JP2012032986A - Compile method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a compile method and a program capable of more reducing a used amount of a storage medium necessary for collection of profile information than before.SOLUTION: A CPU contained in a front end node executes: an insertion step for inserting an aggregation code for aggregating profile information acquired by each parallel computation device and corresponding to a designated range in a parallel computation program in a parallel computation device allocated to each designated range in the parallel computation program and a compression and integration code for compressing and integrating the aggregated profile information into the parallel computation program; and a compile step for compiling the parallel computation program into which the aggregation code and the compression code are inserted.

Description

  The present invention relates to a profile method and a program.

  2. Description of the Related Art Conventionally, a technique for analyzing the performance of a parallel program running on a parallel computer composed of a plurality of processors and outputting profile information is known. In addition, a technique for deleting an overlapping portion of trace data and a technique for designating the number of execution times of trace data are known.

  As a program that operates on a large-scale parallel computer, an SPMD (Single-Program, Multiple-Data) program is known. A large-scale parallel computer includes one front-end node and a plurality of (for example, tens of thousands) computation nodes configured by computers. The front end node and the plurality of calculation nodes are connected to each other via a network. The front-end node includes an SPMD program and a compiler. The compiler creates an executable file by compiling the SPMD program.

  This type of massively parallel computer executes an executable file according to the following procedure. First, the front-end node determines the maximum value of the calculation node number to be used. That is, the maximum number of calculation nodes to be used is determined. Next, the front end node copies the executable file created by the compiler to each compute node. Thereafter, each compute node executes the executable file.

  By the way, in order to investigate the behavior at the time of execution of the executable file created from the SPMD program, the front-end node may acquire profile information. In the existing technology, profile information is acquired according to the following procedure. First, each computation node acquires the number of executions at a certain position of the SPMD program and a certain range of execution time while executing the executable file, and when the execution of the executable file ends, information on the acquired number of executions and execution time is obtained. Is stored as profile information (procedure 1). Next, at the end of execution of the executable file, the front-end node acquires profile information stored in each calculation node and creates one profile information (procedure 2). In this way, the front-end node acquires one profile information.

JP-A-10-63550

JP 2001-147833 A

JP-A-10-11321

  However, since the data size of the detailed profile information collected at each calculation node is huge, one profile information created in the procedure 2 has a huge amount of data. In particular, in the SPMD code, the same processing is often executed by a plurality of calculation nodes. Therefore, a large amount of redundant information is included in the profile information collected individually at each calculation node. Therefore, the capacity of the storage medium necessary for collecting the profile information, that is, the capacity of the memory or the hard disk may be wasted in each calculation node and front-end node.

  In view of the above problems, it is an object of the disclosed compiling method, program execution method, and program to reduce the amount of storage medium used necessary for collecting profile information as compared with the related art.

  In order to achieve the above-described object, the disclosed compiling method is a method in which profile information corresponding to a specified range of a parallel calculation program acquired by each of the parallel computers is stored in a specified range of the parallel calculation program. An insertion step of inserting into the parallel calculation program an aggregated code that is aggregated to a parallel computer assigned for each time and a compressed integrated code that compresses and integrates the aggregated profile information; and the aggregated code and the compressed code are inserted And a compiling step of compiling a parallel computing program.

  The disclosed program is obtained by assigning profile information corresponding to a designated range of a parallel computing program acquired by each of the parallel computers to a parallel computer assigned to each designated range of the parallel computing program. An insertion step of inserting into the parallel calculation program an aggregate code to be aggregated and a compressed integrated code for compressing and integrating the aggregated profile information, and compiling to compile the parallel computation program into which the aggregate code and the compressed code are inserted Process.

  According to the disclosed compiling method and program, it is possible to reduce the usage amount of the storage medium necessary for collecting profile information as compared with the related art.

