types of processing are included, and as the number of types of processing increases, the multiplicity of integration increases.

Next, a performance parameter identification method will be described.
In the present embodiment, the average value μ _ji and the standard deviation value σ _ji of the normal distribution representing the resource usage of processing will be described as performance parameters.

Here, the average value μ _ji and the standard deviation value σ _ji of the normal distribution are expressed by the performance parameter u _i as shown in the following equations (17) and (18).

  The idea of maximum likelihood method is applied to the identification of performance parameters. That is, the parameter update unit 5 corrects the parameters of the performance model so that the probability that the resource usage will be the measured value in each aggregation section is maximized.

The parameter updating unit 5 obtains a probability P _jk (R _jk ) that the usage amount of the resource j in the totaling section k becomes R _jk that is a measured value, and the performance parameter is set so that the probability becomes maximum in all the totaling sections. To correct.

Regarding the setting of the evaluation value for maximizing the probability P _jk (R _jk ) in all the aggregate sections, a plurality of candidates are conceivable, for example, an arithmetic mean or a geometric mean. The arithmetic average is expressed by the following formula (19), and the geometric average is expressed by the following formula (20).

In the present embodiment, the subsequent processing will be described on the assumption that the geometric average is used.
FIG. 24 is a diagram illustrating the concept of parameter correction by evaluation value optimization. As shown in FIG. 24, the algorithm of the hill-climbing method is applied, and the combination of performance parameters that gives the maximum likelihood is obtained.

Next, a specific performance parameter correction method will be described.
As described above, the evaluation function is the geometric mean of the occurrence probabilities of the resource usage of the aggregate interval data calculated by the performance model. This calculation formula is shown in the following formula (21).

The parameter update unit 5 corrects the performance parameter so that this evaluation function is maximized. The sensitivity analysis unit 2 calculates the sensitivity of the evaluation function for all performance parameters u _l = (l = 1, 2,..., 2TU).
Specifically, the sensitivity analysis unit 2 first adds a minute amount Δu to the performance parameter u _{1 according} to the following equation (22).

u _l = u _l + Δu (22)
The evaluation value calculation unit 4 calculates an evaluation function value using the performance parameter. The evaluation function value at this time is defined as _El . Then evaluation value calculating section 4 calculates a sensitivity of the evaluation function value for performance parameters u _l. Specifically, as shown in the following formula (23), the evaluation value calculation unit 4 calculates the difference between the evaluation function value E _l after the minute amount addition and the evaluation function value E before the minute amount addition by the minute amount Δu. The sensitivity s _{l of the} evaluation function with respect to the performance parameter is calculated by dividing by.
s ₁ = (E ₁ −E) / Δu (23)
A vector (s ₁ , s ₂ ,..., S _2TU ) in which the sensitivities are arranged becomes the steepest gradient direction.

The parameter update unit 5 changes the performance parameter by a minute amount in the steepest gradient direction according to the equations (24), (25), and (26).
u ₁ = u ₁ −εs ₁ (1 = 1, 2,..., TU) (24)

ε is an update coefficient and is a small positive value.
That is, the performance calculation device according to the second embodiment repeats the performance parameter sensitivity calculation and update until the designated number of times or, for example, the sum of squares of the sensitivity is equal to or less than a predetermined value.

Next, experimental results of parameter identification are shown.
Here, the experiment was performed under the same conditions as in the experiment of Example 1. The total section data is the same as that in the first embodiment (FIG. 15). FIG. 25 shows the average value of the resource usage amount of each process and the true value of the standard deviation value for the total section data. In addition, FIG. 26 shows initial values of the average value and standard deviation value of the resource usage of each process, which are performance parameters.

Here, the number of corrections was fixed to 200, and parameter correction was performed.
FIG. 27 is a graph showing changes in evaluation function values according to parameter correction, where the horizontal axis represents the number of corrections and the vertical axis represents the evaluation function value.

