JP2006331135A - Performance prediction device, performance prediction method and performance prediction program for cluster system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a cluster system performance prediction device capable of predicting clustering overhead generated in a cluster system sharing data present on a main storage device of each node via a communication line. <P>SOLUTION: In this performance prediction device for the cluster system sharing the data on the main storage device of the plurality of nodes connected via the communication line between the nodes, a frequency of a block copy per unit time performed between the nodes with a memory block inside the main storage device as a unit is calculated, and the clustering overhead generated in the cluster system is calculated on the basis of information about a CPU load per block copy and the block copy frequency. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a performance prediction device for a cluster system, and more particularly to a performance prediction device for a cluster system capable of predicting a clustering overhead generated in a cluster system that shares data on the main storage device of each node via a network. The invention relates to a performance prediction method and a performance prediction program.

  In recent years, with the spread of Web systems, computer systems used for server applications are increasing. In such a computer system, a cluster system using a plurality of computers is used for the purpose of improving the processing capacity of the Web system. For this reason, the importance of the performance design of the cluster system is increasing.

  Although the cluster system seems to achieve a performance improvement commensurate with the number of computers when considered simply, there is actually a clustering overhead that does not occur in a single-node system. It does not improve in proportion to Here, the clustering overhead represents an overhead generated due to a plurality of nodes. Hereinafter, the performance increase rate with respect to the cluster size will be expressed in terms of scalability. The cluster size represents the number of nodes constituting the cluster system.

  When the scalability of the cluster system is low, there is a demand to determine whether the cause is due to clustering overhead or other factors. This is because if the main cause is clustering overhead, it can be determined that rather than aiming to improve performance by scaling out, it should be aimed to scale up by enhancing the performance of a single hardware such as increasing the CPU speed. Thus, the prediction of clustering overhead is important in designing the performance of a cluster system.

  In a cluster system that shares data on the main memory of a node via a network, it is known that block copying for data sharing is a major cause of clustering overhead.

  The following methods have been proposed as performance prediction methods in conventional computer systems.

  The method disclosed in Patent Document 1 relates to a computer system performance prediction program. In this performance prediction program, the performance of the computer system is predicted by performing a simulation for tracking the execution time of the component program using the performance of the component program and the combination information thereof as input parameters. The performance prediction program performs performance prediction for a specific application and workload on a system composed of a single node.

  The method of Patent Document 2 relates to a Web system performance prediction method. In this performance prediction method, model pattern information that is information related to the configuration of the Web system, information related to the characteristics of the Web system, and information related to functions provided by the Web system are registered in advance, and the function name provided by the Web system (for example, When information on document management), priority items (price, performance), and the number of users is input, the configuration, function, server model, price, performance, and server load are output for each server model with different performance. Note that the performance prediction is performed based on past performance information. Therefore, this method can be applied to the performance prediction of a cluster system for which past performance information is available.

  The method of Patent Document 3 relates to a method for realizing various workflows with one system component connection model by dividing a system component connection model and a workflow indicating a data path. In the method of Patent Document 3, first, system components, a workload, and a workflow of a target system are input. The system component includes a master component and a slave component, and the amount of data generated from the master component is input as a workload.

  The method of Patent Document 4 relates to a method for analyzing the number of execution times of a parallel execution program that can confirm the load distribution amount for parallelization numerically and can support parallelization tuning of the parallel execution program. In this execution frequency analysis method, a count instruction for counting the number of program executions is inserted into the analysis target program. Calculate the total number of executions of the executable statement of the program measured by the count instruction, and calculate the performance improvement rate when the parallel degree is changed. Next, using these results, the execution number of execution statements of the specific loop, the average rotation speed of the loop, and the estimated value of the performance improvement rate with respect to the change in the parallel degree are displayed.

  Non-Patent Document 1 describes an example of a performance model of a conventional cluster system, a LogP model. The LogP model is used for finding bottlenecks in cluster systems and analyzing parallel algorithms. The upper limit of communication delay (Latency), the processor time consumed by sending or receiving one message (overhead), and the minimum message sending ( The performance is predicted using four parameters, ie, (reception) interval (gap) and the number of processors (Processors). The LogP model is used, for example, for finding a bottleneck when a fast Fourier transform algorithm is executed and for designing a more efficient fast Fourier transform algorithm.

  Non-Patent Document 2 describes an example of a model obtained by extending Non-Patent Document 1, a LogGPS model. In addition to the input parameters considered in the LogP model, the LogGPS model includes input parameters related to the transmission of long messages and the synchronization that occurs during the transmission. Thus, by adding input parameters, performance can be predicted with higher accuracy than in Non-Patent Document 1.

The techniques described in Non-Patent Document 1 and Non-Patent Document 2 can be applied to a cluster system in which an application that can calculate the calculation amount and the communication amount from an algorithm operates.
JP 2004-272582 A JP 2004-46734 A JP 2004-14937 A JP 2004-102594 A David Culler, Richard Karp, David Patterson, Abhijit Sahay, Klaus Erik Schauser, Eunice Santos, Ramesh Subramonian, and Thorsten von Eicken.LogP: Towards a Realistic Model of Parallel Computation.In Proceedings of the Fourth ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, pages 1--12, May, 1993. Fumihiko Ino, Noriyuki Fujimoto, Kenichi Sugawara. LogGPS: A parallel computing model for high-level communication libraries that considers switching message communication protocols. IPSJ Transactions: High Performance Computing System, Vol.42, No.SIG09.

  All of the conventional techniques described above have the following problems.

  The performance prediction program in the method of Patent Document 1 relates to a system composed of a single node, and is not applicable to a cluster system composed of a plurality of nodes.

  The method of Patent Document 2 can be applied to performance prediction of a cluster system for which past performance information is available. However, the clustering overhead targeted by the present invention cannot be predicted.

  The method of Patent Document 3 is a method for realizing various workflows with one system component connection model, and is not a method for predicting the clustering overhead targeted by the present invention.

  The method of Patent Document 4 is a method of confirming the load distribution amount for parallelization with a numerical value, but is not a method of predicting the clustering overhead targeted by the present invention.

  The technologies described in Non-Patent Document 1 and Non-Patent Document 2 can predict the performance of a cluster system in which an application that can calculate the calculation amount and the communication amount from an algorithm operates. However, the clustering overhead cannot be predicted in the case of a cluster system in which a business application operates and the calculation amount and communication amount cannot be calculated from the algorithm.

  An object of the present invention is to solve the above-described drawbacks of the prior art and to predict the performance of a cluster system that can predict the clustering overhead that occurs in a cluster system that shares data on the main memory of each node via a communication line. An apparatus, a performance prediction method, and a performance prediction program are provided.

  In order to achieve the above object, the present invention provides a performance prediction device for a cluster system in which data on a main storage device of a plurality of nodes connected via a communication line is shared between the nodes. Calculate the number of block copies made between the nodes in units of memory blocks in the device, and based on the block copy number per unit time and the CPU load information per block copy, due to the block copy The resulting clustering overhead of the cluster system is predicted.

  The present invention focuses on the fact that block copying for data sharing is a major cause of clustering overhead in a cluster system in which data on the main memory is shared among nodes.

  In the present invention, first, copy cost information, workload characteristic information, and load distribution characteristic information are input to the cluster system performance prediction apparatus.

  The cluster system performance prediction apparatus generates a copy cost, a workload characteristic table, and a load distribution characteristic table from these pieces of information.

  Copy cost information is CPU load information when the block copy is performed once.