It is a schematic block diagram of the parallel computer which concerns on this Embodiment. It is a schematic block diagram of the front end node 2 and the calculation nodes 3-0 to 3-N. 3 is a flowchart showing processing executed by a compiler 12. 2 is a diagram illustrating an example of an SPMD program 11. FIG. It is a figure which shows the example of the function call graph corresponding to the SPMD program 11, and the control flow graph of each function. It is a figure which shows the example of the function call graph corresponding to the SPMD program 11, and the control flow graph of each function. It is a figure which shows the example of the function call graph corresponding to the SPMD program 11, and the control flow graph of each function. It is a figure which shows the example of the function call graph corresponding to the SPMD program 11, and the control flow graph of each function. It is a figure which shows the example of the function call graph corresponding to the SPMD program 11, and the control flow graph of each function. It is a flowchart which shows the acquisition process of profile information. It is a figure which shows the state of the profile information which each calculation node 3-0 3-3 has. It is a figure which shows the state of each calculation node 3-0 -3-3 after several profile information is compressed and integrated. It is a figure which shows the state in which the profile information aggregated to the calculation node 3-0 and the calculation node 3-3 exists. It is a figure which shows the state which aggregated the profile information compressed and integrated by the calculation node 3-0 and the calculation node 3-3 to one profile information. It is a figure which shows an example of the code | cord | chord containing a profile information processing function (profile_start and profile_end). 4 is a diagram illustrating an example of an executable file 13. FIG. It is a figure which shows an example of the code containing the function PART. It is a figure which shows an example of the code containing function CHOOSE_MAX.

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

  FIG. 1 is a schematic configuration diagram of a parallel computer according to the present embodiment.

  In FIG. 1, a parallel computer 1 includes a front-end node 2 and computation nodes 3-0 to 3-N (N: natural number) as parallel computing devices. The front-end node 2 and the computation nodes 3-0 to 3-N are hardware such as a computer and are connected to each other via the network 4. The front-end node 2 includes an SPMD (Single-Program, Multiple-Data) program 11, a compiler 12, and an executable program 13, which are programs for parallel computation. Note that “K” in FIG. 1 indicates the node number of each computation node. For example, “K = 0” indicates that the node number of the calculation node 3-0 is 0, and “K = N” indicates that the node number of the calculation node 3-N is N (N: natural number). Indicates.

  Each computation node has a CPU 21-K and a memory 22-K. K is a node number corresponding to each calculation node. The detailed configuration of each calculation node will be described later.

  The SPMD program 11 is a program for parallel calculation. The SPMD program 11 uses all node numbers divided into several groups. Therefore, the number of executable programs created based on the processing related to the node number in the SPMD program 11 is smaller than the number of calculation nodes.

  The compiler 12 compiles the SPMD program 11 and creates an executable program 13. Note that the number of executable programs created by the compiler 12 is not limited to one. A plurality of executable programs may be created. The created executable program 13 is transmitted to the computation nodes 3-0 to 3-N and executed by each computation node.

  FIG. 2 is a schematic configuration diagram of the front-end node 2 and the calculation nodes 3-0 to 3-N. Note that the configuration of each of the computation nodes 3-0 to 3-N is the same as that of the front end node 2, and therefore the description thereof is omitted.

  The front end node 2 includes a CPU 21 that controls the entire front end node 2 and a memory 22 that includes various information and programs. The front-end node 2 also includes an OS, a hard disk drive (HDD) 23 having various information and programs, and a communication interface 24 for connecting each calculation node. Further, the front-end node 2 includes an input unit 25 that inputs data from an input device 31 such as a mouse and a keyboard, and an output unit 26 that outputs data and messages to an output device 32 such as a monitor or a printer. . The CPU 21 is connected to the memory 22, HDD 23, communication interface 24, input unit 25, and output unit 26 via the bus 27.

  The SPMD program 11, the compiler 12, and the executable program 13 in FIG. 1 are stored in the memory 22 or the HDD 23. The compiler 12 operates under the control of the CPU 21. The front end node 2 may include a data read / write device such as a CD-R drive, a nonvolatile ROM, and the like in addition to these elements.

