JPH07191882A - Memory access frequency measuring system - Google Patents

Abstract

PURPOSE:To efficiently improve performance by grasping the load distribution of a target software. CONSTITUTION:A trace sampling means 1 is integrated into a target computer system, and the history of memory access at target software operating time is sampled in a trace file F. Continuously, continuous areas on an address space are set by a user, and the number of times of access for each of continuous areas and for each access type is summed up from the address of memory access recorded in the trace file F, edited and displayed. Thus, the heavy processing part of the target software is made clear from the number of times of instruction fetch, and the data of high access frequency are made clear from the number of times of operand access. Since these high load parts are mainly improved, the performance can be efficiently improved.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a memory access frequency measuring method for detecting routines with heavy software processing and data with high access frequency in a computer system.

[0002]

2. Description of the Related Art As a conventional method for measuring the processing distribution of software, there is a method for measuring the number of executions for each instruction code as disclosed in Japanese Patent Laid-Open No. 01-177140. This is a method of accurately counting the number of executions of each instruction code, and it is possible to recognize an instruction having a large number of executions. This method is suitable for determining an instruction that is highly improved by speeding up in hardware improvement. However, it is not very effective in improving software performance. This is because it is difficult to narrow down the tuning target based on the number of executions of each instruction, since general instructions are used everywhere in software.

On the other hand, UNIX has a function of measuring the number of executions for each routine called a profiler. This is shown in FIG.
As shown in, an interrupt is periodically generated during the program execution, the value of the program counter at the time of the interrupt is acquired in step 91 of the interrupt processing routine, and the routine being executed is searched in step 92, and the step At 93, the addition of the number of executions of the routine is repeated. In this method, the relative frequency of the number of execution instructions for each routine is known. Since it is measured by sampling, the exact number of instruction executions of each routine is unknown. However, if sampling is performed for a sufficiently long period, the relative frequency of each routine with respect to the total measured frequency can be known with some accuracy. This makes it possible to identify a heavy-duty routine in the program, and improve the routine to effectively improve the performance of the program. However, in the case of this method, if the sampling period cannot be taken long, or if the program is too large and individual execution frequency is low,
The accuracy becomes worse. Also, as shown in FIG.
No interrupt is generated in the interrupt disabled section, and an interrupt is generated after the interrupt is disabled. Therefore, the routines in the interrupt disabled section are not counted at all, and sampling is performed immediately after the interrupt is released. Thus, in this method, there are some areas that cannot be fundamentally measured.

[0004]

In the case of software, such as an operating system, which has a large program and includes a considerable interrupt prohibited section, there is concern about the accuracy of the relative instruction execution frequency for each routine by sampling. In addition, when it is expected to improve the performance by changing the data structure, the target may not be narrowed down only by examining the instruction execution frequency for each routine. That is, the instruction execution frequency for each routine only shows the aspect of the load distribution.

Therefore, the present invention provides a memory access frequency measuring method which enables accurate measurement of instruction execution frequency for each routine and also grasps load distribution from other aspects such as operand access frequency. With the goal.

[0006]

A memory access frequency measuring method according to the present invention includes a trace file recording a memory access type and an address, an area table including definition information of continuous areas and a counter for each access type. Trace collection means for recording the history of memory access during execution of target software on the target computer system in the trace file, area setting means for registering continuous areas in the address space specified by the user in the area table, and trace file It has a profiling means for counting the number of times of each type access to each area registered in the area table from the recorded memory access information, and a profile printing means for editing and printing the area table.

[0007]

1 is a diagram showing the components of the present invention. Figure 1
, 1 is a trace sampling means, 2 is a region setting means,
3 is a profiling means, 4 is a profile printing means, F is a trace file, and T is an area table.
With the entire configuration described in FIG. 1, the memory access frequency can be measured for each area of software operating on the target computer system and for each access type. Here, the operation of each component will be described.

The trace collection means operates in the target computer system and records the history of memory access in the trace file. One entry is recorded in the trace file with one memory access. This entry includes the type of memory access and the accessed address. That is, when the software is operated on the computer system equipped with the trace collection means,
The history of memory access that occurs when the software operates is recorded in the trace file.

After collecting the trace file, the area setting means registers a plurality of continuous areas in the address space in the area table. Then, the profiling means searches the area table for a continuous area to which the address recorded in each entry of the trace file belongs, and increments the counter of the corresponding area by one. At this time, there is a counter for each access type, and the counter to be added is determined from the type of access recorded in the trace entry. The above operation is performed for all entries in the trace file. When the profiling means is completed, the number of memory accesses included in the input trace file, that is, the number of memory accesses for each continuous area and for each access type, which is generated from the software operating when the trace was collected, is counted. Finally, the profiling means prints these counter values together with the area definition information.

