JPH06274532A - Load distribution supporting device - Google Patents

Abstract

PURPOSE:To improve the working rate of the processors of parallel computers by dynamically distributing the loads of a retrieval processing. CONSTITUTION:A function and the various kinds of parameters, etc., are inputted from an input means 1, the various kinds of the parameters of the function are adjusted by a parameter adjusting means 5 and a retrieval time is estimated from the function for each retrieval, key word by a load prediction means 6. The number of the processors to be allocated to each retrieval key word is calculated from the ratio of the estimated retrieval time by a processor allocation means 7 and the predicted retrieval time of the retrieval key word is compared with a real measured time by a processor allocation verifying means 8. Allocation is changed by interaction with a user by a processor allocation adjusting means 9 and a changed result can be further verified by the processor allocation verifying means 8.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention uses a function peculiar to a target area when a search target having a structured index with a structured search keyword is searched in parallel on a parallel computer. Simulate load balancing of search processing, and provide appropriate load balancing support.
The present invention relates to a load balancing support device, and an information recall system, an idea support system, an expert system, and a CAD that perform ambiguous searches using structured search keywords.
Used for / CAM system, database system, etc.

[0002]

2. Description of the Related Art There is no practical load balancing algorithm for a parallel computer when a database is searched by a parallel computer. Currently used methods include allocating processors at the program level so that users can efficiently distribute the load of search processing.
Alternatively, there is research such as investigating the operating state of a processor and allocating the divided work to a processor with a low operating rate.

[0003]

In the method of allocating processors at the program level so that the user can efficiently distribute the load of search processing, a search keyword having a different structure from the search keyword assumed when the program was created was input. In this case, it becomes necessary to newly change the allocation of processors. However, since it is difficult to dynamically allocate the processors, the search is performed as it is in the processor allocation state set when the program was created. As a result, there is a problem that it is difficult to increase the operating rate of all the processors.

In the method of checking the operating state of a processor and allocating the divided work to a processor with a low operating rate, the setting of the unit for dividing the work of the search processing depends on the domain and the problem. This is effective when the problem is simple, but there is a problem that it is difficult to set an appropriate division unit when the number of processors increases or the target problem becomes complicated.

Therefore, it is an object of the present invention to apply a load balancing support system capable of improving the operating rate of the processors of a parallel computer by dynamically distributing the load of search processing.

[0006]

In order to achieve the above-mentioned object, the present invention utilizes a correlation between a structured search keyword and its search time to construct a structured search keyword, the number of processors used, and parallel processing. A means for inputting the actual measurement time and the function of the search keyword on the computer as character information, a means for outputting the character information, a means for predicting the search time for the search keyword using a function, and a search for the predicted search keyword A means for calculating the number of processors to be allocated to each search keyword from time, a means for comparing the calculated processor allocation information with the actual measurement time, a means for adjusting the processor allocation information, an intermediate result of processing data and control It is composed of a means for temporarily storing the intermediate results of the above, a means for adjusting the parameters of the function, and a means for controlling these.

That is, the present invention estimates the rough search time of the input search keyword from the function obtained from the previous search result by using the function peculiar to the target area, and from the ratio of those search times. Load balancing support for parallel computers characterized by calculating the number of processors to be assigned to search keywords and comparing them with the actual measurement time that has been measured in advance to support the allocation of processors that perform appropriate load distribution. It is a device.

Here, the function peculiar to the target area is a function obtained by performing a search for all indexes in advance using the index as a search keyword and correlating the search time with the search keyword. Further, the structured search keyword is expressed by a tree structure, whereas a normal search keyword is expressed by a pair of an attribute and an attribute value. For example, the search keyword for an optical sensor circuit that has the function of detecting light and converting it into an electrical signal (current) can be expressed as "attribute sensor attribute value light", but with a structured search keyword, "make (insert (light)" , Change (light, current)) ”. Hereinafter, structured search keywords and indexes are abbreviated as search keywords and indexes.

[0009]

First, under the control of the control means, the input means inputs the index, the search keyword and the function, the actual measurement time of the search keyword on the parallel computer and the number of processors to be used in the parallel computer, and stores them in the storage means. To be done.