  FIG. 3 is a flowchart showing processing executed by the compiler 12. FIG. 4 is a diagram illustrating an example of the SPMD program 11. In the SPMD program 11 of FIG. 4, the function H includes a communication process SEND () and a communication process RECEIVE ().

  In FIG. 3, the compiler 12 creates a function call graph of the entire program corresponding to the SPMD program 11 and a control flow graph of each function (step C1). The function call graph is a graph showing a function call relationship written in the source code of the SPMD program 11. The control flow graph is a graph that represents all paths that may be taken when the program is executed.

  Here, FIG. 5 shows a function call graph corresponding to the SPMD program 11 of FIG. 4 and a control flow graph of each function. In FIG. 5, the control flow is represented by a solid line, and the function call and the return from the function are represented by a dotted line. In FIG. 5, the main function includes basic blocks 41 to 44 (code piece A, function G call, function H call, code piece B). The function G includes basic blocks 421 to 424, and the function H includes basic blocks 431 to 434. The basic block is a minimum unit for performing compiler optimization.

  Next, the compiler 12 determines a range for acquiring profile information (hereinafter referred to as a target range) (step C2). The profile information is information that summarizes the number of times the SPMD program 11 is executed and a certain range of execution time. In the present embodiment, the profile information includes a calculation node number, a range name, position identification information, and an execution time. Details of the calculation node number, range name, position identification information, and execution time will be described later. The target range is, for example, a function unit or a basic block unit. The range in which profile information is acquired is specified by the program using compile options. The compile option is an option that is specified at the time of execution in order to accurately inform the compiler of information known to the programmer. Here, it is assumed that the target range is set in function units and basic block units.

  Next, the compiler 12 assigns a range name to the target range, inserts a code for starting the measurement of profile information at the start point of the target range, and inserts a code for ending the measurement of profile information at the end point of the target range ( Step C3). In the example of FIG. 6, the compiler 12 assigns the range name Rn (n = 0 to 11) to the target range. Further, the compiler 12 inserts into the basic block a code (profile_start) for starting measurement of profile information and a code (profile_end) for ending measurement of profile information.

  The compiler 12 uses the function call graph and the control flow graph to determine whether or not communication processing is included in each target range. The compiler 12 changes the round function for rounding the execution time of the program included in the target range, which is set in the profile_start code inserted in step C3, according to the presence / absence of communication processing (step C4). In FIG. 7, since the basic blocks 432 and 433 include communication processing (SEND () and RECEIVE ()), the function H also includes communication processing. In FIG. 7, the basic blocks 432 and 433 including the communication process and the function H are displayed by shading.

The compiler 12 determines the profile information processing position of each target range using the function call graph and the control flow graph (step C5). The profile information processing position is a position where profile information obtained from each target range is dumped (output). The profile information processing position for a certain target range R is a position X on the program that satisfies all of the following conditions (1) to (4).
(1) Program position X reached after execution of the target range R
(2) When the target range R is executed, it always passes through the position X. (3) During the execution of the program, the same position identification information (position identification information will be described later) passes through the position X only once (4). After passing through the position X, the same position identification information does not pass through the target range R again. In general, one profile information processing position covers a plurality of target ranges. An example in which the profile information processing position of each target range is inserted into the function call graph and the control flow graph is shown in FIG. In FIG. 8, the profile information processing position is indicated by a profile_dump code.

  The compiler 12 groups a plurality of profile information processing positions (step C6). FIG. 9 shows an example in which a plurality of profile information processing positions are grouped. In FIG. 9, the compiler 12 sets a plurality of profile information processing positions so that the profile information processing position is set immediately below the loop composed of basic blocks that call the function G or function H or immediately below the last basic block included in the functions G and H. Profile information processing positions are grouped.

  Finally, the compiler 12 compiles the SPMD program 11 into which the profile_start code, profile_end code, and profile_damp code are inserted, and creates an executable file 13 (step C7).