Access types include instruction fetch, operand read, and operand write. By registering the code area of each subroutine as the area and checking the number of instruction fetches, the correct number of instructions executed in the subroutine can be known. Further, by registering each variable area as an area and checking the number of times of operand read / operand write, the accurate number of times of access to the variable can be found. Of course, the number of times of reading and the number of times of writing can be known separately.

Since a plurality of areas can be set in the area table, all the memory access times to be checked can be counted by one operation.

[0012]

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

As shown in FIG. 1, according to the embodiment of the present invention, a trace file F recording the type and address of memory access, an area table T including definition information of continuous areas and a counter for each access type, Trace collection means 1 for recording the history of memory access in the trace file F in the target computer system, area setting means 2 for registering a continuous area in the address space designated by the user in the area table T, and trace file F It comprises a profiling means 3 for counting the number of times each type of access to each area registered in the area table T from the recorded memory access information and a profile printing means 4 for editing and printing the area table T.

In the case of the present invention, it is necessary to embed the trace collecting means 1 in the target computer system. The implementation method of the trace collection unit 1 varies depending on the type and form of the target computer system and the embedding method. Here, an embodiment will be described in which the trace collecting means 1 is incorporated in an instruction level simulator instead of an actual machine.

FIG. 2 shows the internal structure of the trace file F. The trace file F is an array of trace entries 21. Each trace entry 21 has an access type field 22 indicating the type of memory access.
And a field 23 of the accessed address. Various types of access types can be defined according to the target computer system and the purpose of using the profile data. Here, there are three types of instruction fetch, operand read, and operand write, which are considered to be the most common. Instruction fetch is a memory access for reading an instruction word,
Operand read is a read access generated by instruction execution, and operand write is a write access also generated by instruction execution.

FIG. 3 is a view showing the internal structure of the area table T. The area table T is an array in which data necessary for the profiling operation is registered for each area page.
Each area item includes an area name field 31, a head address field 32 defining the area, a size field 33, and a counter for each access type. The counter includes a counter 34 that counts the number of instruction fetches, a counter 35 that counts the number of operand writes, and a counter 36 that counts the number of operand reads.

FIG. 4 is a flow chart showing the processing of the trace sampling means 1. The trace collection means 1 is incorporated in the instruction level simulator of the target computer system and is called when accessing the memory. At this time, the access type and access address are passed from the caller. In step 41, the passed access type and access address are set in the trace entry buffer,
In step 42 it is determined whether the buffer is all full. If there is still space left, the process returns to the reading source as it is. When it is determined in step 42 that the buffer is completely filled, the buffer is written to the trace file in step 43, and the buffer is cleared in step 44. It works functionally without buffering, but the simulation is very slow. Generally, the trace collection means 1 becomes an operation overhead of the target computer system. Therefore, it is necessary to devise to reduce the overhead.

FIG. 5 is a flow chart showing the processing of the area setting means 2. The area setting means 2 receives a command from the user in step 51, and determines the command in step 52. If it is not the area registration request, the registration processing is completed. If the result of the determination in step 52 is the area registration request, a free entry is secured from the area table T in step 53, and the designated name / start address / area size is set in the entry secured in step 53 in step 54. Then, in step 55, the counter group of the entry is cleared to zero. Here, the most basic area setting method has been described. In addition, there is a method of writing the definition of the area in a file in advance and reading the definition, or a method of automatically extracting the area from the symbol information on the program included in the executable file or the like. In the latter case, the area delimited by the address represented by the symbol may be registered, and the symbol name may be registered as the name. The area definition by the symbol name is convenient when collectively registering areas of program subroutine units or data structure units.

FIG. 6 is a flow chart showing the processing of the profiling means 3. At step 61, one trace entry 21 is read from the trace file, and at step 62, the end of the file is judged. If the file has ended, the process ends here. If there is still data in step 62, in step 63, the access address field 2 of the trace entry 21 extracted
The area table T is searched for an area item including the address recorded in No. 3. The area item including the address means an item including the address between the head address 32 and the area size 33. In step 64, the search result is judged. If there is no area item including the access address 23 recorded in the trace entry 21 in the area table T, the process returns to step 61. If there is an area item including the access address 23 recorded in the trace entry 21 in the area table T in step 64, the access type 22 of the trace entry 21 of the area item searched in step 63 is found in step 65. The counter corresponding to is incremented by 1. When the access type 22 indicates instruction fetch, the instruction fetch counter 34
When the operand write is indicated, the operand write counter 35 is added, and when the operand read is indicated, the operand read counter 36 is added. When step 65 ends, the process returns to step 61 and the next trace entry is processed in the same manner.