Next, various parameters of the function are displayed by the parameter adjusting means on the display means, and the user corrects them by using the input means. The adjusted information is stored in the storage means.

Next, the load predicting means estimates the search time of each input search keyword from the function for each search keyword. The estimated search time is stored in the storage means.

Next, the processor allocating means calculates the number of processors allocated to each search keyword from the estimated ratio of the search times. The calculated allocation information of the processors is stored in the storage means.

Next, based on the processor allocation information calculated by the processor allocation verification means, the search time and the actual search time of the predicted search keyword are calculated using the measurement data of the search keyword measured in advance on the parallel computer. The display unit displays the information such as the difference between the search keyword name, the predicted search time, the actual measurement time, and the like in the descending order of the search keywords by comparing with the measurement time.

Next, the processor allocation adjusting means displays the allocation information of the processor to the search keyword on the display means, and the allocation is changed by the dialogue with the user. The change result is stored in the storage means. The change result can be further verified by the processor allocation verification means. Finally, the processor allocation information is output to the output means.

[0015]

The present invention will be described in detail below with reference to the drawings.
FIG. 1 is a block configuration diagram of a load balancing support apparatus according to the present invention. In the figure, reference numeral 1 is an input means such as a keyboard / mouse for inputting character information such as a search keyword. Reference numeral 2 is an output means such as a printer for outputting the input character information and processor allocation information. Reference numeral 3 is a storage means for storing the input information in a semiconductor device, a disk, a floppy disk or the like. Reference numeral 4 is a display means such as a CRT / bitmap display for displaying an input or intermediate processing result as character information or multi-window image information. Reference numeral 5 is a parameter adjusting means for adjusting the parameters of various input functions by interacting with the user.

Reference numeral 6 is a load predicting means for predicting the load of each search keyword from each input search keyword and function. Reference numeral 7 denotes a processor allocation means for estimating the number of processors to be allocated to each search keyword from the predicted search time of the search keyword and the number of used processors. 8 is based on the allocation information of the processor, using the actual measurement time of the search keyword measured in advance on the parallel computer, calculates the comparison between the search time of the predicted search keyword and the search time of the actual measurement time, It is a processor allocation verification means that rearranges in descending order. Reference numeral 9 denotes a processor allocating means for adjusting processor allocation information of the search keyword. The processor 10 controls the CP for controlling each of these means.
Control means such as U.

FIG. 2 is a schematic flow chart showing the function of the load balancing support system of the present invention. Reference numeral 11 is a parameter adjustment module that supports correction of each parameter of the input function. Reference numeral 12 is a maximum search keyword selection module that selects a search keyword having the maximum search time among the search keywords. 13 is the fastest search time of the search keyword with the longest search time (the search time when the speed improvement reaches a peak, but the search time at which the speed improvement does not stop completely but the improvement rate becomes small) and the fastest search time. It is a reference search time setting module for calculating the number of used processors. Reference numeral 14 is a basic search time calculation module for calculating the N processor search time of each search keyword. Reference numeral 15 is a reference search time calculation module for calculating the number of processors used by each search keyword (hereinafter referred to as the reference processor number) and its search time (hereinafter referred to as the reference search time).

Reference numeral 16 denotes a work amount calculation module for calculating the work amount of each search keyword (product of reference search time and reference processor number). A sort module 17 sorts the search keywords in ascending order based on the calculated work amount. Reference numeral 18 denotes a processor allocation module that calculates the work ratio of search keywords and the number of allocated processors. Reference numeral 19 denotes a processor allocation verification module that compares the allocation information of the processor with the actual measurement time and displays the comparison content on the display means. Reference numeral 20 denotes a processor allocation adjustment module that supports correction of processor allocation information.

The correction contents of the processor allocation adjustment module 20 can be verified by using the processor allocation verification module 19. Further, in the modification of the processor allocation adjustment module 20, when the user cannot perform satisfactory load distribution, it is possible to feed back to the parameter adjustment module 11. Further, it is also possible for the user to freely perform the processor allocation processing by directly executing the processing by the processor allocation adjustment module 19 without performing the series of processing subsequent to the parameter adjustment module 11.

