JP2000112782A - Performance measurement device, storage medium recording performance measurement program and performance measurement method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To efficiently sample performance data of a program including a running program with interruption prohibited. SOLUTION: A register window 1 is the one for performing switch of a register group by calling and recovering a program. When a performance measurement start instruction is inputted, a performance measurement start processing means 2 changes environments of the register window 1 so that a trap is generated whenever calling of the program and recovery occur. A performance measurement information sampling means 3 executes a measurement object program 4 by the environments of the register window 1 which the performance measurement start processing means 2 sets, and samples performance measurement information 5 with generation of a trap as an opportunity. A performance measurement completion processing means 6 returns the environments of the register window 1 to the one before performance measurement start when execution of the program by the performance measurement information sampling means 3 is completed.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a software performance measuring device, a computer readable recording medium storing a performance measuring program, and a performance measuring method, and more particularly to a CPU (Central Processin) having a register window.
g unit), a computer-readable recording medium storing a performance measurement program, and a performance measurement method.

[0002]

2. Description of the Related Art In the development stage of software, it is necessary to form a program with an algorithm which can perform a high-speed processing even a little. Therefore, a technique for measuring the performance of the completed program is required. To measure program performance,
What is necessary is just to count how many steps the instruction is executed (how many machine language instructions are executed) to execute a specific process in the program.

In a CPU having a register window,
By using the register window, the number of steps can be counted. Below, SPARC architecture
The register window conforming to Version 9 is described.

FIG. 10 is a diagram showing an example of using a register window. In the architecture using the register window, the register window in the CPU is divided into 16 register sets. Each register set includes eight registers. The register window is used to manage a group of registers available from a program to be executed. Here, eight register windows W0 to W
7 are provided.

Each register window corresponds to three register sets. The register group corresponding to the register window is divided into three types of uses: “in”, “local”, and “out”. At this time, “out” and “in” of adjacent register windows share the same register set. The “in” register is a register in which a value to be received from the calling program is stored. The “local” register is a register that stores a value used only in the program. The “out” register is a register that stores a value to be passed to the calling program. The registers used by the program being executed are determined by the current window pointer (CWP).

[0006] In the register window, OTHER is displayed.
A WIN area is set, and this area cannot be used as a group of registers for use by a program being executed.

[0007] In most program calls, the register window is switched using a SAVE instruction or a RESTORE instruction, and the program transitions. In the case of the SAVE instruction, the register window is “W0 → W1 → W2 → W3 → W4 → W5 → W6 → W7”
It changes in the order of. On the other hand, in the case of the RESTORE instruction, conversely, “W7 → W6 → W5 → W4 → W3 → W2 → W1
→ W0 ”.

In the register window, CANSAVE
And CANRESTORE. CANSAVE is a function of the number of S
This is a value indicating whether the AVE instruction can be executed. Also, C
ANRESTORE is a value indicating how many times a RESTORE instruction can be executed without generating a trap.

When another program is called while a certain program is being executed, the calling program sets an argument in an out register of its own register window, and then performs a process of calling the program. Then, the called program issues a SAVE instruction. As a result, the register window is switched in the SAVE direction. That is, CWP is 1 in the SEVA direction.
Is changed, CANSAVE is decremented, and CANRESTORE is incremented. Then, the called program receives the argument from the in-register and executes the processing described in the program.

When returning from the called program to the calling program, a return value is set in the in-register and a RESTORE instruction is issued. As a result, the register window is switched in the RESTORE direction. That is, CWP becomes 1 in the RESTORE direction.
Is changed, CANSAVE is incremented, and CANRESTORE is further decremented. The return destination program receives the return information through the out register and continues the interrupted processing.

By using such a register window and tracing how many steps have been performed between the time when the environment of a certain register window is established and the time when the next register window is switched, the software Partial performance can be measured. Generally, when the register window is switched, it is controlled by an OS (Operating System) so that a trap does not occur. Therefore, conventionally, an interrupt is generated each time the program executes one step, and by this interrupt processing, the program counter can be traced, and the number of steps and other various information can be traced.