FIG. 28 is a graph showing the parameter transition of the process Trans1. FIG. 29 is a graph showing the transition of parameters of the process Trans2. FIG. 30 is a graph showing the parameter transition of the process Trans3.
FIG. 31 is a table showing the performance parameter values of the experimental results. As a result, it can be seen that the corrected value approaches the true value shown in FIG.

Next, an experimental result in the case where the arithmetic average of the occurrence probabilities P _jk (R _jk ) of the total interval data is used as an evaluation function is shown.
The aggregate section data and performance parameter initial values are the same as in the previous experiment. Here, parameter correction was performed 300 times.

  FIG. 32 is a graph showing changes in evaluation function values according to parameter correction, where the horizontal axis represents the number of corrections and the vertical axis represents the evaluation function value. FIG. 33 is a graph showing the parameter transition of the process Trans1. FIG. 34 is a graph showing the transition of parameters of the process Trans2. FIG. 35 is a graph showing the parameter transition of the process Trans3.

  FIG. 36 is a table showing performance parameter values as experimental results. According to this, it can be seen that the standard deviation values of Trans1 and Trans3 are closer to 0 than the true value. This is because the arithmetic average of the occurrence probabilities of the aggregate interval data is adopted as the evaluation function.

  In the case of arithmetic mean, if the probability density function is infinite for one aggregated section data, the evaluation value is infinite even if the occurrence probability of the other aggregated section data is zero. As the standard deviation value approaches 0, the maximum value of the probability density function approaches infinity. As a result, the standard deviation value approaches 0 by modifying the performance parameter so as to increase the evaluation function value. Will end up.

  Therefore, in order to accurately identify the performance parameters, instead of using the arithmetic mean of the occurrence probability of the aggregated interval data as an evaluation function, evaluate the value obtained by multiplying the geometric mean or the occurrence probability of each aggregated interval data. It needs to be a function.

  Next, Example 3 will be described. Example 3 is an example of the third embodiment described above, and describes a performance monitoring method using a performance model parameter identification method.

  The description of the parts that overlap the contents of the second embodiment is omitted. The performance prediction unit 11 estimates the processing amount at the peak based on the measurement data received from the measurement data input unit 1. For this estimation, for example, a method can be considered in which the processing amount when the processing amount per unit time is the maximum from the past measurement data in a predetermined period is the peak processing amount.

  Further, the performance prediction unit 11 receives the performance parameter from the parameter input / output unit 3 and predicts the probability distribution of the resource utilization rate at the peak based on the peak processing amount and the performance parameter.

Let each processing amount at the peak be (n ₁ , n ₂ , n ₃ ). In the normal distribution, the distribution of the usage amount by n processes is shown in the following formula (27).

  That is, the average value and the standard deviation value are multiplied by n. Therefore, the performance prediction unit 11 can calculate the probability density of the usage amount at the peak according to the following equation (28).

The performance prediction unit 11 transmits the probability density p ^j (R) of the resource usage amount to the notification unit 12.

The notification unit 12 notifies the system administrator by means such as displaying the probability density p ^j (R) of the received resource usage amount on the computer screen or sending an email.
In addition, the notification unit 12 does not always notify the administrator of the probability distribution p ^j (R) of the received resource usage amount. For example, the probability that the usage amount is equal to or greater than a predetermined value is determined in advance. It may be determined whether or not there is a notification, such as notifying the administrator if it exceeds the value, and not otherwise.

The following formula (29) shows a calculation formula for the probability that the usage amount will be T ⁰ or more.

  The administrator of the target system knows that the performance of the system has deteriorated due to the notification from the notification unit 12, and can take measures to prevent it before the performance problem occurs.

  FIG. 37 shows a graph of probability density when the number of processes per second at the peak time is (6, 6, 6). For example, when the resource usage rate is 30% (300 milliseconds) as an upper limit value and the administrator is notified when the probability of exceeding the upper limit value is 20% or more, the graph shown in FIG. Will be notified.

  Next, Example 4 will be described. Example 4 is an example of the above-described fourth embodiment, and describes a performance diagnosis method using a performance model parameter identification method.