A detailed description will be given below for each of these modules. In the parameter adjustment module 11,
The parameters of the function stored in the storage means 3 are modified by the interaction with the user. In this module, the parameter information in the function is cut out and the information is output to the display means 4. Then, the parameter information corrected by the user from the input means 1 is stored in the storage means 3. Figure 3
Shows a display example of the parameter information of the linear function. Reference numeral 21 indicates the name of the function. Reference numeral 22 denotes a function to be modified. Reference numeral 23 denotes a correction target parameter of the function.
A value intended by the user can be input using the input means 1 by clicking the portion * in 23 with the mouse. Reference numeral 24 denotes a menu of operation commands for storing the input correction content in the storage means 3, and do_
Clicking on it with the mouse will set the value. ab
Clicking on ort cancels the entered value. When the parameters are not modified, the value at the time when the function is first input by the input means 1 is used.

Next, the function will be described. It is necessary to reflect the two characteristics of the parallel computer and the search processing in the function. The characteristic of a parallel computer is that the speed of search when one search keyword is executed in parallel by multiple processors is improved by only about the square root of the number of processors used compared to the search time when it is executed by one processor. Is. This is because the communication cost between the processors increases. However, when the number of processors is several, the speed improvement may increase to the same extent as the number of processors used due to the difference in the method of implementing the processing system of the programming language of the parallel computer and the computer architecture.

The characteristic of the search process is that the search time is about several times longer on average than the search keyword having the same structure as the search time of the search keyword composed of a specific word. For example, in the case of a circuit case-based search, most circuits have input / output functions and data conversion functions as basic functions, so inevitably "insert", "change", and "issue" as case-based indexes. The use of words such as "and" make "increases. Therefore, the number of collations at the time of fuzzy search increases. As a result, even if they have the same structure, the search time for a search keyword containing these specific words will be long. Therefore, it is necessary to roughly estimate the search time for the search keyword including these words.

When creating the function, it is necessary to quantify the structural size of the search keyword. The method of quantification depends on the domain and the problem. Here, the quantified value is called an ideal value, and 0.8 is calculated for each of the root node and the intermediate node of the tree structure, and 0.3 is calculated for each of the end nodes of the tree structure. FIG. 4 shows an example of the search keyword. The index has the same structure. The search keyword 25 “count” has one root node, four intermediate nodes, and six end nodes, so the ideal value is: 5 *
It becomes 0.8 + 6 * 0.3 = 5.8.

Both the search keywords 26 "measure" and 27 "make" have one root node and two intermediate nodes,
Since there are three terminal nodes, the ideal value is: 3 * 0.8 + 3 *
0.3 = 3.3.

The search keyword having a larger ideal value has a larger structure of the search keyword. Therefore, a search keyword with a large ideal value has a correlation that the number of times of matching at the time of ambiguous search increases, and the search time also increases.
The function described below is basically a function of the relationship between the ideal value and the search time.

Next, the method of functionalization will be described. First, the search keywords in the index are executed one by one on the parallel computer. FIG. 5 shows the measurement contents of an actual parallel computer. In Figure 5, changing the number of processors from 16 to 1,
The figure shows a situation in which the search process is measured twice. The measurement is performed twice to prevent the measurement time from increasing due to the occurrence of garbage collection in the processor. In FIG. 5, reference numeral 28 shows a specific structure of the search keyword, which represents the search keyword 26 of FIG. 4B in a vector format. 29 is an ideal value. Reference numeral 30 indicates the maximum number of processors used for measurement. In FIG. 5, a maximum of 16 processors are used. Reference numeral 31 is measurement data. Reference numeral 32 indicates the location where garbage collection occurs when the number of measurement processors is 1.
The measurement time for the first time is getting longer. When garbage collection occurs, the smaller value is the appropriate measurement value.

After measuring all indexes, plotting is performed with the number of processors on the X axis and the search time on the Y axis as shown in FIG. Next, the plotted points are approximated by a function including a square root. 33 in FIG. 7 is a function: f (x) = {2 processor search time} / SQR
(1.1 * (x−1)), where x ≧ 3, approximation is performed. x is the number of processors. SQR means take the square root of the value in parentheses. Because of the above-mentioned characteristics, the search time of all processors cannot be obtained by a function including the square root, and thus approximation is performed with the number of processors of 3 or more. The plotted points are functionalized for the measurement data of all indexes, and the function that covers the plotted points of all indexes is adopted.