[0012]

However, if trace data is collected by generating an interrupt, it takes many times longer than usual processing to execute one step, and the amount of trace data becomes large. There is a problem that a large capacity of a storage device for storing data is required.

Further, performance data of a program running with all interrupts disabled cannot be collected. The present invention has been made in view of such a point,
It is an object of the present invention to provide a performance measuring device capable of efficiently collecting performance data of a program, including a program that runs with interrupts prohibited.

Another object of the present invention is to provide a computer-readable recording medium on which a performance measurement program for efficiently collecting performance data of a program, including a program running with interrupts disabled, is recorded. That is.

It is another object of the present invention to provide a performance measuring method for efficiently collecting performance data of a program, including a program running with interrupts disabled.

[0016]

According to the present invention, in order to solve the above-mentioned problems, a plurality of registers for use when an arbitrary program operates are provided, and a program being executed at the time of calling and returning the program is provided. In a performance measurement device that measures the performance of a program using a processor that has a register window that switches the group of registers that can be used in the program, a trap is generated when a program call and return processing occurs due to a performance measurement start command. Performance measurement start processing means for changing the environment of the register window so as to execute the program to be measured according to the environment of the register window set by the performance measurement start processing means, and collect performance measurement information upon occurrence of a trap. Means for collecting performance measurement information that performs Measuring device is provided.

According to such a performance measuring device, when the performance measurement start command is input, the performance measurement start processing means sets the register so that a trap is generated when a program call and a return process occur. The window environment changes. Thereafter, the performance measurement information collecting means executes the measurement target program in the environment of the register window set by the performance measurement start processing means, and collects the performance measurement information triggered by the occurrence of the trap.

In order to solve the above-mentioned problem, a plurality of register groups used when an arbitrary program operates are provided, and a register group usable in the program being executed at the time of calling and returning of the program is provided. When a performance measurement start command is input to a computer-readable recording medium on which a performance measurement program for measuring the performance of a program is recorded by a processor having a register window in which switching is performed, a program call and a return process are performed. When the error occurs, the performance measurement start processing means for changing the environment of the register window so that a trap is generated. The program to be measured is executed according to the environment of the register window set by the performance measurement start processing means. Performance information that collects performance measurement information Information collecting means, a computer-readable recording medium recording a performance measurement program and having a is provided.

By executing the performance measuring program recorded on such a recording medium by a computer, the functions of the performance measuring apparatus according to the present invention are constructed on the computer.

In order to solve the above-mentioned problem, a plurality of register groups used when an arbitrary program operates are provided, and a register group usable in the program being executed at the time of calling and returning of the program is provided. In a performance measurement method for measuring the performance of a program using a processor having a register window in which switching is performed, the environment of the register window is changed so that a trap is generated when a program call and a return process occur. And executing a program to be measured in accordance with the environment of the register window after the change, and collecting performance measurement information upon occurrence of a trap.

According to such a performance measuring method, the program to be measured is executed in an environment of the register window in which a trap is generated when a program call and a return process occur. A trap is generated each time return processing is performed. And
When a trap is generated, performance measurement information is collected.

[0022]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram illustrating the principle of the present invention. Register window 1 is a program call,
This is for switching the register group by return. When the performance measurement start command is input, the performance measurement start processing means 2 changes the environment of the register window 1 so that a trap is generated whenever a program is called or returned. The performance measurement information collecting means 3
The measurement target program 4 is executed according to the environment of the register window 1 set by the performance measurement start processing means 2, and the performance measurement information 5 is collected when a trap is generated. When the execution of the program by the performance measurement information collecting means 3 ends, the performance measurement end processing means 6 returns the environment of the register window to the state before the start of the performance measurement.

According to such a performance measuring device, when a performance measurement start command is input, the performance measurement start processing means 2 causes a trap to be generated whenever a program is called and returned. The environment of the register window 1 is changed. The setting value of the environment to be set in the register window is specifically determined as follows.