  The workload characteristic table is a table including information on the number of appearances of each window size in each memory block set. Here, the window size is the number of occurrences of a series of accesses in a window which is a series of accesses from the time when a specific type of access is made to the memory block set until the next type of access is made. .

  The load distribution characteristic table is a table including information on the arrival probability of access to the partial block set of each node.

  Next, the cluster system performance prediction device determines the number of nodes that access arrives in one window when the window size is given based on the node number information, window size information, and load distribution characteristic table information of the cluster system. The number of arrival nodes that is the expected value is calculated.

  Next, assuming that the access in the window is a write access followed by a read access a plurality of times, the number of block copies when the window size is given from the number of arrival nodes is calculated.

  Next, the block copy count of the entire cluster system is calculated from the workload copy table and the block copy count given the window size.

  Finally, the clustering overhead is calculated from the block copy count and copy cost of the entire cluster system.

  According to the present invention, the clustering overhead generated in the cluster system can be predicted.

  The reason is to calculate the overhead generated by block copy, which is the main factor of clustering overhead, based on the workload characteristics, load distribution characteristics and copy cost of the cluster system.

  Hereinafter, the best mode for carrying out the present invention will be described.

(First embodiment)
Hereinafter, a first embodiment of the present invention will be described with reference to the drawings.

  FIG. 1 is a block diagram illustrating a configuration of a performance prediction apparatus 10 of a cluster system according to the present embodiment.

  Referring to FIG. 1, the cluster system performance prediction apparatus 10 according to the present exemplary embodiment includes an input unit 110, a copy cost generation unit 120, a workload characteristic generation unit 130, a load distribution characteristic generation unit 140, an arrival node. Number calculation means 150, overhead prediction means 160, and output means 170 are provided.

  Before describing each means of the performance prediction device 10 of the cluster system, the cluster system 40 that is a target of the present embodiment will be described.

  FIG. 2 is a diagram illustrating an example of the cluster system 40 that is a target of the present embodiment.

  Referring to FIG. 2, a cluster system 40 that is a target of the present embodiment includes two nodes connected via a communication line. Each node includes a CPU (Central Processing Unit) and a main storage device.

  A client is connected to the cluster system 40 via a communication line. The client performs “write” access (write access) for writing information to the cluster system 40 or “read” access (read access) for reading. The cluster system 40 connects a client to an available node.

  When there is a “read” access from the client and the access destination node detects that the partial block set in the main storage device read by the client is not the latest, a block copy that copies the data of the latest partial block set from another node appear. If it is detected that the block read by the client is the latest in the access destination node, block copy does not occur. The partial block set is a block accessed by the client, and details will be described later.

  As described above, between nodes, the CPU is consumed by performing block copy in units of partial block sets between main storage devices, and clustering overhead occurs. Although the clustering overhead is caused by a plurality of causes, it is known that the main cause is a CPU load caused by block copy for data sharing. The same applies even if the number of nodes is three or more.

  Next, each unit of the cluster system performance prediction apparatus 10 will be described.

  The input unit 110 is a unit for inputting information of the cluster system 40 that is a target of performance prediction, and has a function of inputting the overhead basic information 20 to the performance prediction apparatus 10 of the cluster system. The overhead basic information 20 includes copy cost information 210, workload characteristic information 220, and load distribution characteristic information 230.

  The copy cost generation unit 120 has a function of reading the copy cost information 210 from the overhead basic information 20 and generating the copy cost table 31.

  Here, the copy cost is one CPU load when a block copy occurs once. The CPU load is represented by CPU execution time per unit time, for example.

  The workload characteristic generation unit 130 has a function of reading the workload characteristic information 220 included in the overhead basic information 20 and generating the workload characteristic table 32.

  The load distribution characteristic generation unit 140 has a function of reading the load distribution characteristic information 230 included in the overhead basic information 20 and generating the load distribution characteristic table 33.

  The arrival node number calculating means 150 has a function of calculating the number of nodes where access to the partial block set arrives from the load distribution characteristic and the workload characteristic.

  Here, the partial block set is a set of memory blocks including one or more memory blocks. A partial block set having the same ID (identification number) is arranged in the main storage device of each node and is shared among the nodes.

  The partial block set is set in the cache memory of the main storage device of each node, and is controlled so as to maintain the same information between the nodes. However, even in a partial block set with the same ID, there is a case where data does not match between all nodes due to reasons such as no access from the client.

  The overhead prediction unit 160 is a unit that calculates the clustering overhead of the cluster system 40 and has a function of calculating the clustering overhead and scalability from the number of arrival nodes.

  The output unit 170 is a unit that outputs a result, and outputs the predicted clustering overhead and scalability.

  The cluster system performance prediction device 10 can be configured by a data processing device, a computer, or the like. In this case, a keyboard, a mouse, or the like can be used as the input unit 110. As the output unit 170, a printer, a display, or the like can be used.

  Next, a usage form of the performance prediction apparatus 10 of the cluster system according to the present embodiment will be described.

  Referring to FIG. 2, the performance prediction apparatus 10 of the cluster system according to this embodiment is connected to a node of the cluster system 40 via a communication line. The performance prediction device 10 of the cluster system acquires the overhead basic information 20 from the node and performs performance prediction of the cluster system 40.

  Note that the cluster system performance prediction apparatus 10 does not necessarily have to be connected to a node of the cluster system 40. When the cluster system performance prediction apparatus 10 is not connected to the cluster system 40, the clustering overhead basic information 20 of the cluster system 40 is input to the cluster system performance prediction apparatus 10 manually or by a storage medium.

  Further, the performance prediction device 10 of the cluster system may be mounted in the node. In such a case, the overhead basic information 20 is input from the node to the performance prediction apparatus 10 of the cluster system.

  As for the function of calculating the clustering overhead based on the workload characteristic information 210, the load distribution characteristic information 220, and the copy cost information 230, which is a feature of the present invention, a program for realizing such a function is provided in the computer apparatus. It can also be realized by mounting the incorporated circuit components. However, it is also possible to store the program (application) for realizing the characteristic function of the present invention in a storage medium and execute the program on a computer device, thereby functioning as the performance prediction device 10 of the cluster system.

  FIG. 3 is a block diagram showing a configuration of the contents of the characteristic table 30 according to the present embodiment.

  Referring to FIG. 3, the characteristic table 30 includes a copy cost table 31, a workload characteristic table 32, and a load distribution characteristic table 33.

  A copy cost is written in the copy cost table 31. The information entered in the copy cost table 31 is only one copy cost.

  The workload characteristic table 32 is a table of the number of appearances of windows in each partial block set, and the number of appearances of windows is written for each window size. Here, the window refers to a set of accesses (all accesses) from when a specific type of access is made to the memory block set by the client until the next same type of access is made. The window size represents the number of accesses constituting the window. Details of the window size will be described with reference to FIG.

  The load distribution characteristic table 33 is a table of the arrival probability of access to each partial block set of a node.

  The workload characteristic table 32 and the load distribution characteristic table 33 will be described in detail with reference to FIGS.

  Next, the operation of the present embodiment will be described in detail.

  FIG. 4 is a flowchart for explaining the operation of the performance prediction apparatus 10 of the cluster system according to this embodiment. In the following description, reference is made to the main part of FIG. 1 as necessary.

  Referring to FIG. 4, first, the input means 110 inputs the overhead basic information 20 (step A1).

  The overhead basic information 20 may be input manually or may be input from a node of the cluster system 40 to the performance prediction apparatus 10 of the cluster system via a communication line.

  Next, the copy cost generation means 120 generates the copy cost table 31 from the copy cost information 210 included in the overhead basic information 20 (step A2).