Next, the functionization of the one-processor search time and the two-processor search time, which cannot be approximated by the function including the square root, will be described. The one-processor search time or the two-processor search time of the measurement data of all indexes is plotted to find an approximate function. FIG. 8 is a plot of the 2-processor search time for all eight indexes, where the X-axis is the ideal value and the Y-axis is the 2-processor search time. 34 is a function: f (X) =
It is an approximation of 2.8 * x−0.4. Similarly, the one-processor measurement time is also converted into a function.

As described above, the types of functions used to reflect the characteristics of the parallel computer are summarized as follows.

[Type 1] Function of relationship between ideal value and N processor search time. When the number of processors is several, the speedup may increase to the same extent as the number of processors used due to differences in the implementation method of the processing system of the programming language of the parallel computer and the computer architecture. Is a function for obtaining. This is a function for obtaining the one-processor search time and the two-processor search time described above. This function is a function of the relation that the search time executed by the N processors increases as the ideal value increases.
N is a natural number. Hereinafter, this function is called a function: N processor search time. The input of the function is an ideal value, and the output is the search time when using N processors.

[Type 2] A function of the relationship between the number of processors used and the search time of the search keyword. This is a function that calculates the search time when one search keyword is executed in parallel by an arbitrary number of processors. It is expressed by a function that includes the square root in the denominator (below). In the formula below, x
Is the number of processors, and {x processor search time} is the search time for x processors. a is a constant and x is a rational number. M is a natural number. SQR means take the square root of the value in parentheses. Hereinafter, a function of this relationship will be referred to as a function search keyword search time. Function including the above square root: f (x) = {2 processor search time} / S
QR (1.1 * (x-1)) corresponds to this. f (x) = {x processor search time} / SQR (a * (x-M)) When performing load balancing, it is more convenient to use the fastest search time of the search keyword that takes the longest search time as a reference. Is good. This is because the search time when there are sufficient processors is often the fastest search time of the search keyword that takes the longest search time. Therefore, the following [type 3] function is used. Further, since the present invention uses a value obtained by multiplying the search time by the number of processors used at that time (hereinafter abbreviated as the work amount) as a guide for predicting the search time, the function of [type 4] is used.

[Type 3] Function of relationship between ideal value and fastest search time. This function is a function of the relationship that the fastest search time increases as the ideal value increases. Hereinafter, this function will be referred to as function: fastest search time. The input of the function is an ideal value, and the output is the fastest search time of the search keyword having this ideal value. This function is also converted into a function by the same method as the function for obtaining the two-processor search time described above.

[Type 4] Function of relation between ideal value and required number of processors. The required number of processors is the number of processors used when the above-mentioned fastest search time is reached. This function is a function that the number of required processors increases as the ideal value increases. Hereinafter, this function is called a function: the required number of processors. The input of the function is an ideal value, and the output is the required number of processors of the search keyword having this ideal value.

Regarding the characteristics of the retrieval process, the function of [Type 1] used: N processor For a plurality of retrieval keywords for which an appropriate value cannot be obtained in the retrieval time,
Check whether there is a combination of specific words, and make those combinations a specific word. Then, all the search keywords having the same combination of specific words, which is how many times larger than the value obtained by the function: N processor search time, are examined, and the average value (hereinafter referred to as the correction magnification) is given. This correction factor is used when finding the search time of a search keyword having a specific combination of words of the search keyword. That is, the value obtained by the [type 1] function is multiplied by this correction factor to obtain the search time.

Here, for convenience of explanation, the following will be used as a concrete embodiment of each function. In the following, [Function 1],
[Function 2], [Function 4], and [Function 5] are linear functions, but even if the order increases, this apparatus is not affected.
Further, here, the description will be given by setting the number of usable processors to 50.

