JPH1063516A - Method and device for bottleneck detection - Google Patents

Abstract

PROBLEM TO BE SOLVED: To detect a bottleneck of a task with high precision by measuring the wait times of semaphores suspected to be bottlenecks by individual tasks and obtaining the cumulative time of the wait times. SOLUTION: An all-wait-state cumulative time calculating means 15 in a processor 1 adds the difference between the time that a timer mechanism 10 indicates when a task in a wait state enters an execution state and a state update time in a task control block 20, i.e., the time for which the task is in the wait time to the all-waitstate cumulative time in the task control block 20 to measure the cumulative time for which the task to be measured is in the wait time at an arbitrary semaphore 28. Thus, the wait times of semaphores suspected to be bottlenecks can be measured as to the respective tasks. Further, the semaphore 28 having the longest measured wait-state cumulative time is detected as a bottleneck.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to an information processing apparatus, and more particularly to a bottleneck detecting apparatus and a bottleneck detecting method for detecting a bottleneck in a parallel processing environment.

[0002]

2. Description of the Related Art In recent years, parallel processing has attracted attention as a means of achieving a high level of performance that is difficult to achieve with a single processor.

In such parallel processing, when a plurality of tasks use resources such as various management tables in a shared memory while mutually excluding each other, exclusive control is generally performed using a semaphore corresponding to the resources. It is. Here, a semaphore is a variable used to implement mutual exclusion. At this time, if the use by the task concentrates on a specific resource, the ability to use the specific resource will be limited, and the performance will decrease as a whole even if there is room for other resources. May be. Such a specific resource is called a bottleneck.

Heretofore, due to the existence of such a bottleneck, a situation has arisen in which increasing the degree of parallelism of processing does not directly lead to an improvement in performance as it is, and it has not been possible to achieve the expected improvement in performance.

In addition, it is not easy to detect the existence of the bottleneck, and it is extremely difficult to specify a bottleneck portion.

Therefore, conventionally, bottleneck detection has been performed by selecting a section where parallel operation is not possible in parallel processing and measuring the time spent in that section.

A typical example of the time spent in such a section where parallel operation cannot be performed is a spin-type waiting time in which program execution is repeated until a condition is satisfied, and task information is temporarily stored until the condition is satisfied. A block type waiting time stored in the task control block and waited is exemplified.

Here, in the block type synchronization system,
The program is controlled in units of tasks (processes), and each task transitions between an execution state and a waiting state as the program is executed. Therefore, the block-type synchronous waiting time measurement method measures the waiting time by summing the time spent in each state of the task.

[0009]

However, the conventional bottleneck detecting device has the following problems.

That is, the bottleneck detecting apparatus using the block-type synchronous waiting time measuring method of the related art has a problem that it is impossible to detect a bottleneck with high accuracy. This is because, when considering each task, only the total accumulated time that the task was in the waiting state was measured at each synchronization point, and one of the causes when the task transitioned to the waiting state was a bottleneck. This is because, even in the case where the task has occurred, the task is hidden by the accumulated time in the waiting state due to another cause, that is, the task is buried.

An embodiment of a bottleneck detecting apparatus employing a spin-type synchronous waiting time measuring method is disclosed in
-230226.

The bottleneck detecting device described in Japanese Patent Application Laid-Open No. 3-230226 detects a bottleneck by measuring a synchronization waiting time. There was a problem that behavior changed. This is because a process for measuring the number of repetitions of the process is inserted into the synchronization waiting process, and all the tasks that execute the process perform the measurement process. However, the time required for the inserted processing becomes long, resulting in an operation different from the original operation.

An object of the present invention is to provide a bottleneck detecting apparatus having an effective measurement method in a parallel processing environment using a block-type synchronization method in order to solve the above-mentioned problems.

[0014]

The present invention provides the following means in order to solve the above-mentioned problems.

That is, according to the present invention, there is provided a bottleneck detecting method for detecting a bottleneck in a parallel processing environment using a block-type synchronization wait processing method. The wait time is measured for each of the plurality of suspected semaphores, and the accumulation of the wait time in each of the semaphores is calculated as the measured wait state accumulated time, and the semaphore whose measured wait state accumulated time is the maximum is calculated. A bottleneck detection method of detecting a bottleneck is obtained.