  In this embodiment, it is assumed that what percentage of one CPU is consumed in one block copy is given as copy cost information 210. Therefore, the copy cost generation unit 120 writes the copy cost information 210 into the copy cost table 31 as it is.

  The workload characteristic generation unit 130 generates the workload characteristic table 32 from the workload characteristic information 220 included in the overhead basic information 20 (step A3).

  In this embodiment, since the workload characteristic information 220 is given as the workload characteristic table 32, the workload characteristic generation unit 130 writes the workload characteristic information 22 into the workload characteristic table 32 as it is.

  The load distribution characteristic acquisition unit 140 generates the load distribution characteristic table 33 from the load distribution characteristic information 230 included in the overhead basic information 20 (step A4).

  In this embodiment, since the load distribution characteristic information 230 is given as the load distribution characteristic table 33, the load distribution characteristic acquisition unit 140 writes the load distribution characteristic information 230 into the load distribution characteristic table 33 as it is.

  Among the steps A1 to A4 described above, the order of steps A2, A3, and A4 can be changed.

  FIG. 5 is a diagram showing a workload characteristic table example 32A according to the present embodiment.

  Referring to FIG. 5, in the workload characteristic table example 32A, the number of appearances of windows in each partial block set is shown for each window size. Note that the partial block set is arranged in all nodes.

  The characters A, B, and C shown in each row are the IDs of each partial block set, and the numbers 3, 4, 5, and 7 shown in each column represent the window sizes.

  From the workload characteristic table example 32A in FIG. 5, for example, it is understood that a window of size 5 appears 800 times when accessing the partial block set B.

  FIG. 6 is a diagram showing a load distribution characteristic table example 33A according to the present embodiment.

  Referring to FIG. 6, the load distribution characteristic table example 33 </ b> A shows the probability of access reaching each of the nodes 1, 2, 3, and 4 for each of the partial block sets of A, B, and C. From the load distribution characteristic table example 33A, the access to the partial block set A is 0.1 for the node 1, 0.3 for the node 2, 0.2 for the node 3, and the node 4 It can be seen that there is a probability of 0.4.

  In the present embodiment, it is assumed that two types of access are performed: a “write” access that rewrites the contents of a block from a client to a node and a “read” access that reads the contents of the latest block. The workload characteristic table represents a load caused by a sequence operation in which data written by the “write” access is read by the “read” access. For example, when the window size is 4, it means that an access sequence “write, read, read, read, (write)” appears. Here, (write) at the end of the access sequence is “write” access, but the content is rewritten by the last “(write)” access, and therefore, it is shown in parentheses. The window size is represented by the number of accesses from “write” access to “read” access immediately before the next “write” access.

  The arrival node number calculation means 150 calculates the arrival node number so that the arrival node number monotonously increases in a broad sense as the window size increases with respect to the given load distribution characteristic table 33 (step A5).

  The number of arrival nodes is calculated for each partial block set.

  Here, monotonically increasing in a broad sense means that the number of arrival nodes increases or remains unchanged when the window size increases. In addition, the number of arrival nodes represents an expected value of the number of nodes to which access arrives in one window.

  The reason why the number of arriving nodes is calculated so as to increase monotonically in a broad sense with respect to the window size is that the number of times of accessing the node increases as the window size increases, so that the number of arriving nodes increases or does not change. This is because the number of arrival nodes does not decrease.

  For example, consider a case where an access to a partial block set A occurs with a window size of 5 under a certain load distribution characteristic table 33. If the arrival node number calculation means 150 calculates that these five accesses arrive at three nodes (that is, the arrival node number is 3), the arrival node number calculation means 150 determines that the arrival node number is 3 when the window size is 6. Calculated as above.

  An example of a method for calculating the number of arrival nodes when a cluster size n and a window size w are given will be described. Each node ID (node identification number) is represented by a number from 1 to n. As described above, the number of arrival nodes is calculated for each partial block set.

The probability of access to the node i arrives and _{p i (Σ p i = 1} ). p _i can be acquired from the load distribution characteristic table 33.

Also, let d _i be the number of accesses that have arrived at node i within a window. here

d = (d ₁ , d ₂ ,..., d _n ) (Formula 1)

p = (p ₁ , p ₂ ,..., p _n ) (Expression 2)

Write.

で与えられる。式３でΠは、積を計算する記号である。また、d₁, d₂,… d_nは

ｄ_１＋ｄ_２＋・・・ｄ_ｎ＝ｗ ・・・ （式４）

を満たす自然数である。ここで、自然数は０を含む。 Node ID is 1, 2, the probability that the n-number of events ... arrives access to a is each node n appears, p _1, p _2, respectively, as described above, is given by ... p _n. When the window size is accessed w times, the occurrence probability m (w, d, p) of the event in which the _n events occur d ₁ , d ₂ , ... d _n times is obtained by using the multinomial distribution.

Given in. In Equation 3, Π is a symbol for calculating the product. D ₁ , d ₂ , ... d _n are

_{d 1 + d 2 + ··· d} n = w ··· ( Equation 4)

It is a natural number that satisfies Here, the natural number includes 0.

  Next, calculation of the number of arrival nodes will be described. Here, the number of arrival nodes represents an expected value of the number of arrival nodes.

  First, the number of nodes arriving at the access (hereinafter, the number of nodes arriving at the access is abbreviated as the number of access arrival nodes) is fixed, the expected value of the number of arrival nodes is obtained, and then the number of arrival nodes for all the number of access arrival nodes Sum the expected values of. Hereinafter, the expected value of the number of arrival nodes when the number of access arrival nodes is fixed is described as a partial expected value of the number of arrival nodes.

Among d _i satisfying Equation 4, a set of d whose k is a natural number of 1 or more is described as D (n, w, k). Here, k represents the number of access arrival nodes.

  As an example, if the cluster size n is 5, the number of nodes is 5, but if k is 3, access will arrive at 3 of the 5 nodes, but 2 Access does not arrive at the node.

  First, the partial expected value of the number of arriving nodes is obtained. This partial expected value can be calculated by adding the value obtained by multiplying the probability of Equation 3 by k for the set of D (n, w, k). .

  Next, for all k, the partial expected value of the number of arrival nodes is added, but before that, the maximum value of k is determined.

  First, when the window size w is larger than the cluster size n, the access arrival node count k does not exceed the cluster size n (= node count). Further, when the window size w is smaller than the cluster size n, the access arrival node count k cannot exceed the window size w (= access count). The same applies when the window size w is equal to the cluster size n. Therefore, it can be seen that the maximum value of k is min (n, w).

a(n, w)は、１ウィンドウ中にアクセスが到着するノードの数の期待値になっている。なお、a(n, w) の値や m(w, d, p) の値は、予め計算して表として持っていてもよい。 From the above, the number of arrival nodes is calculated by the following equation.

a (n, w) is an expected value of the number of nodes to which access arrives in one window. The value of a (n, w) and the value of m (w, d, p) may be calculated in advance and stored as a table.

As an example of the calculation method of the arrival node number a (n, w) other than Equation 5,

a (n, w) = αn θ + βw φ + γ ··· ( Equation 6)

a (n, w) = αlogn + βlogw + γ (Expression 7)

and so on. Here, α, β, γ, θ, and φ are real numbers. In these formulas as well, the number of arrival nodes is calculated so as to increase monotonically in a broad sense as the window size increases.

  There is a method of using a multidimensional normal distribution instead of calculating a (n, w) from a multinomial distribution.

  As described above, the number of arrival nodes when the cluster size n and the window size w are given is calculated.