[Type 1] function: 1 processor search time f (x) = 5.7 * x-7.1 ... [Function 1] [Type 1] function: 2 processor search time f (x) = 3.7 * X-7.7 ... [Function 2] [Type 2] Function: Search keyword search time f (x) = {2 processor search time} / SQR (1.1
* (N-1)) [Function 3] [Type 3] Function: Fastest search time f (x) = 0.7 * x-0.1 ... [Function 4] [Type 4] Function: Number of required processors f (X) = 0.8 * x + 1.4 ... [Function 5] Further, the correction magnification is 1 processor search time is 1.3, 2 processor search time is 1.3, and a specific word is created or inserted.
It will be described as a search keyword including change.

Maximum Search Keyword Selection Module 12
Calculates an ideal value for a plurality of input search keywords and selects the search keyword having the maximum ideal value. For example, assume that the three search keywords shown in FIG. 4 are given. The ideal values of the three search keywords are calculated.
The calculation method is as described above. Since the ideal values of the search keyword 25, the search keyword 26, and the search keyword 27 are 5.8, 3.3, and 3.3, the search keyword 25 is selected. The selection result is stored in the storage unit 3.

The reference search time setting module 13 obtains the fastest search time of the search keyword 25 and the required number of processors from [function 4] and [function 5]. From the calculation of the function, the fastest search time is 4.0 (seconds), and the required number of processors is 6.
It becomes 0. The obtained fastest search time is stored in the storage means 3.

The basic search time calculation module 14 calculates the one-processor search time and the two-processor search time of the search keyword 25 from [function 1] and [function 2]. The calculated values are 26.0 and 13.8, respectively. The one-processor search time and the two-processor search time of the search keyword 26 are calculated from [function 1] and [function 2]. The calculated values are 11.7 and 4.5, respectively. Search keyword 27 of 1
When the processor search time and the two-processor search time are calculated from [Function 1] and [Function 2], the ideal value is 3.3, so that the search keyword 26 is 11.7, respectively.
It will be 4.5. However, since the search keyword 27 is a keyword including a specific word, the correction rate is 1.3 times for both the one-processor search time and the two-processor search time, and thus the calculated values are 1.3 times. Therefore, the one-processor search time and the two-processor search time are 15.2 and 5.9, respectively. The calculation result is stored in the storage unit 3.
To be remembered.

The reference search time calculation module 15 obtains the number of processors required for searching in the calculated fastest search time 4.0 and its reference search time for each search keyword based on [function 3]. The processing flow is shown in FIG. Since this process does not always match the search time of a particular search keyword with the fastest search time calculated earlier, it is necessary to reduce the difference between the search time of the search keyword and the size of the fastest search time. It is a thing. The reference search time is usually the fastest search time, but if the N processor search time (in this case, 1 processor search time and 2 processor search time) is shorter than the fastest search time, the N processor search time is the reference search time. It's time.

FIG. 9 is a processing flow when the function: 1 processor search time and the function: 2 processor search time are used. Reference numeral 35 in the figure indicates a condition for setting the search time of the search keyword to the 2-processor search time when the value of the 2-processor search time falls within ± 10% of the value of the fastest search time. Reference numeral 36 indicates a condition for calculating the search time of the search keyword from the function: search keyword search time when the two-processor search time is longer than 1.1 times the fastest search time. Reference numeral 37 is a condition for obtaining the one-processor search time or the two-processor search time as the reference search time of the search keyword.

Function: FIG. 10 shows the processing flow when only one processor search time is used. 38 is a function: if the one-processor search time is more than 1.1 times the fastest time, the reference search time is calculated using the function: search keyword search time, and if it is smaller, the one-processor search time is set as the reference search time Is the condition.

FIG. 11 shows the processing flow when the function: 1 processor search time to the function: N processor search time is used. Reference numeral 39 indicates a condition for setting the search time of the search keyword to the N processor search time when the value of the N processor search time falls within ± 10% of the value of the fastest search time. Reference numeral 40 indicates a condition for calculating the search time of the search keyword from the function: search keyword search time when the N processor search time is longer than 1.1 times the fastest search time. 41,
Reference numerals 42 and 43 are conditions for obtaining the reference search time of the search keyword from any of the N-1 processor search time to the one processor search time.