Further, according to the present invention, for each task, the identification information of the semaphore to be measured is stored as the semaphore to be measured, and the time of transition to the waiting state for each task is stored as the waiting state start time. If the task exits the wait state as a result of the operation on the semaphore, it detects whether the semaphore identification information, which is the identification information of the semaphore, and the measured semaphore identification information match, and matches the result of the detection. If it is, from the current time and the waiting state start time, calculate the time in the waiting state,
The time in the current wait state, which is the result of the calculation, is added to the content of the total wait time that has been in the wait state by the semaphore to be measured, and the measured wait state is accumulated. In this case, a bottleneck detection method of detecting a semaphore having the maximum accumulated time in the measured waiting state as a bottleneck is obtained.

According to the present invention, a first storage area for storing a plurality of task control blocks corresponding to each of a plurality of tasks and a plurality of semaphores corresponding to each of a plurality of events are stored. And a second storage area for storing information on the specific task when the specific one of the plurality of tasks waits for the specific one of the events. A specific one of the blocks is stored in the first storage area, and information indicating the specific task control block is stored in the semaphore of the second storage area. And a processing unit for reading and executing information on the specific task from the specific task control block, thereby realizing parallel processing. In the bottleneck detection apparatus for detecting a bottleneck in the storage area, a storage area corresponding to the specific task control block in the first storage area is identification information on a semaphore to be measured among the plurality of semaphores. A first storage area for storing measured semaphore identification information, a second storage area for storing semaphore identification information that is identification information of a semaphore corresponding to an event that a task to be measured waits for, A third storage area for storing time information at which a certain task has transitioned to the wait state, and a fourth storage area for storing all wait state accumulated time information which is information on the accumulated time that the task to be measured has been in all wait states. And a storage area for storing the measured wait state accumulated time information that is information of the accumulated time in the wait state corresponding to the semaphore to be measured. And a storage area, said processing means may be storing the semaphore identification information measured in the first storage area, the fourth
A measurement condition setting unit for initializing a fifth region, a waiting state start processing unit for storing the semaphore identification information in the second region, and storing the time information in the third region, When the task to be measured ends the waiting state, the difference between the time information stored in the third area and the current time information is calculated, and the total waiting state accumulated time of the fourth area is calculated. Total wait state accumulated time calculating means for adding to the information, the semaphore identification information stored in the first area when the task to be measured ends the wait state, and the second area Measurement condition detection means for detecting whether or not the semaphore identification information stored in the storage device matches the semaphore identification information, and when the measurement condition detection means detects a match between the measured semaphore identification information and the semaphore identification information, Measuring wait state accumulated time calculating means for calculating a difference between the time information stored in the area and the current time information and adding the difference to the measured waiting state accumulated time information in the fifth area, For each task, a wait time is measured for each of a plurality of semaphores suspected of being a bottleneck, and the accumulation of the wait time in each of the semaphores is calculated as the measured wait state accumulated time information. A bottleneck detecting apparatus is characterized in that the semaphore having the largest measured wait state accumulated time information is determined as a bottleneck.

[0018]

Embodiments of the present invention will be described below with reference to the drawings.

The bottleneck detecting device according to the embodiment of the present invention includes a processing device 1, a shared memory 2, input means 3 such as a keyboard, and output means 4 such as a display.

Here, the processing device 1 includes a timekeeping mechanism 10,
Waiting state start processing means 11 and waiting state end processing means 1
2, measurement condition detection means 13, measured wait state accumulated time calculation means 14, and all wait state accumulated time calculation means 15
And a measurement condition setting means 16 for referring to and updating data stored in the shared memory 2 in accordance with a program stored in the shared memory 2.

The shared memory 2 includes a plurality of semaphores 28 (in FIG. 1, only one is shown for simplicity), which is a data structure used to implement mutual exclusion, and a task. It stores a task control block 20 which is a data structure for management, and first measurement control information 29 which is information on whether or not the system is performing measurement. More specifically, the task control block
A state 21 which is an area for storing state information of a task, a state update time 22 which is an area for storing a value indicated by the timer 10 when the state 21 of the task is updated, The semaphore identification information 23, which is an area for storing information for identifying the referenced semaphore, the total wait state accumulated time 24, which is an area for storing the accumulated value of all the times the task has been in the wait state, A second measurement control information 25 which is an area for storing information on whether or not the measurement target; a semaphore identification information 26 which is an area for storing information for specifying a semaphore to be measured; It has a measured wait state accumulated time 27 which is an area for storing the accumulated value of all the times in the wait state in the semaphore. Although only one task control block 20 is illustrated in FIG. 1, when there are a plurality of tasks, the task control block 20 includes a plurality of task control blocks 20 corresponding to each task.

