JP2005110209A - Access response time prediction method, distributed performance data load property constitution method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To predict a Web page access response time in a plurality of candidate network configurations when newly constructing/adding or altering a network, and to constitute a load property corresponding to the response time for an application wherein the response time is distributed even if a load is fixed. <P>SOLUTION: Versatile performance analysis formula is used to predict Web page access response time distribution in a plurality of estimated networks. Round trip time distribution between a server and a client, file size distribution of a Web page, the number of contents in the Web page, a protocol processing time of a client terminal device/server terminal, and a network discard rate are used as input data to analysis formula. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a network design support tool for predicting a web page access response time distribution for a plurality of candidate network configurations and server setting positions during network construction planning. The present invention also relates to a method of processing performance data of a network server system taking an access response time distribution as an example.

  On the World Wide Web (WWW) site, a relational expression between the number of user accesses and access response time is prepared, the future user access fluctuation is estimated based on actual measurement, and the estimated response value is used as an input to determine the access response time by processing at the site. calculate. Using this result, when performance degradation is predicted, preventive measures such as server load balancing and user access limit are taken to maintain a constant quality service on the WWW site (Japanese Patent Laid-Open No. 2002-268922 “WWW Site performance monitoring device ”).

  Alternatively, a test message is sent from the client side to the server, and the time required for a response to be returned is measured. Based on this time, the time required for downloading a WWW page or file is estimated (Japanese Patent Laid-Open No. 10-116220 “Parameters”). See “System and Method for Easily Displaying on a Local Display Device”). In addition, the application running on the LAN is broken down into threads, and the response time of each thread on the WAN is predicted in consideration of the WAN characteristics (link bandwidth, network average queue length, etc.) There is also a method of predicting the overall response time when the application is operated on the WAN from the relationship (see US6393480 Application response time prediction).

  Further, network server system performance data such as response time has been processed in order to determine the performance change with respect to the load according to the purpose. For example, in a Web access application, which is a typical application of a network server system, when the load of the number of accesses increases, the number of accesses and the response are used to determine the number of accesses whose performance of the application response time deteriorates rapidly. The number of accesses at the point (point A in FIG. 23) where the time relationship is constructed and the slope changes rapidly is defined as the limit access number. In the case of an application in which the number of accesses can be controlled, an operational guideline is obtained that the system can be operated at or below the limit number of accesses during system operation. Alternatively, in an application where response time is given as a performance requirement such as a business application, for example, it is determined whether or not the performance requirement is satisfied with this limit access number, and if the required response time is not satisfied, For example, it is decided to introduce a server device with higher capability.

  It is known that the performance data of a network server system shows a distribution when a web access application is taken as an example. (See Non-Patent Document 3) The response time is distributed due to various factors such as the distribution of the page size. For this reason, since the response time is distributed even under a certain network load, the response time distribution is different for each load.

JP 2002-268922 A

JP-A-10-116220 USP 6,393,480 “Application Performance and Network Planning”: NEXT GENERATION NETWORKS October 14-18, 2002 −NetForecast WIDE Project 1998 Research Report Part 15 WWW Cache Technology Peter Sevcik, “Web Performance Not A Simple Number,” Business Communications Review, January 2003, P.8-10.

  In the prior art, actual measurement data is used, and future prediction in a network in actual operation is possible. However, since actual measurement is impossible in new construction or extension of a network, it is difficult to apply the above technique. Although performance can be predicted by simulation, it takes time to model a plurality of network configurations and obtain a predicted value by varying many parameters. The present invention has been made in consideration of the above circumstances, and the first object of the present invention is to reduce the Web page access response time in a plurality of assumed network configurations with less man-hours before the construction of the network. It is to provide a means to predict.

  The prior art has also configured response time characteristics with respect to the load in accordance with purposes such as grasping the limit performance and determining proper equipment introduction. However, since the web access response time is distributed even when the load is fixed, the load characteristic with respect to the response time can only be composed of an average value or one sample value. The second object of the present invention is to configure the load characteristic with respect to the response time of the application in which the response time is distributed even when the load is fixed as described above.