FIG. 7 is a flow chart showing the processing of the profile printing means 4. The profile printing means 4 reads one area item from the area table T in step 71, and in step 72, the name of the area item, the head address 32, the area size 33, and the instruction fetch counter 3
4. Operand write counter 35 and operand read counter 36 are edited and printed. Step 7
In step 3, it is determined whether or not all the area items in the area table T have been processed. If there are still remaining area items, the process returns to step 71 to process the next area item. Step 73,
If it is determined that all area items have been completed, the processing ends.

Finally, an embodiment in which the trace sampling means 1 of FIG. 4 is embedded will be described. FIG. 8 shows the instruction level simulator of the computer system in which the trace collecting means 1 is embedded. First, the operation of the simulator itself before embedding the trace collecting means 1 will be described. Step 81
Then, the instruction word is read, and in step 82, the read instruction word is decoded. In step 83, it is determined whether operand reading is necessary. If operand reading is required, the effective address is calculated in step 84 and the operand is read in step 85. If it is determined in step 83 that operand reading is not necessary, and after the reading in step 85 is completed, in step 86,
The processing corresponding to the instruction word is performed, and in step 87, it is determined whether or not the operand writing is necessary. When it is determined that the operand write is necessary, the effective address is calculated in step 88, and the write is executed in step 89. When it is determined in step 87 that the operand writing is not necessary and after the writing in step 89 is completed, the program counter is advanced by one instruction word in step 8A, and the process returns to step 81. In the case of the instruction level simulator shown in FIG. 8, the trace collection unit 1 is called immediately before the instruction fetch at step 81, immediately before the operand read at step 85, and immediately before the operand write at step 89. When calling before step 81, the instruction fetch as the access type and the address of the instruction word as the access address are passed to the trace collection means 1. When calling before step 85, the operand read is used as the access type and the operand address is used as the access address.
Pass to. When calling before step 89, the write operand is written as the access type, and the address of the operand is passed as the access address to the trace collection means 1.

[0022]

As described above, when the memory access frequency measuring method of the present invention is used, the memory access of the software operating on the target computer system for each access type and each continuous area in the address space is performed. Frequency becomes measurable.

By setting the code part of each subroutine as an area and looking at the number of instruction fetches for each area, the correct number of instructions executed in each subroutine can be found. This reveals the heavy parts of software processing. Also, by setting each variable as an area and looking at the number of operand accesses for each area, the number of times each variable is used can be found. This makes it clear which variables access is concentrated on. In this way, it is possible to accurately measure the number of instruction executions for each routine, and it is also possible to measure the load distribution such as the number of operand accesses.

In order to improve the performance of software, it is more effective to improve the heavy processing part, the frequently accessed variables and the data structure. Therefore, the memory access frequency measuring method of the present invention enables efficient performance tuning of software.

[Brief description of drawings]

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

FIG. 2 is a diagram showing the structure of a trace file F.

FIG. 3 is a diagram showing a structure of a region table T.

FIG. 4 is a flowchart showing the processing of the trace collection unit 1.

FIG. 5 is a flowchart showing the processing of the area setting means 2.

FIG. 6 is a diagram showing processing of profiling means 3.

FIG. 7 is a flowchart showing the processing of the profile printing means 4.

FIG. 8 is a diagram showing a method of embedding the trace collection means 1 in a target computer system.

FIG. 9 is a diagram for explaining a conventional technique.

[Explanation of symbols]

 1 trace sampling means 2 area setting means 3 profiling means 4 profile printing means F trace file T area table

Claims (1)

[Claims] 1. A trace file in which types and addresses of memory access are recorded, an area table including definition information of continuous areas and a counter for each access type, and a history of memory access during execution of target software in the trace file. Trace collecting means for recording, area setting means for registering a continuous area in the address space designated by the user in the area table, and each area registered in the area table from the memory access information recorded in the trace file A memory access frequency measuring method comprising a profiling means for counting the number of accesses to each type and a profile printing means for editing and printing the area table.