Since the two-processor search time of the search keyword 26 is 4.5 (seconds), it does not meet the first condition. Since the condition “1.1 * {fastest search time} <{2 processor processing time}” is satisfied, the number of processors used is: (4.5 / 4.0) * (4 .5 / 4.
0) /1.1+1=2.2.

Since the 2-processor search time of the search keyword 27 is 5.9 (seconds), similarly, the first condition is not met. Since the condition “1.1 * {fastest search time} <{2 processor search time}” is satisfied, the number of processors used is: (5.9 / 4.0) * (5 ．
9 / 4.0) /1.1+1=3.0.

From the above, the reference search time of the search keyword 25 is 4.0 (seconds), the number of reference processors is 6.0, and the reference search time of the search keyword 26 is 4.5.
(Seconds), the reference number of processors is 2.2, the reference search time of the search keyword 27 is 5.9 (seconds), and the number of reference processors is 3.0.

The work amount calculation module 16 calculates the work amount by multiplying the required number of processors and the reference search time for each search keyword. In this case, the workload of the search keywords 25, 26, and 27 is 4.0 * 6.0 =
24.0, 4.5 * 2.2 = 9.9, 5.9 * 3.0 =
It becomes 17.7.

The sorting module 17 sorts the search keywords in the order of smaller work amount based on the calculated work amount values. The search keywords 26, 27, and 25 are sorted in this order. Even if sorting is done in descending order of work, this device is not affected.

In the processor allocation module 18,
Calculate according to the work load ratio of each search keyword,
The number of processors assigned to each search keyword is calculated with respect to the input number of usable processors. Multiplying the ratio of work of search keywords by the number of available processors will calculate the number of processors according to the ratio, but the calculated number of processors will become a rational number and must be converted to a natural number by rounding. If the number of allocated processors is calculated by rounding off, the allocation of processors will be considerably rough. For example, if the calculated number of processors for a certain search keyword is 1.4 and only one processor is assigned by rounding, the calculation time will be about 1.4 times that of other search keywords. I will end up. Also,
Conversely, if one processor is allocated even though the calculated number of processors is 0.5, the actual calculation time will be half. In this way, even if the same number of processors is assigned, the load of search processing is not uniform. In such a case, if the same two processors are assigned to two search keywords, the search processing of two search keywords per processor becomes: 1.4 / 2 + 0.5 / 2 = 0.95 processors, The magnitude of the load can be made uniform.

By thus allowing a plurality of search keywords to be assigned to a plurality of the same processors, the load of the search processing can be made uniform. Furthermore, the calculated number of processors is divided by a natural number obtained by rounding down the decimal point of the calculated number of processors, or a natural number rounded up to the nearest decimal point so that the allocation error of the processor is about 10%. Correct so that it becomes 1. If it is not between 0.9 and 1.1, the number of processors will be calculated by allocating it together with other search keywords. The allocation error of the processor does not have to be 10% because it depends on the architecture and processing system of the parallel computer used and the search problem. The formula for allocating the search keyword i having the above characteristics is shown in Equation 1.

[0051]

[Equation 1]

In this case, since the number of usable processors is 50, the search keyword 26 having the smallest work amount is calculated from the formula as follows.

W1 * N / W = 50 * 9.9 / (9.9 + 17.7 + 24.0) = 9.6 Therefore, N11 = 9 or 10.

Therefore, when N11 = 9, the equation 2 is obtained.

[0055]

[Equation 2]

Even when N11 is 12, it is 0.96, which satisfies the condition. If both the rounded down natural number and the rounded up natural number satisfy the expression, either one may be selected. Here, if the truncated natural number satisfies the condition first, this is set as the allocated processor number.

Therefore, nine processors are assigned to the search keyword 26. Applying the above formula to the search keyword 27, W2 * N / W = 50 * 17.7 / (9.9 + 17.7 + 24.0) = 17.2 Therefore, N22 = 17 or 18.

Therefore, when N22 is 17, Equation 3 is obtained.

[0059]

[Equation 3]

Therefore, 17 processors are assigned to the search keyword 27. From the formula, the search keyword 25 is W3 * N / W = 50 * 24.0 / (9.9 + 17.7 + 24.0) = 23.3 Therefore, N33 = 23, or 24.