  In order to solve the first problem, in the present invention, a general-purpose performance analysis formula is used to predict Web page access response time distribution in a plurality of assumed networks. As input data to the analysis formula, distribution of round trip time between server and client, file size distribution of web page, number of contents in web page, protocol processing time of client terminal / server terminal, and discard rate on the network Used (FIG. 1). In order to achieve the second problem, the performance data value at a certain cumulative probability in the cumulative probability distribution of performance data is set as sample performance data for a certain load, and an approximation is calculated from a plurality of sets of load and sample performance data. A characteristic is that the line has a load characteristic (FIG. 16). Further, an appropriate load is calculated from the configured load characteristics (FIG. 20).

  By using a general-purpose analysis formula, it is possible to quickly predict the Web page access response time for a given network configuration without the need to model the network for each assumed network. Become. Further, by performing performance prediction at the time of network construction planning, it is possible to select an optimal one from a plurality of candidate network configurations / server setting positions. In addition, since the response time characteristic line for the load is configured by fixing the cumulative probability of the response time distribution, it is possible to configure the load characteristic of the response time having a distribution other than the average value or one sample value.

By analyzing the distribution of Web page access response time in each network configuration for a plurality of candidate network configurations at the time of new construction or expansion / reconstruction of the network, it is possible to obtain performance suitable for the purpose of use of the network. Select the network configuration. Also, an approximate line of load characteristics is calculated for the performance data distributed in this way by the method of the present invention.
Embodiments of the present invention will be described below with reference to the drawings.

  FIG. 1 shows a functional block diagram of a server (1) that implements the program of the present invention. The program executor of the present invention stores the program in the hard disk (1-4) by an existing method. Also, various input parameters, which will be described later, are input from the keyboard and stored in the hard disk (1-4) by the existing method via the input control (1-5) and the disk control (1-3). When executing the present invention, the program executor reads the program and input parameters into the memory (1-2) via the bus by an existing method, and calculates the response time distribution under the control of the CPU (1-1). The calculation result of the response time distribution is displayed on the display via the output control (1-6) and stored in the hard disk (1-4).

  Assume the network shown in FIG. 2-1 is a server, 2-2-1 is client A (PC-A), 2-2-2 is client B (PC-B), 2-3-1 to 2-3-4 are clients (2- 2-1, 2-2-2) and a switch (SW) existing between the server (2-1). Here, the access response time between the client A (2-2-1) and the server (2-1) is predicted. Here, an RTT (Round Trip Time) distribution is obtained by a simplified simulation.

  FIG. 3 shows a route for obtaining the RTT distribution. 3-1 is the background load. File transfer was performed from the server (2-1) to the terminal of client B (2-2-2), and a load of 30% was applied to the line. 3-2 is the route measured RTT. RTT was measured between client A (2-2-1) and server (2-1). The RTT density distribution obtained by the simulation is shown in FIG. The horizontal axis shows RTT, and the vertical axis shows probability density. The file size distribution uses the file size distribution of the Web page that is actually provided. Here, we use data from the 1998 WIDE report (Figure 5). The horizontal axis in FIG. 5 represents the file size, and the vertical axis represents the cumulative ratio, which is the total probability density below the value on the horizontal axis.

As an example of implementation,
R (Task Response Time) = Rd (Discovery Time) + Rt (Transfer Time) (All units are sec)
(See Application Performance and Network Planning: NEXT GENERATION NETWORKS October 14-18, 2002-NetForecast).

However,
Rd = 2 (D + Cc + Ct) + (D + (Cc + Cs) / 2) (T-2) / M + DLn ((T-2) / M + 1) + KT (L / (1-L ))
And,
Rt = MAX (8P (1 + OHD) / B, DP / W) / (1-√L)
And
P = Brp + FdDpx
T = FdDtx + I (1 + FdDtxFc) + ISx
It is.

