JPH01145736A - Dynamic performance measuring system - Google Patents

Abstract

PURPOSE:To arbitrarily set and change a collection item by measuring the performance of a program by setting the performance measuring environment. CONSTITUTION:The title system is constituted of a program 1 of a performance measuring object, a performance measuring environment 2 of the program 1, a performance measuring part 3 of a firmware, a log area 4, a performance measurement control part 5 of the firmware and a timer 6. The firmware for executing a dynamic performance evaluation is prepared, and whenever one instruction is executed from the program 1 for running in a VM state, this firmware is allowed to generate an interruption, performance information is collected by simulating an executing state of its instruction by the firmware, and thereafter, the control is returned to the program 1. Also, a period for collecting the performance information is determined by a sampling system, and a sampling interval for prescribing a sampling period therefor and the measurement frequency can be changed arbitrarily. In such a way, in accordance with a state, the performance measurement item and the sampling period can be designated elastically and the performance can efficiently be measured.

Description

DETAILED DESCRIPTION OF THE INVENTION [Summary] This invention relates to a method for measuring the dynamic performance of a program in a data processing device.

An object of the present invention is to diversify the performance evaluation information collected during program execution without using a dedicated hardware circuit or adding a large amount of hardware, and to make it possible to arbitrarily change the settings of collection items. year.

In a system for measuring the performance of a program in a data processing device.

and a timer that controls sampling timing.

and a performance measurement control unit that performs overall execution control for measuring program performance in response to timer interrupts.

The measurement control table manages measurement control information such as the sampling period for performing performance measurement and performance measurement items, and the execution state of program instructions is simulated according to the performance measurement items set in the measurement control table. It is equipped with a performance measurement section composed of firmware that has a function of collecting performance measurement data, and a log area that stores the performance measurement data collected by the performance measurement section.

At least measurement control information such as the sampling period and performance measurement items can be arbitrarily specified or canceled within a certain range.

[Industrial application field]

The present invention relates to a method for measuring the dynamic performance of a program in a data processing device.

Instruction execution frequency and number of pipeline conflicts for programs executed in VM state.

This system uses firmware to collect performance information such as the number of cache misses in real time, making it possible to evaluate the dynamic performance of a program or data processing device itself.

[Prior art and problems to be solved by the present invention] Conventionally, there are two methods for collecting performance information during the operation of a program (target program) for which performance evaluation is to be performed: one using hardware and the other using software. Things were the main thing.

In the hardware method, a circuit dedicated to performance measurement is provided within the data processing device, such as a selector that selects a signal at a point to be measured, and a counter that measures the duration and number of times the signal is measured.

If you increase the number of measurement items, the amount of hardware will increase.

Furthermore, there is a problem in that flexibility is lacking because the performance measurement items are fixed on the hardware.

In addition, the software method is performed using O8, for example, but in this case, the operating state of the program differs between when performance evaluation is performed and when performance evaluation is not performed, and the OS itself There was a problem in that it was extremely difficult to collect performance evaluation information.

In addition to the above-mentioned method, there was also a method of collecting performance evaluation information using firmware, but the measurement target was limited, such as measuring the appearance frequency of execution instructions.

The present invention diversifies the performance evaluation information collected during program execution without using a dedicated hardware circuit, and allows collection items to be changed arbitrarily. The objective is to be able to collect performance information about dynamic states, such as the number of line conflicts and cache misses, in real time with high reliability.

[Means for solving problems]

The present invention provides firmware that performs dynamic performance evaluation, generates an interrupt to this firmware every time an instruction is executed from a program running in a VM state, and simulates the execution state of that instruction using the firmware. After collecting information, control is returned to the program.The period for collecting performance information is determined by a sampling method, and the sampling interval and measurement frequency (measurement rate) are used to define the sampling period. and can be changed arbitrarily.

FIG. 1 is a diagram explaining the principle of the present invention.

In the figure, 1 is the program whose performance is to be measured.

2 is a performance measurement environment for the program 1, 3 is a firmware performance measurement section, 4 is a log area for storing performance measurement result data, 5 is a firmware performance measurement control section, 6 is a timer used for sampling control, and 12 is a performance This is a measurement control table that manages measurement conditions and performance measurement items.

When the performance measurement environment 2 is set, when one instruction of the program 1 is executed, the performance measurement unit 3 raises an interrupt and causes the performance measurement unit 3 to execute a predetermined performance measurement. The performance measurement unit 3 is composed of firmware that operates in an HPV (hypervisor) state.