Steps C <b> 1 to C <b> 6 are processes in which the compiler 12 automatically incorporates a mechanism for acquiring profile information into the SPMD program 11. Steps C1 to C6 can be performed directly by the programmer instead of being executed by the compiler 12. That is, the specification of the target range and the specification of the profile information processing position can be realized by one of the following three methods.
(1) The compiler automatically inserts the range specification function and processing position into the source code of the SPMD program 11 (automatic).
(2) The programmer writes the range specification function and processing position in the source code of the SPMD program 11 (manual).
(3) The programmer writes the range specification function and processing position of the necessary part in the source code of the SPMD program 11, and the compiler automatically processes the remaining part (semi-automatic).

  Next, a process in which each of the calculation nodes 3-0 to 3-N executes the executable program 13 and acquires profile information will be described. FIG. 10 is a flowchart showing the profile information acquisition process.

  First, each calculation node 3-0 to 3-N starts execution of the executable program 13 (step E1). The executable program 13 is a program obtained by compiling the SPMD program 11 according to the processing of FIG.

  When each calculation node 3-0 to 3-N reaches the profile information processing function, it determines the type of the function (step E2). The profile information processing function is profile_start, profile_end, or profile_dump.

  When the profile information processing function is profile_start, each of the calculation nodes 3-0 to 3-N starts measuring the execution time of the program included in the target range (step E3).

  When the profile information processing function is profile_end, each of the calculation nodes 3-0 to 3-N ends the measurement of the execution time of the program included in the target range and stores the profile information in the memories 22-0 to 22-N. (Step E4).

  When the profile information processing function is profile_dump, each of the calculation nodes 3-0 to 3-N transmits profile information of a plurality of target ranges to each of the predetermined calculation nodes, and the acquired plurality of profiles The information is compressed and integrated (step E5). Here, calculation nodes 3-0 to 3-N are used in order as nodes for storing profile information. For example, the profile information of the range name (R0) is transmitted to the calculation node 3-0, and the profile information of the range name (R1) is transmitted to the calculation node 3-1. The profile information of the range name (Rn) is transmitted to the calculation node 3-N. When the number n of target ranges is larger than the total number N of calculation nodes, the profile information of the target range corresponding to the remainder when the number n of target ranges is divided by the total number N of calculation nodes is the calculation node corresponding to the remainder. Sent to the number calculation node. For example, if the target range is 102 and the total number of calculation nodes is 100, the profile information of the 101st and 102nd target ranges is transmitted to the calculation nodes of calculation node numbers 1 and 2, respectively.

  Each of the computation nodes 3-0 to 3-N repeats the processes of steps E2 to E5 until the executable program 13 ends (step E6). Finally, at the end of execution of the executable program 13, the front-end node 2 aggregates the profile information compressed and integrated by the calculation nodes 3-0 to 3-N into one (step E7).

  Next, profile information acquisition processing will be described in detail with reference to FIGS. Here, four calculation nodes 3-0 to 3-3 are used.

  After the execution of the executable program 13 is started, each of the computation nodes 3-0 to 3-3 determines that the profile information processing function that arrives first is profile_start (R0,...) From FIG. Step E1,2). Therefore, each calculation node 3-0 to 3-3 starts measuring the execution time of the program included in the target range of the range name R0 (step E3). Next, since each of the calculation nodes 3-0 to 3-3 reaches the profile information processing function profile_end (R0,...), The measurement of the execution time of the program included in the target range of the range name R0 is terminated. (Step E4). The profile information obtained in steps E3 and E4 is shown in Table 1 below.