The register window is always "CANSA
The value of “VE + CANRESTORE + OTHERWIN” is controlled to be “the number of register windows−2”. Therefore, the value to be set in OTHERWIN is saved in memory, and when creating the environment of the register window, the environment of the register window is created by using the memory value and the number of saved register windows. I do. That is, the number of saved register windows is a value that should be included in CANSTORE. Therefore, when there are eight register windows, “6-memory value−the number of saved register windows” is CA
This is the value to be set in NSAVE. Therefore, when returning from the software trap handler to the trap generation source, the environment of the register window is created by the above equation. When the system is operating regularly,
The value of OTHERWIN is set to 0, and when the performance measurement is started, the value of OTHERWIN is set to 6.

FIG. 2 is a diagram showing a state of a register window at the time of performance measurement according to the present invention. Thus, CA
The values of NSAVE and CANSTORE are both set to 0. Also, in this example, the value of CWP is 2,
The register window W2 is a valid register group.

In such an environment, the performance measurement information collecting means 3 executes the program 4 to be measured.
Then, every time the program is called or returned, the register window needs to be switched. However, since there is only one available register group, a trap always occurs when switching to an adjacent register group is performed, the register group of the caller is saved on the stack, and the program of the callee registers the register. Window W3 or W
Make 1 available. Then, the performance measurement information collection unit 3 collects the performance measurement information 5 triggered by the occurrence of the trap. That is, the performance measurement information 5 is collected every time the program is called or restored.

When the execution of the measurement target program 4 by the performance measurement information collecting means 3 is completed, the performance measurement termination processing means 6 returns the environment of the register window 1 to the state before the start of the performance measurement. Thereafter, the register window 1 is used as usual. That is, even if the program is called or returned, the register window is switched without generating a trap within a certain number of times.

As described above, the performance measurement information 5 can be collected. By analyzing the performance measurement information 5, the performance in a specific section in the program 4 to be measured can be measured. For example, if the number of steps of an instruction is collected as performance measurement information, a section having a large processing load can be specified by comparing the number of steps in each section.

Next, embodiments of the present invention will be described more specifically. FIG. 3 is a system configuration diagram for implementing the present invention. FIG. 1 shows the configuration of a multi-CPU system. In this system, a plurality of processor modules 10, 20, 30 are connected to each other via a bus 41. Each processor module 10, 20, 3
0, the CPU 11, 21, 31, and the local memory 1
2, 22, and 32 are mounted. In this embodiment, the CPUs 11, 21, and 31 use the SPARC architecture V
It is assumed that the CPU conforms to ersion9. further,
To the bus 41, a shared memory 42, a hard disk device (HDD) 43, an input device 44, and a display device 45 are connected. The shared memory 42 stores data commonly used by the processor modules 10, 20, and 30. The HDD 43 stores data to be stored even when the power is turned off, such as the OS and a newly created program.

In such a system, when power is turned on, O
S is started. The OS controls distributed processing of application programs and the like. Then, the user operates the input device 44.
When the instruction for starting the performance measurement is input, the OS changes the environment of the register window, and the specified programs are processed by the CPUs 11, 21, and 31. During the processing, trace data in each of the CPUs 11, 21, and 31 is collected, and the respective local memories 12, 22, and 3 are collected.
2 is stored. When the execution of the program is completed, the trace data collected for each of the processor modules 10, 20, and 30 is collected in the shared memory 42 and stored in the HDD 43. Then, by analyzing the collected trace data, it is possible to know which process of the executed program and how many steps were required.

FIG. 4 shows the CPU of each processor module.
FIG. 2 is a block diagram showing a performance measurement function constructed by activating an OS. This processing function is a function provided as a part of the OS, and is realized by each of the CPUs 11, 21 and 31 executing a program for realizing this function.

The performance measurement start processing unit 51 has a function for preparing a performance measurement environment when a performance measurement start instruction is issued. That is, the environment of the register window as shown in FIG. 2 is prepared. Also, buffer areas for tracing performance information are secured in the local memories 12, 22, 32. Further, for the CPUs 11, 21 and 31,
Instructs to count the performance information, and sets a flag indicating that the performance is being measured.

