2 Performance

2.3 approach for the regulation of performance

2.3.5 measurings of performance of multi-layer architectures

After bench mark tests are limited mainly to the measurements of hardware and/or computers, techniques are still missing, in order to determine the performance of more complex system landscapes, like multi-layer architectures. In complex system landscapes plays not only the performance of each individual component a large role with the determination of the entire system performance, but also the interaction of the components among themselves. That is, not sufficiently the performance of each component of the system landscape is to be determined to project and these on the basis of assumptions on the overall system. It is necessary to test the entire system and determine possible bottlenecks. Since it is already difficult to define and measure the performance from individual components to, it is necessary to find for the performance analysis of multi-layer architectures new beginnings.

2.3.5.1 definition - measure of the performance

As already with the measurings of performance seen by hardware components, it is difficult to define a uniform measure for the performance. From there it is to be first specified unabdinglich before each measurement, which data describe the performance of the system best. In the last sections most diverse definitions were presented to define the tried the performance of the systems. With the MIPS and FLOP the number implemented of the instructions or operations was selected as measure for the performance, introduced with the Dhrystone bench mark its own measure that the number of program sequences per second represented. But in multi-layer architectures such a measure cannot be used, but new considerations must be employed.

In the multi-layer architecture of an application of Web selected here a suitable measure for the performance is the time some user to wait must, until its inquiry was worked on and it the result of its inquiry receives. As previously mentioned this period of reply sits down together out it to Browserantwortzeit, the network period of reply and the server periods of reply. This period of reply in connection with the number of users, who work at the same time on the system, defines a measure of the performance of an application of Web. However no statement about can be still met whether the appropriate system is performant or not. This is only made possible by the definition of expectations to the system.

2.3.5.2 expectation definition

Related to applications of Web, which can consist of many sides, this means that for each possible inquiry of the user a period of reply must be defined, which may not be exceeded, so that the system can be called performant.

But the definition of expectations is a large problem, since many parameters must be considered. It begins with the different complexity of the inquiries and results, which are to return the delivery these inquiries. The expected period of reply of a short log in dialogue should be substantially shorter than the expected period of reply of an inquiry, which accomplishes computations over the temperatures of the letzen centuries and plots these. In addition come different conditions and subjective impressions of the users, who must be considered. A user on the application of Web with a 56K bit the modem accesses is used to it that the treatment of its inquiries needs a certain time. For it are not 10 seconds possibly performant enough, whereby however a user accesses by optical waveguide cables the application of Web and asks the same inquiry, after 5 seconds a word of performance to already hear would like.

By it is to be recognized that the expectation definition of the performance measure must probably be considered and again be abstracted and defined for each system.

2.3.5.2.1 example of a possible expectation definition

After in the last section the necessity for the definition of an expectation definition was described, now on the basis the example of an application of Web a possible definition is indicated.

With the access to an application of Web different parameters must be considered, in addition belong among other things:

• the number of parallel accessing users
• the range, with which a user accesses the application of Web
• the point of network entrance of the user with the access to application

On the basis these parameters ideally expectancy values for the periods of reply of the user inquiries must be defined for each combination, at which application within the demanded performance borders works.

Since determining of the demanded periods of reply is not very promising through treasures and is not possible a computation for all cases, procedures must be found to specify these.

A possible approach is a definition of a reference value, which can be determined by measurement of a selected scenario. In addition a basis configuration of all parameters will become specified and it the periods of reply all sides of the application of Web determined.

For the parameters specified above the optimum assumptions should be selected. That is, as basis parameters for example the largest range and a very small user number should be consulted.

According to the determination of the reference value now an acceptance factor can be specified, to that dependent on the parameters, which defines expected and/or demanded periods of reply, which may not be exceeded, in order to call the system performant. From this now the maximally permitted periods of reply for each side of the application of Web can be computed. The acceptance factor can be built up from the definitions of the different parameters.

Assumed, 56K modem, ISDN modem with channeling and 2 a MBit DSL line seizes users over 3 different ranges on the application of Web. The reference value was selected and defined with the basis of the 2MBit-DSL-Leitung thereby the partial acceptance factor for the range parameter 2MBit-DSL with AF-Bandbreite2MBit=1. for the access over a ISDN modem fixed that the periods of reply may be so high 6-fach, for a modem even 9-fach. From this AF-BandbreiteIsdn=6 and AF-Bandbreite56k=9. follow.

The dependence of the acceptance factor on the parallel user number to seize at the same time on the system must be specified proportionally.

Assumed that for 100 users each the periods of reply may rise around 2% it is defined, as partial acceptance factor for the user number:

AF-pUser = 0.02 x parallele user number / 100.

The acceptance factor defined for this example only from two parameters can be formulated now as:

AFGesamt = AF-pNutzer x AF-range.

Here the partial factors, on the tests which can be accomplished depend to begin.

In this way the acceptance factor is expandable on further partial factors, which can to affect and into the expectation definition to be included be supposed a system.

The acceptance factor can be multiplied now by the determined reference values for each side (AZ-Ref-side) of the application of Web and supplies by it for each side dependent on the accomplished tests the expectation and/or limit values, which may not be crossed.

EWSeite = AFGesamt x AZ-Ref-side

By these limit values it is possible after the execution of the tests to decide concretely whether a system corresponds to the requirements or not. If the expectancy value is larger than the determined periods of reply of the appropriate side, then the side is to call performant is smaller the expectancy value, does not correspond the side not to the demanded performance.

The case that inquiries are because of a side partly over and partly below the expectancy value, will arise. From there it is to be availed meaningfully of statistic procedures, which can filter the standard case. This should be used then as reference value. Procedures for the computation of the strewing measure present themselves from the statistics, which can be consulted with the determination of a representative value from the quantity of the periods of reply of a side and compared with the expectancy value. While the arithmetic means is already sufficient in many cases, offer variance, standard deviation, Variationskoeffizient or Interquartialsabstand a larger security that few incorrect measured values do not falsify the standard case too much.

Translated by Yahoo Babelfish Service