The input means 3 is for inputting measurement conditions to the measurement condition setting means 16 in the processing device 1.

The output means 4 displays the measurement conditions set by inputting to the input means 3.

Hereinafter, each component will be described in more detail.

The time counting mechanism 10 outputs a value added at a constant cycle as time information.

The waiting state start processing means 11 updates the state 21 of the task stored in the task control block 20 from the execution state to the waiting state, and updates the time indicated by the timer 10 at the time of the update. Update time 22
Is stored in the task control block 20.
Further, the wait state start processing unit 11 outputs the identification information on the semaphore 28 that caused the wait to the semaphore identification information 2.
3 is stored in the task control block 20.

The wait state end processing means 12 updates the state 21 of the task stored in the task control block 20 from the wait state to the execution state, and updates the time indicated by the timer 10 at the time of the update. Update time 22
Is stored in the task control block 20.
Further, the waiting state end processing unit 12 instructs the measurement condition detection unit 13 to trigger the detection of the measurement condition.

The measurement condition detecting means 13 detects that the first measurement control information 29 is in the ON state. Further, the measurement condition detecting means 13 outputs the second measurement control information 25
Is detected as being in the ON state. Further, the measurement condition detecting means 13 detects that the semaphore identification information 23 is the same as the semaphore identification information 26 to be measured.

The measured waiting state accumulated time calculating means 14
The difference between the time indicated by the timer 10 and the state update time 22 is added to the measured waiting state accumulated time 27.

The all wait state accumulated time calculating means 15 adds the difference between the time indicated by the timer 10 and the state update time to the all wait state accumulated time 24.

The measurement condition setting means 16 sends the identification information of the semaphore to be measured input by the input means 3 to the task control block 20 corresponding to the task to be measured input by the input means 3. Identification information 2
6 is stored. The measurement condition setting means 16
Resets the total wait state accumulated time 24 and the measured wait state accumulated time 27, and sets the second measurement control information 25 to the ON state. Further, the measurement condition setting means 16 reflects the measurement start instruction input by the input means 3 in the first measurement control information 29.

Hereinafter, the operation of the bottleneck detecting apparatus having the above configuration will be described in detail with reference to FIGS.

First, the measurement condition setting processing in the measurement condition setting means 16 will be described with reference to the flowchart shown in FIG.

The measuring condition setting means 16 determines whether or not a task to be measured is specified by the input means 3 (step S301). If a task to be measured is specified, the task corresponding to the task is determined. The task control block 20 is specified (step S302). On the other hand, if the task to be measured has not been specified, the process ends.

Next, the measurement condition setting means 16 determines whether or not the semaphore to be measured has been designated by the input means 3 (step S303). Here, when the semaphore to be measured is not specified, the measurement condition setting unit 16 performs the process of step S306 described later.

If it is determined in step S303 that the semaphore to be measured is specified, the measurement condition setting means 16 determines the identification information of the specified semaphore to be measured in the task control block 20. It is set in the identification information 26 (step S304).

Next, the measurement condition setting means 16 initializes the total wait state accumulated time 24 and the measured wait state accumulated time 27 in the task control block 20 to zero, and sets the second wait state accumulated time 24 in the task control block 20 to zero. The measurement control information 25 is set to the ON state (step S305).

Next, the measurement condition setting means 16 determines whether or not a measurement start is designated by the input means 3 (step S306).

If the start of measurement is specified in step S306, the measurement condition setting means 16 sets the first measurement control information 29 to the ON state (step S30).
7) If measurement start is not specified, the measurement condition setting means 16 sets the first measurement control information to the off state (step S308).

When the first measurement control information is set in either step S307 or step S308, the measurement condition setting means 16 causes the output means 4 to display the measurement conditions (step S309).

Next, the wait state start processing executed by the wait state start processing means 11 after the measurement conditions (measurement environment) are set will be described with reference to the flowchart shown in FIG.

First, the wait state start processing means 11 identifies the semaphore 28 in the task control block 20 corresponding to the task which enters the wait state by identifying information of the semaphore 28 in the shared memory 2 which caused the entry into the wait state. The information is stored in the information 23 (step S401).