  Here, parameters used in the present embodiment and their meanings will be described. B is an abbreviation for Bandwidth and means the lowest bandwidth in the network. In the present embodiment, 100 Mbps (bits per sec) is assumed. Brp is an abbreviation for Browser payload, which means the file size of the page to be displayed. The unit is bytes. In this embodiment, the actual measurement data shown in FIG. 5 is used. Cc is an abbreviation for Client processing time and means the processing time in the client terminal. In this embodiment, 0.00025 (sec) is used. Cs is an abbreviation for Server processing time, and means processing time other than TCP session at the server terminal.

In this embodiment, 0.01 (sec) is used. Ct is an abbreviation for Server TCP proc. Time, which means the TCP session processing time at the server terminal. In this example, it was assumed to be 0.001 (sec). D refers to a Round trip D elay, it means a round-trip delay of the network only. In this example, the simulation values shown in FIG. 4 were used. Dpx is DNS Turn Tax, which is the number of DNS servers used on the communication path. The unit is (count). In this embodiment, 0 (count) is assumed. Fc stands for Discovery Factor, which is the probability of using a DNS server.

  In this embodiment, 0 is assumed. I is the number of objects in the page. In this embodiment, it is assumed that the number of objects is two. K is TCP Timeout and means the time until the TCP session is disconnected due to timeout. In this company, 2 (sec) is assumed. L is Packet loss, which means the rate of lost packets during packet transfer. In this embodiment, it is assumed that packet loss is 0, that is, L = 0. M is an abbreviation for Multiplexing factor, which means the number of threads between the client terminal and the server.

  In this embodiment, it is assumed that the number of threads is four. OHD is an abbreviation for Overhead and means overhead by protocol. In this embodiment, 0.1 is assumed. P is an abbreviation for Payload, and its unit is (bytes). In this embodiment, Brp + FdDpx is used. Sx is an abbreviation for Security Tax, and 0 or 1 specifies whether to use SSL (Secure Sockets Layer). In this embodiment, SSL is not used and the value is assumed to be 0. T is Application Turns (count). In this embodiment, FdDtx + I (1 + FdDtxFc) + ISx is assumed. W means Effective window size and means the execution size of the TCP window. The unit is (bytes). In this embodiment, 8000 bytes is assumed.

  The RTT shown in FIG. 4 is divided as shown in FIG. 6, and the probability density of the section is obtained. Next, the Web page access response time (hereinafter referred to as response time) when the file size is changed is obtained using the maximum value in the section as an input to the analytical expression. Here, since the response time monotonically increases with respect to the file size, the cumulative distribution of response times when RTT is fixed can be obtained from the cumulative distribution of file sizes shown in FIG. FIG. 7 shows a response time distribution when RTT = 0.0003 (sec). The horizontal axis shows the response time, and the vertical axis shows the cumulative ratio, which is the total probability density below the value on the horizontal axis.

  Next, FIG. 8 shows a result of obtaining a similar response time distribution for each RTT value and multiplying by the RTT probability density shown in FIG. Next, the sum of these distributions is obtained and shown in FIG. The horizontal and vertical axes in FIGS. 8 and 9 are the same as those in FIG. From the cumulative distribution of response times, a complementary distribution of response times (FIG. 10) and a density distribution (FIG. 11) are obtained. In FIG. 10, the horizontal axis indicates the response time, and the vertical axis indicates the ratio exceeding the value on the horizontal axis. In FIG. 11, the horizontal axis represents response time, and the vertical axis represents probability density.

  FIG. 12 shows a series of procedures. First, the probability density distribution of RTT and the cumulative distribution of file size are acquired (13-2, 13-3). Next, the RTT is fixed, the file size is changed, and the response time cumulative distribution is obtained (13-4). Then, the cumulative distribution of response times for each RTT is multiplied by the probability density of the RTT, and the sum of each distribution is obtained and displayed (13-5). If necessary, obtain and display the complementary response time distribution and density distribution (13-6).

  In this embodiment, the server (1) itself is treated as a web access response time prediction tool by installing a program (software) on the server (1). However, other information such as a PC or workstation is used. It is also possible to realize a prediction tool by installing software on a processing device.