  The overhead prediction unit 160 calculates the number of block copies from the number of arrival nodes (step A6).

  The block copy count is the block copy count per unit time. The block copy count calculated in step A6 is the block copy count when the window size w is fixed. Hereinafter, the block copy count when the window size w is fixed is described as the partial block copy count.

As an example of a method of calculating the partial block copy count b (n, w) when the load distribution characteristic table 33, cluster size, and window size are specified,

b (n, w) = a (n, w) −1 (Expression 8)

There is. b (n, w) is equal to the number of nodes where only “read” access has arrived. Equation 8 is a calculation method that assumes that after the partial block set is updated by “write” access, block copy occurs when “read” access arrives for the first time at a node to which “write” access has not arrived. is there.

  As described at the beginning of the present embodiment, since “write” access is performed by the client, the contents of the partial block set that has been written are not the latest. Therefore, when the “read” access arrives for the first time, it is detected that the partial block set in the main storage device read by the client is not the latest, and the block copy that copies the latest partial block set data from another node Will occur.

  For example, if the number of nodes is 5 and the window size is 3, the access constituting the window is “write, read, read”, but the first “write” access arrives. If the “read” access arrives at one node and the “read” access arrives at another node among the four nodes that did not exist, a (n, w) is 3, but b (n, w) represents 2;

  Also, in this example, if two “read” accesses arrive at the same node, the first “read” access causes a block copy to make the partial block set up-to-date. Block read does not occur in "read" access. In this case, a (n, w) is 2 and b (n, w) is 1.

  Next, the total number of partial block copies for all window sizes is calculated to calculate the number of block copies generated for the partial block set.

とする。b(n, W) は、部分ブロック集合に対して発生するブロックコピー回数を表している。 Let W be the set of non-zero window sizes, f (w) be the number of occurrences of window size w,

And b (n, W) represents the number of block copies generated for the partial block set.

  Next, the total number of block copies is calculated by summing up the number of block copies for each partial block set.

と算出する。ここで、ブロックコピー回数は単位時間当たりのブロックコピー回数である。 Each partial block set appearing in the characteristic table 30 is defined as S _i.

b (S _i ) = b (n, W) (Equation 10)

And Also, let S _i ∈S. Here, S is a set of each partial block set S _i . Here, in order to acquire all the data of the workload characteristic table 32, the period during which the window information is measured is described as period. And the block copy count B (S) is

And calculate. Here, the number of block copies is the number of block copies per unit time.

により、ブロックコピー回数からクラスタリング・オーバーヘッドとスケーラビリティを算出する（ステップＡ７）。 The overhead predictor 160 describes the copy cost as copy, the clustering overhead as overhead, and the scalability as scalability.

Thus, the clustering overhead and scalability are calculated from the number of block copies (step A7).

  Finally, the output unit 170 outputs the clustering overhead and scalability (step A8).

  Next, the hardware configuration of the performance prediction device 10 of the cluster system will be described.

  FIG. 7 is a block diagram showing a hardware configuration of the performance prediction apparatus 10 of the cluster system according to this embodiment.

  Referring to FIG. 7, a cluster system performance prediction apparatus 10 includes a CPU 101, a main storage unit 102 including a RAM and a ROM, a communication unit 103 that performs data communication via a communication line, an output unit 104 such as a printer and a display, An input unit 105 such as a keyboard and a mouse and an auxiliary storage unit 106 such as a hard disk device are provided.

As described above, the performance prediction device 10 of the cluster system according to the present embodiment is
A device for predicting the performance of a cluster system 40 that shares data on main storage devices of a plurality of nodes connected to each other via a communication line between the nodes. Calculates the number of block copies per unit time made in step 1, and calculates the clustering overhead that occurs in the cluster system based on the information on the number of block copies and the information on the CPU load when one block copy is made It is.

(Effects of the first embodiment)
According to the embodiment described above, the clustering overhead generated in the cluster system 40 that shares data via the communication line can be predicted.

  This is because the overhead generated by block copy, which is the main factor of clustering overhead, is calculated from the workload characteristics, load distribution characteristics, and copy cost of the cluster system 40.

  In addition, the scalability that occurs in the cluster system 40 described above can be predicted.

  This is because the scalability is calculated from the clustering overhead on the assumption that the determining factor of the scalability of the cluster system is the clustering overhead.

(Second Embodiment)
The second embodiment of the present invention will be described below with reference to the drawings.

  FIG. 8 is a block diagram showing a configuration of the performance prediction apparatus 11 of the cluster system according to the present embodiment.

  Referring to FIG. 8, the configuration of the cluster system performance prediction apparatus 11 according to the present embodiment is the same as the configuration of the cluster system performance prediction apparatus 10 according to the first embodiment shown in FIG. However, the function of the load distribution characteristic generation unit 141 and the function of the arrival node number calculation unit 151 are different from the function of the load distribution characteristic generation unit 140 and the function of the arrival node number calculation unit 150 according to the first embodiment, respectively. It has become. Further, the load distribution characteristic information 231 included in the overhead basic information 21 is also different from the load distribution characteristic information 230 included in the overhead basic information 20 according to the first embodiment.

  The function of the load distribution characteristic generation unit 141, the function of the arrival node number calculation unit 151, and the load distribution characteristic information 231 will be described with reference to FIG.

  FIG. 9 is a diagram showing an example of the transition characteristic table set 36 according to the present embodiment. In this embodiment, a transition characteristic table set 36 is used instead of the load distribution characteristic table 33 of FIG. 3 described in the first embodiment.

  The transition characteristic table set 36 includes transition characteristic tables 36A, 36B,..., And access information related to the partial block sets A, B,... Is written in the transition characteristic tables 36A, 36B,.

  Referring to FIG. 9, the transition characteristic table is represented as a two-dimensional table for convenience. However, in general, a transition characteristic table (not shown) is a table of k-dimensional node IDs (identification numbers). In the first embodiment, k represents the number of access arrival nodes, but in the present embodiment, k is a natural number representing the dimension of the node ID table.

  First, an expression for calculating the probability of occurrence of an access sequence to a node will be described.

Access sequence _{(v 1, v 2, ...} , v k) the occurrence probability q of _{(v 1, v 2, ...} v k) represented by. Here, in the access sequence (v ₁ , v ₂ ,..., V _k ), the node ID accessed last time is v ₁ , the node ID accessed last time is v ₂ , and the node ID accessed k times before is v represents an access sequence that is _k .

Also, suppose that v _i is a node ID accessed i times before, and i is a natural number of k −1 or less.

により算出できる。 The appearance probability of the access sequence (v ₁ , v ₂ ,…, v _i ) is

Can be calculated.

  Expression 14 is an expression for calculating the appearance probability of the access sequence to the node from the previous time to i times before from the appearance probability of the access sequence to the node from i to the previous time i + 1 times or more before. When i is k-1, Equation 14 calculates the appearance probability of the access sequence from the previous time to k-1 times from the previous probability of the access sequence to the node from the previous time to k times.

  FIG. 10 is a diagram showing a two-dimensional transition characteristic table example 37 in the performance prediction device 11 of the cluster system according to the present embodiment.

  In the transition characteristic table example 37, it is assumed that the cluster system 40 includes four nodes. The node IDs 1 to 4 at the left end of each row represent the IDs of the nodes where the previous access occurred, and the node IDs 1 to 4 at the upper end of each column represent the IDs of the nodes where the current access occurred.

  In the following, the access sequence (n) represents the probability of an access sequence in which the current node ID is n. Also, the access sequence (m, n) represents the appearance probability of an access sequence in which the current node ID is m and the previous node ID is n.