Next, the waiting state start processing means 11 sets the time indicated by the timer 10 at the time of this step (S402) (synchronously with the time of entering the waiting state) in the state in the task control block 20. The state is stored in the update time 22 (step S402), and the state 21 in the task control block 20 is updated to the waiting state (step S403).

Next, a description will be given, with reference to the flowchart shown in FIG. 5, of a wait state ending process when the task enters the wait state and ends the wait state.

First, the measurement condition detecting means 13 reads the first measurement control information 29 in the shared memory 2 and determines whether the first measurement control information 29 is on or off (step S501). . Here, when it is determined that the first measurement control information 29 is in the off state, the process of step S507 described later is performed.

If it is determined in step S501 that the first measurement control information 29 is in the ON state, the measurement condition detecting means 13 reads the second measurement control information 25 in the task control block 20, and It is determined whether 25 is on or off (step S5).
02). Here, when it is determined that the second measurement control information 25 is in the off state, the process of step S507 described later is performed.

If it is determined in step S502 that the second measurement control information 25 is in the ON state, the measurement condition detecting means 13 reads out the semaphore identification information 23 in the task control block 20 (step S503) and performs task control. Measured semaphore identification information 26 in block 20
Is read (step S504).

Next, the measurement condition detecting means 13 determines whether or not the semaphore identification information 23 and the semaphore identification information 27 to be measured are the same (step S505). here,
If it is determined that they are not the same, step S507 described later
Is performed.

If it is determined in step S505 that the semaphore identification information 23 and the semaphore identification information 26 to be measured are the same, the measured wait state accumulated time calculation means 14
Is the time indicated by the timer 10 at the time of this step (S506) (which is synonymous with the time when the task shifts from the waiting state to the execution state) and the task control block 20
By adding the difference from the value stored in the state update time 22 in the above, that is, the time spent in the waiting state to the accumulated time in the waiting state to be measured 27 in the task control block 20, the task to be measured is The accumulated time in the waiting state in the semaphore is measured (step S506).

Next, the all waiting state accumulated time calculating means 15
Is the time indicated by the timer 10 at the time of this step (S507) (accurately equivalent to the time when the task transitions from the wait state to the execution state) and the task control block 20
Is added to the total wait state accumulated time 24 in the task control block 20 by adding the difference between the value of the state update time 22 in the wait state and the total wait state accumulated time 24 in the task control block 20. The accumulated time is measured (step S507).

Next, the waiting state end processing means 12 sets the time indicated by the timer 10 at the time of this step (S508) (similarly to the time when the task shifts from the waiting state to the execution state). It is stored in the state update time 22 in the block 20 (step S508).

Finally, the wait state end processing means 12 updates the state 21 in the task control block 20 to the execution state (step S509), and ends the wait state end processing.

As described above, in the present embodiment, it is possible to measure the waiting time for a plurality of semaphores suspected to be bottlenecks for each task. In addition, as a result, the semaphore in which the measured waiting state accumulated time 27 is the maximum is detected as a bottleneck.

[0054]

As described above, according to the present invention, for each task, the waiting time is measured for a plurality of semaphores suspected to be bottlenecks, and a plurality of waiting accumulations are performed. Since the time can be obtained, if there is a specific bottleneck for a certain task, the bottleneck of the task can be determined with high accuracy regardless of the size of the waiting time when the task is in a waiting state due to other causes. Thus, a bottleneck detecting device which can be detected by the above is obtained.

Further, according to the present invention, only the waiting time of the semaphore to be measured in the task to be measured can be measured. Since there is no influence on the operation of the system due to the increase, a bottleneck detection device having less influence on the operation of the measurement target system by performing the measurement as compared with the conventional example can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a bottleneck detection device according to an embodiment of the present invention.

FIG. 2 is a diagram showing data stored in a task control block shown in FIG.

FIG. 3 is a flowchart showing a measurement condition setting process according to the embodiment of the present invention.

FIG. 4 is a flowchart illustrating a waiting state start process according to the embodiment of the present invention.