  Using the probability density of the file size and the cumulative distribution of RTT, obtain the cumulative distribution of response time by changing the RTT with the file size fixed, multiply the response time distribution for each file size by the probability density of the file size, The probability distribution of the response time is obtained by obtaining the sum of them.

  Instead of the simple simulation used in Example 1, in order to obtain the distribution of RTT, pay attention only to the relationship between PC-A (2-2-1) and server (2-1). A simplified simulation model as shown in FIG. 13 is created by replacing with a queue, and a pseudo load is applied to obtain an RTT distribution. Here, 14-1-1 represents the upstream queue of SW (2-3-1), 14-1-2 represents the downstream queue of SW (2-3-1), and 14-2- Reference numerals 1 and 2 respectively indicate upstream and downstream queues of SW (2-3-2). 14-3-1 to 14-3-4 indicate pseudo loads on each queue (14-1-1, 2 / 14-2-1, 2).

  In order to obtain the distribution of RTT, the delay time distribution in one stage of the queue is obtained by simulation and is convolved by the number of round-trip queue stages in FIG.

  In order to obtain the RTT distribution, the delay time distribution in one stage of the queue is obtained by analysis, and is convolved by the number of round-trip queue stages in FIG.

  In order to obtain the server processing time, a table (FIG. 14) showing the relationship between the capacity / load of the network and the server processing time is prepared, the server processing time is obtained from the given capacity / load, and the performance analysis formula is obtained. Input.

  One or more threshold values are provided in advance for the response time, and the ratio of the response time satisfying the threshold value is obtained from the response time distribution as the output result and displayed graphically. FIG. 15 is an example of display. Here, for the two threshold values T1 and T2, ratios X and Y at which the response times are equal to or less than T1 and T2, respectively, are displayed.

FIG. 16 shows the steps for configuring the load characteristics according to the present invention. The variable i is set to an initial value 1 so that the sample value of the load characteristic is n, and the cumulative probability is fixed to P (step S11). The response time distribution f (t, L _i ) at the load L = L _i is calculated by the method shown in the first embodiment (step S12). t = t _{i such} that f (t, L _i ) = P is calculated (step S13). A set (L _i , t _i ) of L _i and t _i obtained in step 13 is stored (step S14). The variable i is incremented (step S15) and repeated until i = n (step S16). An approximate line of the stored n sets of values (L _i , T _i ) (i = 1, 2,..., N) is calculated (step S17). The approximate line is calculated by a method such as a straight line approximation between two adjacent points or a polynomial approximation of n points.

In the eighth embodiment, the method of obtaining the response time distribution f (t, L _i ) at the load L = L _i by calculation is shown, but it may be actually measured instead of being calculated. FIG. 17 shows an example of the load characteristic configuration by actually measuring the response time distribution. The variable i is set to an initial value 1 so that the sample value of the load characteristic is n, and the cumulative probability is fixed to P (step S11). The response time distribution f (t, L _i ) at the load L = L _i is actually measured (step S121). t = t _{i such} that f (t, L _i ) = P is calculated (step S13).

A set (L _i , t _i ) of L _i and t _i obtained in step 13 is stored (step S14). The variable i is incremented (step S15) and repeated until i = n (step S16). An approximate line of the stored n sets of values (L _i , T _i ) (i = 1, 2,..., N) is calculated (step S17). The approximate line is calculated by a method such as a straight line approximation between two adjacent points or a polynomial approximation of n points.

18 and 19 show the results obtained in the steps of FIG. 16 or FIG. In FIG. 18, when the response time t is taken on the X axis and the cumulative probability is taken on the Y axis, the response time distribution f (t, L ₁ ) is illustrated as a result of step S12 or S121 when i = 1. As a result of S13, a response time t = t ₁ corresponding to the cumulative probability P is calculated. As the variable i is incremented, the response time distributions f (t, L ₂ ), f (t, L ₃ ), ..., f (t, L _n ) are the same as when i = 1 as a result of step S12 or S121. As a result of step S13, response times t = t ₂ , t = t ₃ ,..., T = t _n are illustrated as in the case of i = 1. FIG. 19 is an illustration of the approximate line calculated in step S17. The load on the X-axis in FIG. 19 takes the response time in the Y-axis, the stored n sets of values in the step _{S14 (L i, T i)} (i = 1,2, ..., n) of approximated line f _P ( L) is shown.