  Referring to FIG. 10, it can be seen that access sequence (1) occurs with probability 0.25 and access sequence (2) occurs with probability 0.1. For example, since the access sequence (1) is the probability that the previous node ID is 1, the value is calculated by adding the numerical value of the row where the left end of the row is 1 in the example of the transition characteristic table 37 to the side.

  It can also be seen that the sequence (4, 3) consisting of two accesses occurs with a probability of 0.1.

  In the two-dimensional transition characteristic table, for example, a load distribution method that can determine the probability of which node the current access arrives from the node that the previous access arrived, or a static load such as round robin Describe the distribution method.

  In the k-dimensional transition characteristics table, a load balancing method that can determine the probability of which node the current access arrives from the node that has reached the access up to (k −1) times before, and the way It is also possible to describe a static load distribution method such as Ted Round Robin that statically determines the next access arrival node depending on not only the previous access arrival node but also the previous access arrival.

  The function that describes the load distribution characteristics in the transition characteristic table 36 that specifies the access sequence to the node where the block copy occurs, which is a feature of the present invention, and obtains the number of times of block copying is provided inside the computer device. It is also possible to implement by mounting a circuit component incorporating a program for realizing the above. However, it is also possible to store the program (application) for realizing the characteristic functions of the present invention in a storage medium and execute the program on a computer device, thereby functioning as the performance prediction device 11 of the cluster system.

  Next, the operation of the present embodiment will be described.

  FIG. 11 is a flowchart for explaining the operation of the performance prediction apparatus 11 of the cluster system according to this embodiment.

  In the operation of this embodiment, the steps B1 to B3 and the steps B6 to B8 are the same as the steps A1 to A3 and the steps A6 to A8 of the first embodiment described in FIG. However, the steps B4 and B5 are different from the steps A4 and A5. Hereinafter, steps B4 and B5 will be described.

  The load distribution characteristic creating unit 141 generates a transition characteristic table set 36 (step B4).

  However, in this embodiment, since the transition characteristic table set 36 is given as the load distribution characteristic information 231, the load distribution characteristic creation unit 141 writes the load distribution characteristic information 231 as it is in the transition characteristic table set 36.

  The arrival node number calculating means 151 calculates the number of arrival nodes from the occurrence probability of the access sequence (step B5).

When the w-dimensional space of the node ID is D ^w , w access sequences can be represented by a vector v ∈ D ^w . Here, the vector v is a w-dimensional vector and represents an access sequence. Let the i-th component of v be v _i, and let v _{i be} the node ID that has been accessed i times before. Also,

q (v) = q (v ₁ , v ₂ ,..., v _w ) (Equation 15)

And

However, if the dimension k of the transition characteristic table is smaller than w,

q (v) = q (v ₁ ,..., v _k ) × q (v _{k + 1} ,..., v _2k ) × q (v _{hk + 1} ,..., v _w ). (Formula 16)

It becomes. Here, h is a positive or 0 integer.

  At this time, the arrival node number calculating means 150 calculates the arrival node number a (w) for each access sequence v and window size w. However, note that the number of arrival nodes increases monotonically in a broad sense as w increases.

a(w) は、第１の実施の形態における a(n, w) に相当する。Card(v)はアクセスシーケンス vにおける、ノードＩＤの数を表す。ノードＩＤの数は、v の成分に同じノードが出現した場合には１であると計数し、２つのノードが出現した場合には２であると計数する。 An example of a method for calculating the number of arrival nodes is shown in the following equation. The number of arrival nodes calculated by the following formula is an expected value of the number of nodes to which access arrives in one window when the window size is w.

a (w) corresponds to a (n, w) in the first embodiment. Card (v) represents the number of node IDs in the access sequence v. The number of node IDs is counted as 1 when the same node appears in the component of v, and is counted as 2 when two nodes appear.

(Effects of the second embodiment)
According to the embodiment described above, the block copy count can be calculated more accurately than in the first embodiment.

  The reason is that many load distribution methods can be described more accurately and in detail because the load distribution characteristics are described in the transition characteristic table 36 that specifies the access sequence to the node where the access occurs.

(Third embodiment)
The third embodiment of the present invention will be described below with reference to the drawings.

  FIG. 12 is a block diagram showing a configuration of the performance prediction device 12 of the cluster system according to the present embodiment.

  Referring to FIG. 12, the configuration of the cluster system performance prediction apparatus 12 according to the present embodiment is the same as the configuration of the cluster system performance prediction apparatus 10 according to the first embodiment shown in FIG. However, the function of the workload characteristic generation unit 131 and the function of the arrival node number calculation unit 151 are different from the functions of the workload characteristic generation unit 130 and the arrival node number calculation unit 150 according to the first embodiment, respectively. Yes. Further, the workload characteristic information 221 included in the overhead basic information 22 is different from the workload characteristic information 220 included in the overhead basic information 20 according to the first embodiment.

  In the present embodiment, the workload characteristic generation unit 131 has a function of converting the workload characteristic information 221 into the workload characteristic table 32. The arrival node number calculation means 151 has a function of calculating the arrival node number even when the window size is not a natural number.

  In the first embodiment, the workload characteristic information 220 is given in the workload characteristic table. In the present embodiment, the workload characteristic information 221 is given as information for generating the workload characteristic table 32. Specifically, the workload characteristic information 221 is given by the number of accesses for each access type arriving at the partial block set, that is, the number of “read” accesses and the number of “write” accesses. The workload characteristic information 221 is converted into the workload characteristic table 32 by the workload characteristic generation unit 131.

  Next, the operation of the present embodiment will be described.

  FIG. 13 is a flowchart for explaining the operation of the performance prediction device 12 of the cluster system according to this embodiment.

  In the operation of this embodiment, the steps C1, C2, C4, and C6 to C8 are the same as the steps A1, A2, A4, and A6 to A8 of the first embodiment described in FIG. However, the steps C3 and C5 are different from the steps A3 and A5. Hereinafter, steps C3 and C5 will be described.

  In the first embodiment, the operation of the workload characteristic generation unit 13 in step A3 only writes the workload characteristic information 220 to the workload characteristic table 32.

  In step A5, the arrival node number calculating means 150 calculates the number of arrival nodes only from the natural window size.

  In this embodiment, in step C3, the workload characteristic generation unit 131 calculates the window size and the appearance frequency from the number of “read” accesses and the number of “write” accesses, and generates the workload characteristic table 32 ( Step C3).

  In step C3, the workload characteristic generation unit 131 sets the window size to a value obtained by dividing the sum of the number of “read” accesses and the number of “write” accesses by the number of “write” accesses, and the appearance frequency of the window size. Is written in the workload characteristic table 32 as the number of “write” accesses. This calculation method can be calculated particularly accurately when it is assumed that all “write” accesses arrive at equal intervals.

及び

であるときの到着ノード数から w に対する到着ノード数を算出する（ステップＣ５）。 In step C5, in order to be able to calculate the number of arrival nodes from a positive real window size w which is not a natural number, the arrival node number calculation means 160 has a window size of

as well as

The number of arriving nodes for w is calculated from the number of arriving nodes at step S5 (step C5).

は、それぞれ w 以下の最大の整数、w 以上の最小の整数を表す記号である。 here,

Are symbols representing the largest integer less than or equal to w and the smallest integer greater than or equal to w.

上記の式は、

と

の間で線形補間を行っているが、他の一般に知られた補間方法を用いてもよい。なお、式１８及び式１９で h の記号が使用されているが、第２の実施の形態における式１６に記載された h とは、別の量を表している。 In step C5, for example, the arrival node number calculating means 151 calculates the number of arrival nodes for a positive real window size by the following equation.