The performance measurement information collecting section 52 traces performance information triggered by a trap generated each time the register window is switched. In particular,
The performance information values counted by the CPUs 10, 20, and 30 are processed into information for tracing, and the performance additional information setting unit 5
Trace along with the information of 4.

The performance measurement termination processing unit 53 restores the environment before the performance measurement is started when the performance measurement termination instruction is issued. That is, the environment as shown in FIG. 2 is returned to the normal register window environment, and the trace information stored in the buffers in the local memories 12, 22, 32 is integrated in the shared memory 42, and the contents Is stored in the HDD 43. Then, while releasing the buffer, the CPU 11, 21, 31 is instructed not to count the performance information, and the flag indicating that the performance is being measured is cleared.

The additional performance information setting section 54 collects certain additional information indicating the operation status of the system, information to be constantly collected, and the like. Examples of the additional information include information on a context switch, information on a virtual memory space, bind information on a load module, and the like. The information that is regularly collected includes time information.

FIG. 5 is a flowchart showing a processing procedure of the performance measurement start processing unit. The processes in this flowchart are all processes performed by the performance measurement start processing unit 51. [S1] A dedicated software trap for issuing a performance measurement start instruction for the entire system is issued. [S2] An instruction is issued to the software trap handler, and the operating environment of the issue source is saved. [S3] The operating environment of the issue source that has been saved is changed so that a trap is generated each time the register window is switched. That is, if the number of valid register windows from which traps are generated is other than 1, only the current window is left
The environment of another register window is forcibly saved in the stack. The number of valid register windows at the trap source is 1
If this is the case, a trap is generated each time the register window is switched as it is, so that the state is not changed. [S4] A buffer area for collecting trace data is prepared on the local memory. [S5] Space information of the virtual memory at the start of the performance measurement, load module bind information, and the like, which are additional information of the performance information, are stored in a buffer for collecting trace data. [S6] The performance measurement start flag is set in the memory. Further, the CPU sets a register of the CPU to count the performance information. [S7] The operating environment of the program that issued the software trap is restored, and the program returns to the trap source. After the return, a trap is generated every time the register window is switched.

Next, the contents of the processing performed by the performance measurement information collecting section 52 will be described. FIG. 6 is a flowchart illustrating the processing of the performance information collection unit. Performance measurement information collection unit 52
Starts operation when a trap that is generated each time the register window is switched or a trap that constantly occurs such as an external interrupt occurs. All the processes in this flowchart are executed by the performance measurement information collecting unit 52. [S11] It is determined whether the performance is being measured. This is determined based on whether or not a flag indicating that the performance is being measured is set. If the performance is being measured, the process proceeds to step S12,
Otherwise, the process proceeds to step S21. [S12] The performance information counted by the CPU is temporarily saved in the local memory. [S13] The environment of the register window is set for the information collection program so that a trap for performance measurement does not occur during performance information collection. [S14] Context information, virtual memory space information,
It is determined whether it is necessary to collect additional information such as the binding information of the load module. If necessary, the process proceeds to step S15, and if not, the process proceeds to step S16. [S15] The additional information collected by the performance additional information setting unit 54 is stored in the buffer. [S16] The saved CPU performance information is processed and stored in the buffer. [S17] The information that the system constantly collects is acquired from the additional performance information setting unit 54, and is included in the performance information and stored. [S18] It is determined whether a simulation is necessary.
This is determined by whether or not the trap is generated each time the register window is switched. That is, S
The AVE instruction or the RESTORE instruction is an instruction for switching the register window. However, at this point,
Since the operating environment for the information collection program is set, if it is executed as it is, the consistency of the environment of the register window cannot be maintained. Therefore, SAVE instruction or RES
It is necessary to simulate the TORE instruction. If the trap is generated each time the register window is switched, it is determined that the simulation is necessary, and the process proceeds to step S19. Otherwise, the simulation is determined to be unnecessary and the process proceeds to step S21. [S19] Instruction simulation is performed. [S20] Return to the trap source. [S21] An instruction is issued to the software trap handler to perform the normal handler process.