It includes functions such as creating statistics on the number of times execution instructions appear for the program 1 and simulating the execution of instructions. In the instruction execution simulation, occurrences of pipeline conflicts and cache misses are detected according to information in the measurement control table 12, performance measurement data is collected, and stored in the log area 4.

The performance measurement control section 5 stores the measurement control table 12 in the measurement control table 12.

A function to set or cancel the sampling period, measurement frequency, and performance measurement items that define the sampling period, as well as fixed-cycle sampling pulses and measurement control table (1) supplied using interrupts from timer 6.
2) A function that generates a sampling period for performing performance measurement according to the measurement frequency specified in It has a control function to operate the performance measurement unit 3 according to control information such as performance measurement items specified in advance in the measurement control table 12.

[Effect]

In FIG. 1, the performance measurement control unit 5 operates in either the continuous mode or the sampling mode, and when the sampling mode is instructed, the performance measurement control unit 5 performs the specified sampling based on the timing pulse from the timer 6. Generate sampling periods based on interval and frequency to control performance measurement execution.

When the sampling period starts, the performance measurement control unit 5 first controls the performance measurement unit 3 when each instruction of the program 1 is executed and the execution of one instruction is completed in the setting state of the performance measurement environment 2. A hardware interrupt is raised.

The performance measurement unit 3 analyzes the interrupt, uses the performance measurement function of the firmware to perform a simulation of instruction execution, etc. according to specified performance measurement items, collects predetermined performance measurement data, and stores it in the log area. After that, control is returned to program 1 to execute the second instruction.

The performance measurement control unit 5 releases the performance measurement environment 2 for the program 1 when the sampling period ends. As a result, the program 1 executes instructions under the normal execution environment until the secondary performance measurement environment 2 is set.

What is the sampling interval, measurement frequency, and performance measurement items?
It can be specified arbitrarily within a predetermined range, making it possible to efficiently measure performance depending on the situation.

As described above, the performance measurement unit 3 is configured by firmware. Firmware is 1 normal.

Because it runs in HPV (hypervisor) mode with strong execution rights, it is less affected by the operating status of the data processing device than when measuring performance using O8, and runs in VM mode. This provides stable and reliable data on the dynamic performance of programs running.

〔Example〕

FIG. 2 shows the configuration of one embodiment of the present invention.

In the figure, l is the program whose performance is to be measured.

2 is a performance measurement environment, 3 is a performance measurement section configured with firmware, 4 is a log area, 5 is a performance measurement control section configured with firmware, and 6 is a timer.

7 is a performance measurement instruction unit, 8-1 to 8-4 are performance measurement setting commands, and 8-1 is a MODE-3ET
(mode cent) instruction, 8-2 is 5TART (7, start) instruction, 8-3 is STOP instruction, 8-
4 is a MODE-RESET (mode reset) command. 9 is a PSW (program status word), 10 is a program interrupt processing section, 11 is an external interrupt processing section, 12
13 is a measurement control table, 13 is a state control information setting area, and 14 is a measurement control table.
15 represents a performance measurement item setting area, and 15 represents an instruction measurement frequency setting area.

The performance measurement instructing unit 7 consists of a group of machine language instructions for instructing performance measurement, and in particular issues performance measurement setting instructions 8-1 to 8-4 to perform firmware performance measurement, which will be described later. operate the section. MODE-3
The ET command 8-1 is a command for setting details 1 of performance measurement, such as performance measurement items and sampling method.

The 5TART command 8-2 is a command for instructing the start of performance measurement. 5TOP command 8-3 is a command to instruct the end of performance measurement. The MODE-RESET command 8-4 is a command for canceling settings of performance measurement items and the like.

When the performance measurement setting instructions 8-1 to 8-4 are executed, an interrupt to the firmware occurs due to an instruction check.

When a firmware interrupt occurs due to execution of the performance measurement setting instructions 8-1 to 8-4, the performance measurement control unit 5 configured by firmware is activated.

The performance measurement control section 5 stores the measurement control table 12 in the measurement control table 12.

The control information specified by the MODE-3ET command 8-1 is set, and the 5TART command 8-2 is used as a trigger to set the Kuimaro, and processes such as controlling the start of measurement are performed. Detailed processing will be described later.