Therefore, when N33 = 23, Equation 4 is obtained.

[0062]

[Equation 4]

Then, the calculation is such that 23 processors are allocated. However, since only 26 processors are assigned to the search keywords 26 and 27 in total,
The remaining 24 unused processors will be allocated. In this way, the number of processors of the search keyword finally assigned becomes the number of remaining unused processors. The calculation result is stored in the storage means. From the above, the number of processors assigned to the search keyword 26 is 9, and the assigned processor nodes on the parallel computer are 49 to 41. The number of processors assigned to the search keyword 27 is 17, and the assigned processor nodes on the parallel computer are 40 to 24. The number of processors assigned to the search keyword 25 is 24
The number of processor nodes on the parallel computer to be assigned is 23 to 0.

Here, it is assumed that the search keywords 25, 26,
The work of 27 is 163.5, 8.9 and 17.2 respectively.
The following is an example of the case. In this case, the search keyword 26 having the smallest work amount is calculated as follows.

W1 * N / W = 50 * 8.9 / (8.9 + 17.2 + 163.5) = 2.3 Therefore, N11 = 2 or 3.

Therefore, when N11 = 2, the equation 5 is obtained.

[0067]

[Equation 5]

When N11 = 3, the equation 6 is obtained.

[0069]

[Equation 6]

Therefore, the above condition is not satisfied. Therefore, it is assigned to the processor together with the search keyword 27. When the search keywords 26 and 27 are applied to the above formula, W1 * N / W + W2 * N / W = 50 * 8.9 / (8.9 + 17.2 + 63.5) + 50 * 17.2 / (8.9 + 17) .2 + 163.5)) = 6.8 Therefore, N12 = 6, or 7.

Therefore, when N12 = 6, the equation 7 is obtained.

[0072]

[Equation 7]

When N12 = 7, the equation 8 is obtained.

[0074]

[Equation 8]

By the above calculation, the search keywords 26 and 2
Seven processors are allocated to seven.

From the formula, the search keyword 25 is W3 * N / W = 50 * 163.5 / (8.9 + 17.2 + 163.5) = 43.1 Therefore, N33 = 43 or 44.

Therefore, when N33 = 43, the equation 9 is obtained.

[0078]

[Equation 9]

As a result, 43 processors are allocated. From the above, the number of processors assigned to the search keyword 26 is 7, and the assigned processor nodes on the parallel computer are 49 to 43. The number of processors assigned to the search keyword 27 is 7, and the assigned processor nodes on the parallel computer are 49 to 4
It will be 3. The number of processors assigned to the search keyword 25 is 43, and the assigned processor nodes on the parallel computer are 42 to 0.

Next, the processor allocation verification module 19 compares the measured time with the actual measured time. The actual measurement time is data measured in advance for all indexes and all tree structures that can be created from the indexes. FIG. 12 shows an example of reference measurement data of the search keyword 25. 44 is an ideal value of the search keyword of this measurement data. 45 is a measurement time. In the data of FIG. 5, two processor measurement times exist, but in FIG. 12, the processor search time is only one (the search time is smaller). If there is more than one with the same ideal value, at each processor search time,
The average value of them is used as the processor search time. The actual measurement time used for reference has the same ideal value. if,
If there is no same, the actual measurement time having an ideal value within the range of plus or minus 10% of the ideal value is referred to.

Information such as the predicted search time, the number of allocated processors, and the actual measurement time is displayed on the display means 4 for each search keyword.

FIG. 13 shows a display example. 46 is a concrete structure of the search keyword. 47 is the ideal value of the search keyword. 48 is the node name of the processor to be assigned. 49 is a predicted search time of the search keyword. 50 is an error between the predicted search time and the actual measurement time. 5
1 is the number of allocated processors. 52 is the actual measurement time. 53 is the total predicted search time. 54 is the total actual measurement time. 55 is the total error.
Reference numeral 56 is a boundary for grouping search keyword groups having the same search keyword allocation node.
Here, the total predicted search time and the total actual measurement time are the sum of the predicted search time and the actual measurement time of all the search keywords assigned to the same processor. The total error is the error between the total predicted search time and the total actual measurement time.