Next, the processing performed by the performance measurement termination processing section 53 will be described. FIG. 7 is a flowchart illustrating a processing procedure of the performance measurement instruction unit. The processes in this flowchart are all processes performed by the performance measurement termination processing unit 53. [S31] A software trap for issuing an instruction to end performance measurement of the entire system is issued. [S32] An instruction is issued to the handler of the software trap, and the operating environment of the source is saved. [S33] The saved contents are changed so that a trap does not occur every time the register window is switched. [S34] The contents of the buffer area are stored on the disk. [S35] The buffer area is released. [S36] The performance measurement start flag is cleared. [S37] The CPU operates the register to stop counting the performance information. [S38] Return to the trap source based on the operating environment of the program that issued the software trap.

After the return, no trap is generated every time the register window is switched, and the operation returns to the normal system operation. FIG. 8 is a diagram illustrating an example of collecting performance measurement information. In this example, a program whose performance is to be measured is composed of three modules 61 to 63.
Then, the module 62 is called from the module 61. Further, after a trap is generated during the execution of the module 62 and the processing is performed by the handler 63, the process returns to the module 61. In the figure, A1 to A4, B1 to B4, H
1, H2 indicates the value of the program counter.

At the time T1, a CALL instruction is issued in the module 61. In the CALL instruction, the transition of the program is performed after executing the next instruction, and the module 62 is called. Then, a trap is generated at B1, and the performance information when traveling to A1 to A3 is traced.

The point in time T2 is the point in time when a trap occurs in B3, and the control is transferred to the handler 63, and B1 to B2
The performance information when traveling is traced. The time point of T3 is when returning from the handler 63 to the module 62, and the performance information when traveling from H1 to H2 is traced.

The time T4 is the time when the module 62 returns to the module 61. A trap is generated at B4, and the performance information when traveling from B3 to B4 is traced. FIG. 9 is a diagram showing the contents of the traced performance information. The performance information 71 includes the time at which the information was collected, the caller PC (program counter) value, and the callee P
The C value, the number of running steps, and the number of other various events are registered. The number of events includes the number of instruction cycles (the number of cycles of the CPU operating clock), the number of cache references, the number of cache updates, the number of cache hits,
There are TLB (Translation Lookaside Buffer) mistakes. Although not shown in the figure, the number of events such as the number of cache misses, the number of copy backs, the number of uses of the bus, the number of references to the shared memory, and the number of updates of the shared memory can also be collected.

By analyzing these pieces of information collected each time the register window is switched, it is possible to know which section of the executed program has been executed and how many steps have been executed. Further, the number of instruction cycles required to execute a specific section, the number of cache references, the number of cache updates, the number of cache hits,
The number of LB misses is also known.

As a result, the performance can be measured without changing the system and the program to be measured, and the time spent for the performance measurement can be reduced. Moreover,
Since it is possible to know the performance of each program in detail, it is possible to specify the program that is affecting the performance.

Furthermore, since the collection of performance measurement information according to the present invention does not require interrupt processing, the performance of a program in which interrupts are prohibited can be measured. In addition,
The above-described processing functions are included in the OS, and are realized by executing, by a computer, a program describing the processing contents of the corresponding functions. In this case, the program in which the processing contents of the functions that the performance measuring device should have are described is recorded on a computer-readable recording medium.
Examples of the computer-readable recording medium include a magnetic recording device and a semiconductor memory. When distributing to the market, CD-ROM (Compact Disk Read Only Memory)
The program may be stored and distributed on a portable recording medium such as y) or a floppy disk, or stored in a storage device of a computer connected via a network, and transferred to another computer via the network. When executing the program on a computer, store the program on a hard disk device in the computer,
Load into main memory and execute.

[0046]

As described above, according to the present invention, in the performance measuring apparatus of the present invention, the environment of the register window is changed so that a trap is generated when a program call and a return process occur. When a trap occurs during execution of the measurement target program, the performance measurement information at that time is collected, so that the performance measurement information can be collected without changing the measurement target program. As a result, the performance can be measured efficiently.