The above formula is

When

However, other generally known interpolation methods may be used. Note that the symbol h is used in Equation 18 and Equation 19, but h described in Equation 16 in the second embodiment represents a different amount.

  The subsequent steps after C6 are the same as those after A6 in the first embodiment, and the clustering overhead and scalability can be calculated using the number of arrival nodes calculated by linear interpolation.

  Even if the window size w obtained from the number of “read” accesses and the number of “write” accesses, which is a feature of the present invention, is not an integer, the number of arrival nodes when the window size is the largest integer equal to or less than w; For the function to calculate the number of arrival nodes for w by interpolation from the number of arrival nodes when the minimum integer is greater than or equal to w, a circuit component incorporating a program that implements such a function is installed inside the computer device. It can also be realized. However, it is also possible to store the program (application) for realizing the characteristic function of the present invention in a storage medium and execute the program on the computer device, thereby functioning as the performance prediction device 12 of the cluster system.

(Effect of the third embodiment)
According to the embodiment described above, the clustering overhead and scalability of the cluster system 40 can be predicted even when the workload characteristic information 221 is given by the number of “read” accesses and the number of “write” accesses.

  The reason is that even if the window size w obtained from the number of “read” accesses and the number of “write” accesses is not an integer, the number of nodes that arrive when the window size is the largest integer less than or equal to w This is because the number of arrival nodes for w is calculated from the number of arrival nodes when the number is the smallest integer.

(Fourth embodiment)
The fourth embodiment of the present invention will be described below with reference to the drawings.

  In the first embodiment, the second embodiment, and the third embodiment described above, there are two types of access from the client, “read” and “write”. However, in the present embodiment, clustering overhead and scalability prediction when there are three types of access, “read”, “write”, and “get” will be described.

  In the “get” access (acquisition access) described in the present embodiment, a block is read in the same manner as in the “read” access, and when a block that does not have the latest block is reached, a block copy is generated. Unlike the case where the client accesses the node that accessed “write” or the node that accessed “read”, even if the node that accessed “get” is accessed, the node can obtain information that the current partial block set is the latest. I can't.

  Therefore, block copy occurs when “get” access arrives at a node where only “get” access has arrived in the past. The reason why this happens is that the node that accessed "get" cannot obtain information about whether it is the latest, so in "get" access, it is determined that a block that does not have the latest block has arrived, and a block copy occurs. Because. On the other hand, even if “get” access arrives at a node where “write” or “read” access has occurred in the past, block copy does not occur. In the above description, the past description means that the time is traced back within one “write” access window including the determination time for a node that is to determine whether or not block copy occurs.

  FIG. 14 is a block diagram illustrating a configuration of the performance prediction apparatus 13 of the cluster system according to the present embodiment.

  Referring to FIG. 14, the configuration of the cluster system performance prediction apparatus 13 according to the present embodiment is the same as the configuration of the cluster system performance prediction apparatus 10 according to the first embodiment shown in FIG. However, the functions of the copy cost generation unit 121 and the workload characteristic generation unit 132 are different from the functions of the copy cost generation unit 120 and the workload characteristic generation unit 130 according to the first embodiment. Yes. The copy cost information 211 and the workload characteristic information 222 included in the overhead basic information 23 are different from the copy cost information 210 and the workload characteristic information 220 included in the overhead basic information 20 according to the first embodiment.

  The function of the copy cost generation unit 121, the function of the workload characteristic generation unit 132, the copy cost information 211, and the workload characteristic information 222 will be described with reference to FIG.

  FIG. 15 is a block diagram showing a configuration of contents of the characteristic table 35 according to the present embodiment.

  Referring to FIG. 15, the characteristic table 35 according to the present embodiment is based on a copy cost table 31, a workload characteristic table 32, a load distribution characteristic table 33, a copy cost table (GET) 38, and a workload characteristic table (GET) 39. It has a configuration. Details of the configuration of the workload characteristic table 32 and the load distribution characteristic table 33 are not shown, but the configuration of these tables is the same as that shown in FIG.

  In the characteristic table 35 according to the present embodiment, a copy cost table (GET) 38 and a workload characteristic table (GET) 39 are added to the characteristic table 30 of the first embodiment.

  The copy cost (GET) written in the copy cost table (GET) 38 represents a rate at which one CPU is consumed when one block copy by “get” access occurs. The ratio of CPU consumption is represented by CPU execution time per unit time, for example, as in the first embodiment.

  Similar to the workload characteristics table 32, the workload characteristics table (GET) 39 is a table in which the partial block set ID is arranged in each row and the window size of “write” access related to “read” access is arranged in each column.

  In the present embodiment, the window size of “write” access related to “read” access represents the window size of an access sequence such as “write, read, read, read, (write)”. Note that “get” access may be sandwiched between access sequences, but in this case, the window size does not include the number of “get” accesses.

  In the section (record) written as window size (write & read) in the workload characteristic table (GET) 39, the window size of “write” access or “read” access related to “get” access is recorded. Here, the window size of “write” access or “read” access related to “get” access means “get” access window for “read” access that sandwiches “get” access between access sequences. The window size of “write” access related to “access” or the window size of “read” access related to “get” access. Here, the window size of “write” access related to “get” access represents the window size of an access sequence such as “write, get, get, (read)”.

  For example, in the access sequence “write, get, get, read, get, get, read, get, get, read, get, get, (write)”, the window size of “write” access related to “read” access is 4 The window size of “write” access or “read” access relating to “get” access is 3.

  In the workload characteristic table (GET) 39 where the above-described sequence occurs, 3 is written in a section where the row of the corresponding partial block set ID and the column where the window size (write) is written as 4. .

  In the above access sequence, the window size of “write” access related to “get” access and the window size of “read” access are both 3. However, if the window size for “get” access does not match between the window for “write” access and the window for “read” access, one representative value is assigned to “write” access and “read” for “get” access. Use as window size for access. W ′ calculated by Expression 20 described later is an example of such a representative value.

  Regarding the function of calculating the clustering overhead from the number of block copies of “read” access and the number of block copies of “get” access, which is a feature of the present invention, a program for realizing such a function is incorporated in the computer device. It can also be realized by mounting circuit components. However, it is also possible to store the program (application) for realizing the characteristic function of the present invention in a storage medium and execute the program on the computer device, thereby functioning as the performance prediction device 13 of the cluster system.

  Next, the operation of the present embodiment will be described.

  FIG. 16 is a flowchart for explaining the operation of the performance prediction device 13 of the cluster system according to this embodiment. In the following description, reference is made to the main part of FIG. 14 as necessary.

  Referring to FIG. 16, the flowchart of the present embodiment is obtained by adding steps D3, D5, D8, D10, and D12 to the flowchart of the first embodiment shown in FIG. These steps are steps for calculating the number of block copies by “get” access.

  First, the input means 110 inputs the overhead basic information 23 (step D1).

  According to the present embodiment, the copy cost information 211 of the overhead basic information 23 includes information on the copy cost caused by block copy by “read” access and the copy cost caused by block copy by “get” access.

  In addition to the workload characteristics related to “read” access, the workload characteristics information 222 includes the number of “get” accesses for each partial block set and the sum of the numbers of “write” accesses and “read” accesses. Suppose that Below, the number of “get” accesses for each partial block set is represented by g, and the sum of the numbers of “write” accesses and “read” accesses is represented by wr.