  The description of the parts that overlap the contents of the second embodiment is omitted.

  The performance determination unit 21 receives the performance parameter from the parameter input / output unit 3 at the stage where the modification of the performance parameter has been completed, compares it with the value of the performance parameter received last time, and sends a notification signal to the notification unit 12. Determine whether to send.

  Specifically, for example, the performance determination unit 21 compares the value of the received performance parameter with the value of the performance parameter received last time, and determines a certain value that has an average value of the resource usage of a certain process. If it has changed more greatly, the processing name of the performance parameter is transmitted as a notification signal to the notification unit 12, and otherwise, the notification signal is not transmitted to the notification unit 12.

  FIG. 38 is a graph showing changes in the average value and standard deviation value of the process Trans1. The horizontal axis is the date, and performance parameters are identified every week. As a result of identifying the performance parameter values on September 19, the average value has changed greatly. When the average value change is larger than 1.0 and the process name is transmitted to the notification unit 12, the performance determination unit 21 notifies on September 19th.

  As another example of the determination method, the performance determination unit 21 compares the value of the received performance parameter with the value of the performance parameter received last time, and has a standard deviation value of the resource usage of a certain process. If the change is larger than the determined value, the processing name of the performance parameter may be transmitted as a notification signal to the notification unit 12, and otherwise, the notification signal may not be transmitted to the notification unit 12.

  FIG. 39 is a graph when the standard deviation value of Trans1 changes greatly. The horizontal axis is the date, and performance parameters are identified every week. When the standard deviation value changes more than 1.0 compared to the previous time and the process name is transmitted to the notification unit 12, the performance determination unit 21 notifies on July 20th.

  When the notification signal is received, the notification unit 12 notifies the system administrator by means such as displaying the processing name on the computer screen or sending a mail. As a result, the administrator of the target system can know that there is a process whose performance has changed greatly, and can take measures to prevent it before the performance problem occurs by investigating the cause.

  The present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. Various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the embodiment. For example, some components may be omitted from all the components shown in the embodiment. Furthermore, you may combine suitably the component covering different embodiment.

The block diagram which shows an example of a structure of the performance calculating apparatus according to the 1st Embodiment of this invention. The flowchart which shows an example of the processing operation of the performance calculating apparatus according to the first embodiment of the present invention. The flowchart which shows an example of the processing operation of the performance calculating apparatus according to the 2nd Embodiment of this invention. The block diagram which shows an example of a structure of the performance calculating apparatus according to the 3rd Embodiment of this invention. The flowchart which shows an example of the processing operation of the performance calculating apparatus according to the 3rd Embodiment of this invention. The block diagram which shows an example of a structure of the performance calculating apparatus according to the 4th Embodiment of this invention. The flowchart which shows an example of the processing operation of the performance calculating apparatus according to the 4th Embodiment of this invention. The figure which shows the 1st cut-out example of the total area when the number of processes is comparatively few. The figure which shows the calculation result of the resource usage-amount in each area in a table format. The figure which shows the calculation result of a resource usage-amount as a time series graph. The figure which shows the 2nd example of a cutting-out of a total area. The figure which shows the example of the calculation result of the processing number of each total area, and a resource utilization factor. FIG. 3 is a diagram illustrating an example of a performance model in the first embodiment. FIG. 3 is a conceptual diagram of a performance parameter correction method according to the first embodiment. FIG. 6 is a diagram illustrating an example of total section data in the first embodiment. The figure which shows the average value of the resource usage-amount in Example 1. FIG. The figure which shows the initial value of the resource usage-amount in Example 1. FIG. The figure which shows the transition of the evaluation function value in Example 1 with a graph. The figure which shows the transition of the average value parameter of process Trans1 in Example 1 with a graph. The figure which shows the transition of the average value parameter of process Trans2 in Example 1 with a graph. The figure which shows the transition of the average value parameter of process Trans3 in Example 1 with a graph. The figure which shows the performance parameter value of the experimental result in Example 1 as a table | surface. FIG. 10 is a diagram illustrating an example of a performance model in the second embodiment. FIG. 10 is a diagram illustrating a concept of parameter correction by optimization of evaluation values in the second embodiment. The figure which shows the average value of the resource usage-amount in Example 2, and the average value of a standard deviation value. The figure which shows the initial value of the average value of a resource usage-amount in Example 2, and a standard deviation value. The figure which shows the change of the evaluation function value according to parameter correction with a graph. The figure which shows the transition of the average value parameter of process Trans1 in Example 2 with a graph. The figure which shows the transition of the average value parameter of process Trans2 in Example 2 with a graph. The figure which shows the transition of the average value parameter of process Trans3 in Example 2 with a graph. The figure which shows the performance parameter value of the experimental result in Example 2 as a table | surface. The figure which shows the change of the evaluation function value according to parameter correction with a graph. The figure which shows the transition of the average value parameter of process Trans1 at the time of using the arithmetic mean of an evaluation function in Example 2 with a graph. The figure which shows the transition of the average value parameter of process Trans2 at the time of using the arithmetic mean of an evaluation function in Example 2 in a graph. The figure which shows the transition of the average value parameter of process Trans3 at the time of using the arithmetic mean of an evaluation function in Example 2 with a graph. The figure which shows the performance parameter value of the experimental result at the time of using the arithmetic mean of an evaluation function in Example 2 as a table | surface. The figure which shows the resource usage-amount at the peak time estimated in Example 3 with a graph. The figure which shows the change of the average value of process Trans1 in Example 4 with a graph. The figure which shows the change of the standard deviation value of process Trans1 in Example 4 with a graph.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Measurement data input part, 2 ... Sensitivity analysis part, 3 ... Parameter input / output part, 4 ... Evaluation value calculating part, 5 ... Parameter update part, 6 ... Memory | storage device, 11 ... Performance prediction part, 12 ... Notification part, 21 ... Performance judgment unit.