FIG. 20 shows the steps for calculating the proper load according to the present invention. Normally, even when operating with an appropriate load, the load fluctuates within a certain range. Therefore, the maximum load that satisfies the required response time even if there is a load change is defined as the appropriate load. In order to obtain an appropriate load, the load fluctuation ΔL is first stored together with the required response time T (step S21). L _{T where} f _P (L) = T is calculated (step S22). According to the definition, L _O = L _T −ΔL / 2 is set as an appropriate load (step S23).
FIG. 21 is an illustration of the results obtained in the step of FIG. In Figure 21, indicating that the value obtained by subtracting the half-value [Delta] L / 2 from the load L _T load change [Delta] L corresponding to the request response time T of the approximation line f _P (L) shown in FIG. 19 proper load L _O .

Examples 8, 9, 10, and 11 are examples that focus on the method of configuring the load characteristics. Next, FIG. 22 shows an embodiment focusing on the relationship between the system constituting this load characteristic and its user. FIG. 22 shows the input / output relationship of the user to the load characteristic configuration system. The user inputs a plurality of loads L ₁ , L ₂ ,..., L _n and a cumulative probability P (10) to the system. The load characteristic configuration system 20 calculates the load characteristic at the steps of the embodiment shown in FIG. 16, and outputs the response time characteristic (30) for the load to the user.

3 shows functional blocks of a server on which a Web page access response time prediction program according to the present invention is implemented. The assumed network configuration is shown. The simulation model used to obtain the RTT distribution is shown. The RTT distribution obtained from the simulation by the model of FIG. 3 is shown. The file size distribution of the 1998 WIDE report is shown. The density distribution of RTT is shown. The cumulative distribution of response time when RTT is fixed is shown. The cumulative distribution of response time for each RTT is shown. The cumulative density distribution of Web page access response time in the assumed network is shown. The complementary distribution of Web page access response time in the assumed network is shown. The density distribution of Web page access response time in the assumed network is shown. The procedure of web page access response time distribution analysis according to the method of the present invention will be described. A simplified simulation model of the assumed network is shown. It is an example of the database for calculating | requiring server processing time from the capacity | capacitance number and load of a network. It is an example of a result display method when a threshold value is set for the Web page access response time. The processing step figure which shows the method of the load characteristic structure of 1st Example of this invention. The processing step figure which shows the method of the load characteristic structure of the 2nd Example of this invention. The 1st figure which graphed the result of FIG. 17 or FIG. The 2nd figure which graphed the result of FIG. 17 or FIG. The processing step figure which calculates | requires the appropriate load of one Example of this invention. The figure which graphed the result of FIG. The figure which shows the input / output of the load characteristic structure system of one Example of this invention. The figure which graphed the conventional load characteristic.

Explanation of symbols

1 server
2-1 server
2-2-1,2 clients
2-3-1 to 4 switches
Step to calculate S17 load characteristics
S23 calculating the appropriate load.

Claims (15)