Here, it is assumed that the search keywords 26 and 27 are used.
Is assigned to 26 processors, and the search keyword is 25
Is assigned to 24 processors, FIG.
As described above, the search keyword 26 and the search keyword 27 are in the same group.

The processor allocation adjusting module 20 corrects the processor allocation by looking at the displayed contents. The present processor allocation information as shown in FIG. 15 is displayed on the display means 4. Reference numeral 57 is a specific structure of the search keyword, and the 21st and subsequent characters are displayed in abbreviated form in which * is substituted. Reference numeral 58 is the number of allocated processors. 5
Reference numeral 9 indicates information on whether to search together by the same processor. Reference numeral 60 denotes a menu of operation commands for storing the correction contents in the storage means. When do_it is clicked with the mouse, a value is set, and when the abort is clicked, the input value is canceled. In 59, the same character string is added when searching the same processor together.

When the user clicks the number of processors to which each search keyword is assigned or the processor shared search in FIG. 15 with the mouse, the clicked value disappears and an appropriate value can be input using the input means 1 instead.

After the correction of the value, the processing of the processor allocation verification module 19 can be executed to perform the verification.

When this module is first executed, all columns are marked with * as shown in FIG.

The output module outputs according to the specified format.

[0089]

According to the load balancing support apparatus of the present invention, it is possible to refer to the processor information dynamically assigned to a plurality of input search keywords by using a function. It has a unique effect that the work can be reduced.

Further, since it is possible to compare with the past actual measurement data, it is possible to execute the actual parallel computer without executing the actual parallel computer.
A unique effect that load distribution work can be performed is achieved.

Further, there is an effect that the operating rate of the processor of the parallel computer can be easily increased.

[Brief description of drawings]

FIG. 1 is a block configuration diagram of a load balancing support apparatus according to the present invention.

FIG. 2 is a schematic flow diagram of a load balancing support apparatus according to the present invention.

FIG. 3 is a diagram showing an example of a screen for modifying function parameter information.

FIG. 4 is a diagram showing an example of a search keyword.

FIG. 5 is a diagram showing an example of data measurement.

FIG. 6 is a diagram showing an example of a measurement time plot.

[Figure 7] Example of functionalization (function: search keyword search time)
FIG.

FIG. 8 is a diagram showing an example of functionization (function: two processor search time).

FIG. 9 is a diagram showing a reference search time calculation flow from the function: 1 processor search time to the function: 2 processor search time.

FIG. 10 is a diagram showing a reference search time calculation flow when a function: 1 processor search time is used.

FIG. 11 is a diagram showing a reference search time calculation flow when a function: N processor search time is used from a function: 1 processor search time.

FIG. 12 is a diagram showing an example of reference measurement data.

FIG. 13 is a diagram showing a comparative example of predicted measurement time and actual measurement time.

FIG. 14 is a diagram showing a comparative example of predicted measurement time and actual measurement time.

FIG. 15 is a diagram illustrating an example of processor allocation adjustment.

FIG. 16 is a diagram illustrating an example of processor allocation adjustment.

[Explanation of symbols]

 1 Input Means 2 Output Means 3 Storage Means 4 Display Means 5 Parameter Adjusting Means 6 Load Predicting Means 7 Processor Allocation Means 8 Processor Allocation Verifying Means 9 Processor Allocation Adjusting Means 10 Control Means 11 Parameter Adjusting Module 12 Maximum Search Keyword Selection Module 13 Standard Search Time setting module 14 Basic search time calculation module 15 Standard search time calculation module 16 Work load calculation module 17 Sort module 18 Processor allocation module 19 Processor allocation verification module 20 Processor allocation adjustment module

Claims (1)

[Claims] 1. A means for inputting a structured search keyword and function as character information, a means for outputting character information, a means for predicting a search time for a search keyword using a function, and an intermediate result of processed data. Or a means for temporarily storing the intermediate results of control, a means for adjusting the parameters of the function, and a means for controlling these, and by using a function specific to the target area, the search keyword input is roughly Load balancing support apparatus, which estimates a specific search time from a function obtained from a previous search result.