Timer 6 is similar to a general CPU timer and generates an external interrupt when a preset time has elapsed. Good but.

For example, an existing timer such as a so-called RVM timer used for dispatching virtual machines may be used.

When an external interrupt occurs, the external interrupt processing section 11 is activated. The external interrupt processing unit 11 is.

When it is detected from the interrupt code that the second interrupt is caused by the timer 6, the performance measurement unit 3 is activated.

The performance measurement unit 3 is composed of firmware with strong execution rights that operates in the HPV state, and as described later, collects performance measurement information based on the contents set in the measurement control table 12 and stores it in the log area 4. Perform the storage process.

External interrupts by the timer 6 are generated at intervals arbitrarily specified by the MODE-3ET instruction 8-1.

Therefore, the performance measurement unit 3 uses the external interrupt as a trigger to sample and collect performance measurement information.

FIG. 3 shows the configuration of the log area 4. The log area 4 includes a data part in a map data format divided by type of performance measurement item, and a data part in a trace data format that stores PSW, instructions, and hardware information in the order of execution instructions,
in the VM's memory.

Can be set so that it can be referenced by VM (not shown)
.

As shown in the data part of the map data format in Fig. 3, the performance measurement items in this example are SPV (
The number of instruction executions in supervisor) mode, the number of cycles that caused register conflict (■), the number of cycles that applied the ink clock in store/feft, (■) the number of cycles that caused memory conflict, etc.
(problem) mode, the number of cycles in which register conflicts occurred (■), and the worst instruction sequence that caused the most pipeline conflicts (■ to ■) in Figure 9, for example, in units of 256 instruction sequences. , buffer hint rate, etc. can be specified.

When these performance measurement items are specified, they are stored in their respective designated positions within the log area.

It can be output and referenced at any time, such as during program execution or at the end of the program.

Next, details of the measurement control table 12 will be explained.

FIG. 4 shows a configuration example of the state control information setting area 13 in the measurement control table 12.

In the illustrated state control information setting area 13, MODE-3ET
It is set to °1" by instruction 8-1, and MODE-RESE
Status flag 13-1 set to O'' by T instruction 8-4
It is set to “1” by the 5TART instruction 8-2, and the 5T
Status flag 1 returned to “0” by OP instruction 8-3
3-2 is provided.

In addition, sampling control bit group 13-3゜13-4
is provided. In this embodiment, bit group 13-3
For example, from 1.0 to 1,000 m by the 4 bits of
Four types of sampling periods up to 5 can be selected, and the frequency at which performance measurement data is collected in the specified sampling period can be selected in 16 steps using the four bits of bit group 13-4.

FIG. 5 shows an example of control when the sampling period is 1.0 ms and the measurement frequency is 174. in this case.

Repeated measurements are performed once every four sampling pulses generated by the timer at a period of 1.0 tu. That is, the period between two consecutive sampling pulses is defined as a measurement period (i.e., sampling period),
The period between the following three sampling pulses is a rest period.

FIG. 6 shows a configuration example of the performance measurement item setting area 14 in the measurement control table 12.

The illustrated performance measurement item setting area 14 consists of 16 bits,
Each bit can specify the number of instruction executions (frequency), register conflict, store/fetch interlock, memory conflict, buffer hint rate, worst instruction sequence, etc.

Each bit becomes valid when set to "1", and if a plurality of bits are "1", measurements of the corresponding items are performed in parallel.

FIG. 7 shows an example of the structure of the instruction measurement number setting area 15 in the measurement control table 12.

The illustrated command measurement number setting area 15 consists of 16 bits,
The bit position O indicates the validity of the information in this setting area, and the bit positions 1 to 15 represent the value of 0 to 2+s-1,
Specifies the upper limit for the number of measurements of consecutive instructions within the sampling period.

FIG. 8 fat, (bl) illustrates the effect of setting the number of instruction measurement times.

FIG. 8 [al] indicates that when the sampling period is 10, the measurement frequency is 1/2, and the sampling period is 10m5, when the number of consecutive instructions n is instructed to be measured within 5 sampling periods (measurement periods) Ions, 10 Indicates a state in which the number of instructions executed within Hiuchi has not reached n, and the sampling period is 10.
The whole ugly part is used for measurement.

Figure 8 (g)) shows the case where the number of consecutively executed instructions within 10 cycles of the sampling period is greater than n when measurements are performed under the same conditions as in Figure 8 (al). In this case, when the number of executed commands reaches n, storage of measurement data in the log area is discontinued, and the remaining data is ignored.