In a computer-readable recording medium on which the performance measurement program of the present invention is recorded, the computer executes the recorded performance measurement program so that a trap is generated when a program call and a return process occur. By changing the environment of the register window so that it occurs, if a trap occurs during the execution of the program to be measured, the computer can perform processing such as collecting performance measurement information at that time, Efficient performance measurement using a computer becomes possible.

In the performance measurement method of the present invention, the program to be measured and executed is executed by executing the program to be measured in an environment of a register window in which a trap is generated when a program calling and returning process occurs. Since a trap is generated each time the return processing is performed, and the performance measurement information is collected when the trap occurs, the performance can be measured efficiently.

[Brief description of the drawings]

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

FIG. 2 is a diagram showing a state of a register window during performance measurement according to the present invention.

FIG. 3 is a system configuration diagram for implementing the present invention.

FIG. 4 is a block diagram illustrating a performance measurement function constructed by activating an OS by a CPU of each processor module.

FIG. 5 is a flowchart illustrating a processing procedure of a performance measurement start processing unit.

FIG. 6 is a flowchart illustrating a process of a performance information collecting unit.

FIG. 7 is a flowchart illustrating a processing procedure of a performance measurement instruction unit.

FIG. 8 is a diagram showing an example of collecting performance measurement information.

FIG. 9 is a diagram showing the contents of traced performance information.

FIG. 10 illustrates an example of using a register window.

[Explanation of symbols]

 1 register window 2 performance measurement start processing means 3 performance measurement information collection means 4 measurement target program 5 performance measurement information 6 performance measurement end processing means

Claims (8)

[Claims] A register window is provided in which a plurality of registers used when an arbitrary program operates is provided, and a register group used in a program being executed is switched at the time of calling and returning of the program. A performance measurement device for measuring the performance of a program by a processor, wherein a performance measurement start command changes the environment of the register window so that a trap is generated when a program call and a return process occur. Means, and a performance measurement information collecting means for executing a program to be measured according to the environment of the register window set by the performance measurement start processing means, and collecting performance measurement information in response to the occurrence of a trap. Performance measurement device. 2. A performance measurement end processing means for returning the environment of the register window to a state before the start of performance measurement by a performance measurement end command when execution of the measurement target program by the performance measurement information collecting means is completed. The performance measuring device according to claim 1, wherein: 3. The performance measurement start processing means sets the number of register windows that can be used by the program of the entire system without trapping to one, so that a trap is generated whenever a program is called or returned. 2. The performance measuring device according to claim 1, wherein the performance is measured. 4. The performance measurement information collecting unit collects, as the performance measurement information, the number of steps of an instruction processed when a specific section on the measurement target program is executed. Item 2. The performance measuring device according to Item 1. 5. The performance according to claim 1, wherein the performance measurement information collecting means collects, as the performance measurement information, the number of events during execution of a specific section on the measurement target program. measuring device. 6. The performance measurement apparatus according to claim 1, wherein said performance measurement start processing means simultaneously collects a plurality of types of information as said performance measurement information. 7. A register window provided with a plurality of registers for use when an arbitrary program operates, and a register window for switching registers available for the program being executed when the program is called and returned. When a performance measurement start command is input to a computer-readable recording medium that records a performance measurement program for measuring the performance of a program by a processor, a trap is generated when a program call and return processing occur. A performance measurement start processing means for changing an environment of the register window so as to generate, executing a program to be measured according to the environment of the register window set by the performance measurement start processing means, and generating performance measurement information upon occurrence of a trap. Means for collecting performance measurement information to be collected A computer-readable recording medium having recorded thereon a performance measurement program. 8. A register window in which a plurality of registers used when an arbitrary program operates is provided, and a register window in which registers usable in a program being executed are switched at the time of calling and returning of the program. In a performance measurement method for measuring the performance of a program using a processor provided with the processor, the environment of the register window is changed so that a trap is generated when a program call and a return process occur, and the register after the change is changed. A performance measurement method comprising: executing a program to be measured according to a window environment; and collecting performance measurement information when a trap is generated.