  Next, the copy cost generation unit 121 generates the copy cost table 31 from the copy cost information 211 included in the overhead basic information 23 (step D2). Step D2 corresponds to step A2 in FIG.

  Next, the copy cost generation unit 121 generates a copy cost table (GET) 38 from the copy cost information 211 included in the overhead basic information 23 (step D3).

  In the present embodiment, the copy cost that occurs in one block copy of “get” access is included in the copy cost information 211 as a copy cost (GET). Therefore, the copy cost generation unit 121 writes the copy cost (GET) from the copy cost information 211 into the copy cost table (GET) 38.

  Next, the workload characteristic generation unit 132 generates the workload characteristic table 32 from the workload characteristic information 222 included in the overhead basic information 23 (step D4).

  Step D4 corresponds to step A3 in FIG.

Next, the workload characteristic generation unit 132 sets, for each partial block set, a window size w ′ related to “write” access and “read” access related to “get” access.

w ′ = (g + wr) / wr (Equation 20)

Are written in all the records in the row of the partial block set targeted by the workload characteristic table (GET) 39 (step D5).

  The load distribution characteristic generation unit 140 generates the load distribution characteristic table 33 from the load distribution characteristic information 230 included in the overhead basic information 23 (step D6).

  Step D6 corresponds to step A4 in FIG.

  The arrival node number calculation means 150 calculates the number of arrival nodes of “read” access so that the number of arrival nodes monotonically increases (in a broad sense) with respect to the window size, for the given load distribution characteristic table 33 ( Step D7).

  Step D7 corresponds to step A5 in FIG.

  The arrival node number calculating means 150 calculates the number of arrival nodes of “get” access in the “write” access window related to “read” access from the workload characteristic table 22 and the workload characteristic table (GET) 39. However, unlike the case of the number of arrival nodes for “read” access, the number of arrival nodes for “get” access indicates that “write” access or “read” access has arrived in the past in one window. The number of arrival nodes is calculated assuming that the number of “get” accesses that have arrived at a node that does not exist (step D8).

  The number of arrival nodes for “get” access is the cluster size n, the window size w for “write” access for “read” access, and the window size for write access or “read” access for “get” access g (w) Calculate from In the following, the window size of “write” access or “read” access related to “get” access is represented by g (w).

  g (w) is obtained from the workload characteristic table (GET) 39. Then, for the given load distribution characteristic table 33, the number of arrival nodes a ′ (n, w) of “get” access is calculated so as to increase monotonously with respect to w and g (w).

オーバーヘッド予測手段１６０は、「read」アクセスの到着ノード数からブロックコピー回数を算出する（ステップＤ９）。 An example of a calculation method of the number of arrival nodes a ′ (n, w) of “get” access when the window size w of “write” access is given is shown in the following formula.

The overhead prediction unit 160 calculates the number of block copies from the number of arrival nodes of “read” access (step D9).

  Step D9 corresponds to step A6 in FIG.

  The overhead prediction unit 160 calculates the number of block copies (GET) from the number of arrival nodes of “get” access (step D10).

とする。B'(n,W) は、部分ブロック集合に対して発生する「get」アクセスによるブロックコピー回数を表している。 Let W be the set of window sizes for “write” access with non-zero occurrences, and f (w) be the number of occurrences of window size w for “write” access.

And B ′ (n, W) represents the number of times of block copy by “get” access generated for the partial block set.

と算出する。ここで、ブロックコピー回数 B'(S)は単位時間当たりのブロックコピー回数である。 Here, each partial block set appearing in the characteristic table 35 is defined as S _i.

b ′ (S _i ) = b ′ (E, W) (Equation 23)

And Also, let S _i ∈ S. Then, the block copy count B '(S) generated by the "get" access is

And calculate. Here, the number of block copies B ′ (S) is the number of block copies per unit time.

  Next, the overhead prediction unit 160 calculates a clustering overhead from the number of block copies generated by the “read” access (step D11).

  Step D11 corresponds to step A7 in FIG.

を用いて、「get」アクセスによるクラスタリング・オーバーヘッドと、スケーラビリティを算出する（ステップＤ１２）。 Assuming that the copy cost (GET) by “get” access is copy (get), the clustering overhead by “get” access is overhead (get), and the scalability is scalability, the overhead predicting means 160

Is used to calculate the clustering overhead and scalability due to “get” access (step D12).

  Finally, the output unit 170 outputs the clustering overhead by the “read” access, the clustering overhead by the “get” access, and the scalability (step D13).

(Effect of the fourth embodiment)
According to the embodiment described above, the clustering overhead and scalability can be calculated even in the case of the cluster system 40 having “get” access in addition to “write” access and “read” access.

  This is because the clustering overhead is calculated from the number of block copies for “read” access and the number of block copies for “get” access.

  The performance prediction apparatuses 10, 11, 12, and 13 of the cluster system of the present invention not only realize the operation in hardware, but also execute a performance prediction program (application) 300 for executing each of the above-described means as a computer processing apparatus. It can be realized by software by being executed by the performance prediction apparatuses 10, 11, 12, and 13 of the cluster system. The performance prediction program 300 is stored in a magnetic disk, a semiconductor memory, or other recording medium, loaded from the recording medium to the performance prediction apparatuses 10, 11, 12, and 13 of the cluster system, and controls the operation thereof. Each function is realized.

  Although the present invention has been described above with a plurality of preferred embodiments, the present invention is not necessarily limited to the above-described embodiments, and various modifications may be made within the scope of the technical idea. Can do.

It is a block diagram which shows the structure of the performance prediction apparatus of the cluster system by the 1st Embodiment of this invention. It is a figure which shows an example of the cluster system made into the object of the 1st Embodiment of this invention. It is a block diagram which shows the structure of the content of the characteristic table by the 1st Embodiment of this invention. It is a flowchart for demonstrating operation | movement of the performance prediction apparatus of the cluster system by the 1st Embodiment of this invention. It is a figure which shows the example of a workload characteristic table by the 1st Embodiment of this invention. It is a figure which shows the example of a load distribution characteristic table by the 1st Embodiment of this invention. It is a block diagram which shows the hardware constitutions of the performance prediction apparatus of the cluster system by the 1st Embodiment of this invention. It is a block diagram which shows the structure of the performance prediction apparatus of the cluster system by the 2nd Embodiment of this invention. It is a figure which shows an example of the transition characteristic table set by the 2nd Embodiment of this invention. It is a figure which shows the example of a two-dimensional transition characteristic table in the performance prediction apparatus of the cluster system by the 2nd Embodiment of this invention. It is a flowchart for demonstrating operation | movement of the performance prediction system by the 2nd Embodiment of this invention. It is a block diagram which shows the structure of the performance prediction apparatus of the cluster system by the 3rd Embodiment of this invention. It is a flowchart for demonstrating operation | movement of the performance prediction apparatus of the cluster system by the 3rd Embodiment of this invention. It is a block diagram which shows the structure of the performance prediction apparatus of the cluster system by the 4th Embodiment of this invention. It is a block diagram which shows the structure of the content of the characteristic table by the 4th Embodiment of this invention. It is a flowchart for demonstrating operation | movement of the performance prediction apparatus of the cluster system by the 4th Embodiment of this invention.