Claims (6)

Measurement data input means for inputting measurement data of the operation of the target device;
A performance parameter input means for inputting a performance parameter that is a parameter of the performance model;
Evaluation value calculation means for calculating an evaluation function value based on the measurement data input by the measurement data input means and the performance parameter;
Sensitivity calculation means for calculating the sensitivity of the evaluation function value calculated by the evaluation value calculation means for the performance parameter;
Updating means for updating the performance parameter according to the sensitivity calculated by the sensitivity calculating means so that the evaluation function value changes according to a predetermined condition;
A performance calculation apparatus comprising: an output unit that outputs a performance parameter updated by the performance update unit when a predetermined condition is satisfied. The performance model includes a processing resource usage model,
The usage amount of the resource usage model is represented by a probability distribution,
The evaluation value calculation means calculates the evaluation function value based on the occurrence probability of the resource utilization rate of the input measurement data,
The performance calculation apparatus according to claim 1, wherein the updating unit updates the performance parameter by correcting the probability distribution so that the evaluation function value changes according to a predetermined condition. In the resource usage model of the processing, the usage is expressed in a normal distribution,
The performance calculation device according to claim 2, wherein the performance parameter is an average value and a standard deviation value of a normal distribution representing the usage amount. The evaluation value calculation means includes
The measurement data input by the measurement data input means is divided into a plurality of time zones, and a geometric average of occurrence probabilities of resource utilization rates of the divided measurement data is calculated as the evaluation function value. 2. The performance calculation device according to 2. Estimating means for estimating a processing amount under a predetermined condition from measurement data input by the measurement data input means;
Performance prediction means for predicting the performance of the target device based on the processing amount estimated by the estimation means and the performance parameter output by the output means;
The performance calculation apparatus according to claim 1, further comprising notification means for notifying information on performance predicted by the performance prediction means. A performance determination unit that determines whether or not notification of the change is necessary based on a change amount of the performance parameter updated by the update unit;
The performance calculation apparatus according to claim 1, further comprising a notification unit that notifies the change when the performance determination unit determines that the change notification is necessary.

Images