A method for predicting a web page access response time distribution for a client device of the server in a network including at least one server device and a plurality of client devices,
Input data including Web page file size distribution, server / client round trip time distribution, server / client device protocol processing time, and packet discard rate on the network as data for the response time distribution prediction Web page access response time prediction method characterized by using In the web page access response time prediction method according to claim 1,
A web page access time response prediction method characterized by further using a packet discard rate on a network and a round trip time distribution between a server and a client according to a network load as the data for prediction. The web page access response time prediction method according to claim 1,
Web page access response time characterized in that, as the prediction data, protocol processing time in the server device estimated from the number of clients and the load on the server in the assumed network configuration is used. Prediction method. The web page access response time prediction method according to claim 1,
Create density distribution of round trip time between server and client,
Change the file size of the web page for each round trip time,
Web page access response time is calculated using the above performance analysis formula,
Find the cumulative distribution of web page access response time from the cumulative distribution of web page file size,
A web page access response time prediction method for predicting a web page access response time distribution by calculating a sum of distributions obtained by multiplying a cumulative distribution of web page access response times for each round trip time by the density of the round trip time. The web page access response time prediction method according to claim 1,
Create a density distribution of the web page file size,
Change the round trip time between server and client for each file size,
Web page access response time is calculated using the above performance analysis formula,
Find the cumulative distribution of web page access response time from the cumulative distribution of round trip time,
2. The web page access response time distribution according to claim 1, wherein the web page access response time distribution is predicted by calculating a sum of distributions of the web page access response time cumulative distribution for each file size of the web page and the density of the file size. Prediction method. In a web page access response time prediction tool comprising a CPU capable of executing software, a storage means for storing the software, and an input means for inputting data necessary for execution of the program,
The software is
Inputting input data including, and
Obtaining a round trip time distribution between a server and a client in the network and a packet discard rate on the network from the network configuration to be analyzed and load data input by the input unit;
Obtaining a protocol processing time in the server device from the load;
Web page from the file size distribution of the Web page input by the input means, the number of contents in the Web page, the various protocol processing time data in the client device, and the round trip time distribution and the packet discard rate data A Web page access response time prediction tool characterized by executing an access response time distribution. In the web page access response time prediction tool according to claim 6,
With the software
Setting one or more thresholds for the web page access response time;
A Web page access response time prediction tool, comprising: calculating a probability that the Web page access response time is equal to or less than a given threshold value based on the Web page access response time distribution. An application that distributes performance data of network server system applications,
A step of calculating the distribution of performance data for a certain load, a step of calculating performance data corresponding to the fixed value of the cumulative distribution of performance data, and the above steps for a plurality of loads. A load characteristic construction method comprising: a step of repeating, a step of storing a set of performance data corresponding to a load and a fixed value in the previous period, and a step of calculating an approximate line of the previous period set. An application that distributes performance data of network server system applications,
A step of calculating the distribution of performance data for a certain load, a step of calculating performance data corresponding to the fixed value of the cumulative distribution of performance data, and the above steps for a plurality of loads. A load characteristic configuration program which operates a step of repeating, a step of storing a set of performance data corresponding to a load and a fixed value in the previous period, and a step of calculating an approximate line of the previous period. An application that distributes performance data of network server system applications,
The step of actually measuring the distribution of performance data for a load, the step of calculating the performance data corresponding to the fixed value by fixing the cumulative probability of the distribution of performance data, and the above steps for multiple loads A load characteristic construction method comprising: a step of repeating, a step of storing a set of performance data corresponding to a load and a fixed value in the previous period, and a step of calculating an approximate line of the previous period set. An application that distributes performance data of network server system applications,
The step of actually measuring the distribution of performance data for a load, the step of calculating the performance data corresponding to the fixed value by fixing the cumulative probability of the distribution of performance data, and the above steps for multiple loads A load characteristic configuration program which operates a step of repeating, a step of storing a set of performance data corresponding to a load and a fixed value in the previous period, and a step of calculating an approximate line of the previous period. A system that can obtain the load characteristics of the performance data of the network server system application,
It comprises the steps of calculating the load corresponding to the required performance data from the load characteristics, and the step of setting the load corresponding to the previous period required performance data to a value obtained by subtracting half of the assumed load fluctuation. Load characteristics configuration method. A system that can obtain the load characteristics of the performance data of the network server system application,
The steps of calculating the load corresponding to the required performance data from the load characteristics and the step of setting the appropriate load by subtracting half of the assumed load fluctuation from the load corresponding to the previous required performance data Load characteristic configuration program to be performed. An application that distributes performance data of network server system applications,
A load characteristic configuration system, wherein a user inputs a plurality of loads and one cumulative probability and outputs a response time characteristic for the load to the user. 15. The load characteristic configuration system according to claim 14, wherein the response time characteristic with respect to the load is displayed in a graph.

Images