Explanation of symbols

10, 11, 12, 13: Cluster system performance prediction device 20, 21, 22, 23: Overhead basic information 30, 35: Characteristic table 31: Copy cost table 32: Workload characteristic table 32A: Workload characteristic table example 33 : Load distribution characteristic table 33A: Load distribution characteristic table example 36: Transition characteristic table set 36A: Transition characteristic table 36B: Transition characteristic table 37: Transition characteristic table example 38: Copy cost table (GET)
39: Workload characteristics table (GET)
40: Cluster system 101: CPU
102: main storage unit 103: communication unit 104: output unit 105: input unit 106: auxiliary storage unit 110: input unit 120, 121: copy cost generation unit 130, 131, 132: workload characteristic generation unit 140, 141: load Distribution characteristic generation means 150, 151: Arrival node number calculation means 160: Overhead prediction means 170: Output means 210, 211: Copy cost information 220, 221, 222: Workload characteristic information 230, 231: Load distribution characteristic information 300: Performance Prediction program

Claims (28)

A performance prediction device for a cluster system that shares data on main storage devices of a plurality of nodes connected via a communication line between the nodes,
Calculating the number of block copies per unit time made between the nodes in units of memory blocks in the main storage device;
A cluster system performance prediction apparatus, wherein clustering overhead generated in the cluster system is calculated based on the block copy count and CPU load information per block copy.   The cluster system performance prediction apparatus according to claim 1, wherein scalability of the cluster system is calculated based on the calculated clustering overhead and the number of nodes.   The block copy count is calculated for each window size, which is the number of accesses from the time when a specific type of access is made to the memory block set until the next access of the same type, and the block copy count for each window size. The cluster system performance prediction apparatus according to claim 1, wherein the number of block copies per unit time performed between the nodes is calculated based on the number of partial block copies.   4. The performance prediction apparatus for a cluster system according to claim 3, wherein the specific type of access is a write access, and access other than the specific type of access is a read access.   5. The performance of the cluster system according to claim 3, wherein the partial block copy count is calculated from an arrival node count that is an expected value of the number of nodes arriving at the window size count. Prediction device.   Calculating the probability of occurrence of an event in which access to the plurality of nodes is made a natural number of times under the condition that the sum of the number of accesses to all nodes is equal to the window size, and the occurrence of the calculated event The cluster system performance prediction apparatus according to claim 5, wherein the number of arrival nodes is calculated based on a probability of performing the operation.   The cluster system performance prediction apparatus according to claim 6, wherein the probability of occurrence of the event is calculated using a multinomial distribution or a multidimensional normal distribution.   6. The cluster system according to claim 5, wherein the number of arrival nodes is calculated using a polynomial, an exponential function expression, or a series expression in which the number of arrival nodes monotonously increases as the window size increases. Performance prediction device.   The number of arriving nodes is calculated based on a probability of appearance of an access sequence to the node realized by the number of accesses of the window size and a number of nodes receiving access by the access sequence. Item 6. The cluster system performance prediction device according to Item 5.   The probability of the appearance of an access sequence to a node from the previous several times to the natural number of times is calculated from the appearance probability of the access sequence from the previous time to the number of times obtained by adding 1 to the natural number of times, and based on the calculated result, The cluster system performance prediction apparatus according to claim 9, wherein a probability of appearance of an access sequence to the node realized by accessing the window size times is calculated.   The window size, which is the number of accesses from the time when a particular type of access is made to a set of memory blocks until the next type of access is made, is the number of read accesses and the number of write accesses included in the access. The cluster system performance prediction apparatus according to claim 1, wherein the sum is calculated by dividing the sum by the number of write accesses.   When the window size is an integer equal to or larger than the window size and when the window size is an integer equal to or smaller than the window size, the access to the plurality of nodes is performed under the condition that the total number of accesses is equal to the window size. Calculating the probability of occurrence of an event that is made any number of times, and calculating the number of arrival nodes, which is an expected value of the number of nodes arriving at the number of accesses of the window size, based on the calculated result. The performance prediction device for a cluster system according to claim 11.   The specific type of access is a write access, the access other than the specific type of access is a read access and an acquisition access, and the acquisition access is an access made only by the acquisition access in the past. 4. The cluster system performance prediction apparatus according to claim 3, wherein a block is read by a node that has been read and a block that has been previously accessed for write or read is not read by a block. .   The block copy is a block copy by the read access and a block copy by the acquisition access, and the CPU load is a CPU load by the read access block copy and a CPU load by the acquisition access block copy. The cluster system performance prediction apparatus according to claim 13. A performance prediction method for a cluster system in which data on a main storage device of a plurality of nodes connected via a communication line is shared between the nodes,
Calculating the number of block copies per unit time made between the nodes in units of memory blocks in the main storage device;
Based on the block copy count and CPU load information per block copy, the clustering overhead generated in the cluster system is calculated,
A performance prediction method comprising: calculating scalability of the cluster system based on the calculated clustering overhead and the number of nodes.   The block copy count is calculated for each window size, which is the number of accesses from the time when a specific type of access is made to the memory block set until the next access of the same type, and the block copy count for each window size. The performance prediction method according to claim 15, wherein the number of block copies per unit time made between the nodes is calculated based on the number of partial block copies.   17. The performance prediction method according to claim 16, wherein the specific type of access is a write access, and an access excluding the specific type of access is a read access.   18. The performance prediction method according to claim 16, wherein the partial block copy count is calculated from an arrival node count that is an expected value of the number of nodes arriving at the window size count.   The probability of the occurrence of an event in which the access to the plurality of nodes is made a natural number of times is calculated under the condition that the sum of the number of accesses in all the nodes is equal to the window size, and the calculated occurrence of the event 19. The performance prediction method according to claim 18, wherein the number of arrival nodes is calculated based on a probability. A performance prediction program for predicting the performance of a cluster system that is executed on a computer processing device and that shares data on main storage devices of a plurality of nodes connected via a communication line between the nodes,
Calculating the number of block copies per unit time made between the nodes in units of memory blocks in the main storage device;
Based on the block copy count and CPU load information per block copy, the clustering overhead generated in the cluster system is calculated,
A performance prediction program comprising a function of calculating scalability of the cluster system based on the calculated clustering overhead and the number of nodes.   The block copy count is calculated for each window size, which is the number of accesses from the time when a specific type of access is made to the memory block set until the next access of the same type, and the block copy count for each window size. 21. The performance prediction program according to claim 20, further comprising a function of calculating the number of block copies per unit time made between the nodes based on the number of partial block copies.   The performance prediction program according to claim 21, wherein the specific type of access is a write access, and access other than the specific type of access is a read access.   23. The performance according to claim 21, further comprising a function of calculating the number of partial block copies from an arrival node number that is an expected value of the number of nodes arriving at the number of accesses of the window size. Prediction program.   The probability of the occurrence of an event in which the access to the plurality of nodes is made a natural number of times is calculated under the condition that the sum of the number of accesses in all the nodes is equal to the window size, and the calculated occurrence of the event The performance prediction program according to claim 23, having a function of calculating the number of arrival nodes based on a probability.   The performance prediction program according to claim 24, having a function of calculating the probability of occurrence of the event using a multinomial distribution or a multidimensional normal distribution.   It has a function of calculating the number of arrival nodes based on the probability of appearance of an access sequence to the node realized by the number of accesses of the window size and the number of nodes that are accessed by the access sequence. The performance prediction program according to claim 23.   The window size, which is the number of accesses from the time when a particular type of access is made to a set of memory blocks until the next type of access is made, is the number of read accesses and the number of write accesses included in the access. 21. The performance prediction program according to claim 20, further comprising: a function of calculating the sum of the above and the number of write accesses. The specific type of access is a write access, the access other than the specific type of access is a read access and an acquisition access, and the acquisition access is an access made only by the acquisition access in the past. 23. The performance prediction program according to claim 21, wherein said node has a function of reading a block at a node and not reading a block at a node to which said write access or said read access has been made in the past. .

Images