FIG. 5 is a flowchart illustrating a wait state end process according to the embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Processing apparatus 2 Shared memory 3 Input means 4 Output means 10 Clocking mechanism 11 Waiting state start processing means 12 Waiting state end processing means 13 Measurement condition detecting means 14 Cumulative time calculating means for measured waiting state 15 Meaning for calculating total waiting state accumulated time 16 Measurement condition setting means 20 task control block 21 state 22 state update time 23 semaphore identification information 24 total wait state accumulated time 25 second measurement control information 26 measured semaphore identification information 27 measured wait state accumulated time 28 semaphore 29 first measurement control information

Claims (5)

[Claims] 1. A bottleneck detection method for detecting a bottleneck in a parallel processing environment using a block-type synchronization wait processing method, wherein a plurality of semaphores suspected to be a bottleneck are identified for each task. The waiting time is measured for each, and the accumulation of the waiting time in each of the semaphores is calculated as the measured waiting state accumulated time, and the semaphore having the largest measured waiting state accumulated time is detected as a bottleneck. Bottleneck detection method. 2. For each task, the identification information of the semaphore to be measured is stored as the semaphore to be measured, the time of transition to the waiting state for each task is stored as the waiting state start time, and the result of the operation on the semaphore is stored. As the task exits the wait state, it detects whether the semaphore identification information, which is the semaphore identification information, and the measured semaphore identification information match, and if the result of the detection matches, From the current time and the waiting state start time,
The time in the waiting state is calculated, and the content of the total waiting time in the waiting state by the semaphore to be measured is stored in the content storing the sum of the waiting times. A bottleneck detection method for detecting a semaphore having the largest cumulative time under measurement in the waiting state as a bottleneck. 3. A first storage area for storing a plurality of task control blocks corresponding to each of a plurality of tasks, and a second storage area for storing a plurality of semaphores corresponding to each of a plurality of events. And storage means comprising
When a specific one of the plurality of tasks waits for a specific one of the events, information on the specific task is stored in the first storage area as a specific one of the task control blocks. Storing information indicating the specific task control block in the semaphore in the second storage area, and when the specific event that has been waiting occurs, the specific task control block sends the information related to the specific task. A bottleneck detection device for detecting a bottleneck in an information processing device that realizes parallel processing by including a processing unit that reads and executes information, wherein the specific task control block in the first storage area Is a storage area corresponding to the semaphore to be measured, which is identification information on the semaphore to be measured among the plurality of semaphores. Information, a second storage area for storing semaphore identification information that is identification information of a semaphore corresponding to an event that the task to be measured waits for, and a task for waiting for the task to be measured. A third storage area for storing time information that has transitioned to a state, and a fourth storage area for storing all wait state accumulated time information that is information on the accumulated time that the task to be measured has been in all wait states. And a fifth storage area for storing measured wait state accumulated time information that is information on the accumulated wait time corresponding to the semaphore to be measured, wherein the processing unit includes the measured semaphore identification. Measuring condition setting means for storing information in the first storage area and initializing the fourth and fifth areas; and storing the semaphore identification information in the second area. A wait state starts processing means for storing the time information in the third region, when to end the task waiting state which is the measurement object,
An all-waiting-state cumulative-time calculating unit that calculates a difference between the time information stored in the third area and current time information and adds the difference to the all-waiting-state cumulative time information in the fourth area; When the task to be measured ends the waiting state,
Measuring condition detecting means for detecting whether or not the semaphore identification information stored in the first area matches the semaphore identification information stored in the second area; When the means detects coincidence between the semaphore identification information to be measured and the semaphore identification information, the difference between the time information stored in the third area and the current time information is calculated, and the measurement of the fifth area is performed. A measured wait state cumulative time calculating means for adding to the wait state cumulative time information, and measuring a wait time for each of a plurality of semaphores suspected to be a bottleneck for each task. Calculating the accumulation of the waiting time in each of the semaphores as the measured waiting state accumulated time information, and regards the semaphore whose measured waiting state accumulated time information is the largest as a bottleneck. Determining Te bottleneck detection device according to claim. 4. The measurement condition setting means, which is unique to the information processing apparatus, sets first measurement control information which is information indicating whether or not measurement processing is performed for the entire information processing apparatus, and stores the first measurement control information in the storage means. 4. The bottleneck detecting device according to claim 3, wherein: 5. The method according to claim 1, wherein the measurement condition setting unit sets second measurement control information, which is information indicating the presence or absence of a measurement process of the task, for each of the task control blocks and stores the second measurement control information in the task control block. The bottleneck detection device according to claim 4, wherein