FIG. 9 shows an example of the configuration of the performance measuring section 3. In the figure,
When the hardware interrupt is raised due to the completion of execution of the last executed instruction 16, each function of the performance measurement unit shown in FIG. 9 is activated.

The instruction execution count measurement routine 17 identifies the last executed instruction 16 from the PSW 9 in FIG. 2, counts the number of executions for each instruction type, and calculates the instruction execution count 18 in the logout area.
Update and record the value of .

Pipeline simulation routine 19.

Simulated pipeline (pseudo pipeline) 2
0 and performs a simulation of instruction execution. 256 instructions for which simulation has been completed are stored in the instruction string buffer 21.

The main dynamic performance obtained during the pipeline simulation process is the Regisf Conflict Measurement Routine 2.
2. Store fetch interlock measurement routine 23. Memory conflict measurement routine 24. Buffer simulation routine 25. It is measured by the buffer hint rate measurement routine 27.

The buffer simulation routine 25 performs the buffer simulation using a simulated buffer (pseudo buffer) 26, and the buffer hit ratio at that time is measured by the buffer simulation routine 27.

Furthermore, the worst instruction sequence measurement routine 28 determines whether each of the 256 instructions accumulated in the instruction sequence buffer 21 is the worst instruction group.

The data of each performance measurement item measured by simulating instruction execution in this manner is recorded in the logout area 29.

〔Effect of the invention〕

The present invention does not require the addition of significant hardware.

Various conflicts and interlocks that occur during program execution in data processing equipment.

Alternatively, it is possible to collect detailed performance measurement data showing dynamic conditions such as buffer misses in real time.
In addition, performance measurement items and sampling periods can be flexibly specified depending on the situation, making efficient performance measurement possible.

[Brief explanation of the drawing]

FIG. 1 is a diagram explaining the principle of the present invention. FIG. 2 is a configuration diagram of one embodiment of the present invention. FIG. 3 is an explanatory diagram of a configuration example of a log area. FIG. 4 is an explanatory diagram of a configuration example of a state control information setting area. FIG. 5 is an explanatory diagram of an example of sampling control. FIG. 6 is an explanatory diagram of a configuration example of a performance measurement item setting area. FIG. 7 is an explanatory diagram of a configuration example of a command measurement number setting area. FIG. 8 is an explanatory diagram of the effect of setting the number of command measurements. FIG. 9 is an explanatory diagram of a configuration example of the performance measuring section. In Figure 1. l is the program whose performance is to be measured. 2 is the performance measurement environment for program l. 3 is the firmware performance measurement section. 4 is a log area 5 for storing performance measurement result data; 5 is a firmware performance measurement control section; 6 is a timer used for sampling control. 12 is a measurement control table for managing performance measurement conditions and performance measurement items; Output material θnI) +, familiarity [89min $IU! J 443 Ming's 1st company 1 groove A゛ Figure $ 2 Figure 4 Figure Y 7"I 7'1llabu ii /l ift, a
pt rzt name 5 Figure ← Clause 10 -- qny" ni rlfl'lRto",;tlaHA
The effect of the number of times the entire song has been sent is impressive. Figure 8.

Claims (1)

[Claims] A system for measuring the performance of a program (1) in a data processing device, comprising: a timer (6) for controlling sampling timing; and a program (1) in response to an interrupt by the timer (6).
a performance measurement control unit (5) that performs overall execution control for performance measurement; and a measurement control table (12) that manages measurement control information such as the sampling period for performance measurement and performance measurement items.
and a performance measurement system configured with firmware that has the ability to simulate the execution state of the program (1) instructions and collect performance measurement data according to the performance measurement items set in the measurement control table (12). (3) and a log area (4) for storing performance measurement data collected by the performance measurement unit (3), and at least measurement control information such as the sampling period and performance measurement items can be arbitrarily set within a certain range. The performance measurement control unit (5) sets the measurement control information on the measurement control table (12) in accordance with the specification or cancellation of the control information such as the sampling period and performance measurement items. Or, cancel the setting and while program (1) is running,
When there is an interrupt from the timer (6), the performance measurement environment (2) is set according to the sampling period control information set in the measurement control table (12), and the performance measurement section (3) executes the program (1). ) is a dynamic performance measurement method characterized by performing performance measurement.