表１の計算ノード番号は、計算ノードの番号Ｋを表す。範囲名は、関数profile_startと関数profile_endで指定される対象範囲の名前を表す。位置識別情報は、対象範囲に到達した履歴を表す履歴情報である。ここでは、対象名Ｒ０の対象範囲が関数main()内に存在することを示す（main）が位置識別情報である。実行時間は、各計算ノード３−０〜３−３で対象範囲のプログラムを実行するのにかかった時間を表す。各計算ノード３−０〜３−３が、表１の実行時間の１の位を四捨五入して、実行時間を１０の倍数に丸めた場合の、各計算ノード３−０〜３−３が有するプロファイル情報の状態を図１１に示す。

The calculation node number in Table 1 represents the calculation node number K. The range name represents the name of the target range specified by the function profile_start and the function profile_end. The position identification information is history information representing a history of reaching the target range. Here, (main) indicating that the target range of the target name R0 exists in the function main () is the position identification information. The execution time represents the time taken to execute the program in the target range at each of the calculation nodes 3-0 to 3-3. Each calculation node 3-0 to 3-3 has the execution time rounded to a multiple of 10 by rounding the execution time in Table 1 to the first decimal place. The state of profile information is shown in FIG.

  When each of the calculation nodes 3-0 to 3-3 reaches the profile information processing function profile_dump (R0, R1), the calculation node 3-0 aggregates the profile information of the target range of the range name R0, and displays the aggregated profile information. Compress and integrate (step E5). FIG. 12 shows a state of each of the calculation nodes 3-0 to 3-3 after a plurality of profile information is compressed and integrated. In FIG. 12, the computation node 3-0 holds profile information having the same execution time in a compressed format. At this time, since the profile information of the target range of the range name R0 is all collected in the calculation node 3-0, the profile information of the target range of the range name R0 is stored in the other calculation nodes 3-1 to 3-3. not exist.

  The above-described processing of steps E1 to E5 is also executed for the program in the target range of the target names R1 to R3 (step E6). Therefore, at the end of execution of the executable program 13, the profile information aggregated in each of the calculation nodes 3-0 to 3-3 is distributed. For example, FIG. 13 shows a state where profile information aggregated in the computation nodes 3-0 and 3-3 exists. In FIG. 13, the computation node 3-3 holds the profile information of the target range of the range name R3 in a compressed format.

  At the end of execution of the executable program 13, the front end node 2 aggregates a plurality of profile information compressed and integrated by the calculation nodes 3-0 and 3-3 into one profile information (step E7). FIG. 14 shows a state in which the profile information compressed and integrated by the calculation node 3-0 and the calculation node 3-3 is integrated into one profile information.

  As described above, since each of the calculation nodes 3-0 to 3-3 partially compresses and integrates the profile information at step E5, the data amount of the profile information finally created at step E7 is greatly increased. Can be reduced. In addition, since a plurality of calculation nodes are used for the processing of steps E1 to E5, the processing of aggregation, compression, and integration of profile information can be executed in parallel.

  Next, the profile information processing function will be described.

  The compiler 12 determines a range (target range) in which the execution time profile information is acquired in the SPMD program 11 in step C2. At this time, the compiler 12 designates the target range using a code including profile information processing functions (profile_start and profile_end) shown in FIG.

  Here, the meaning of each parameter of the profile information processing function used for specifying the target range is as follows. REGION_NAME in FIG. 15 represents a name for distinguishing the target range. SELECTOR in FIG. 15 specifies a function for selecting profile information. The argument of the function is position identification information. If the calculation nodes 3-0 to 3-3 do not select the profile information, NULL is specified for SELECTOR.

  COMPRESSOR in FIG. 15 specifies a function for compressing and integrating profile information. The code including the function specified by COMPRESSOR of the function profile_start functions as a compression integrated code. The function argument is the execution time. For example, when the function ROUND10 is designated as COMPRESSOR, the execution time is t, the 1's place of the measured execution time t is rounded off, and the execution time t is rounded to a multiple of 10. When the function ROUND10000 is designated as COMPRESSOR, the thousands of the measured execution time t is rounded off, and the execution time t is rounded to a multiple of 10,000.

  ANALYZER in FIG. 15 specifies a function for extracting profile information to be analyzed from the compressed profile information. The function argument is a table of compressed profile information. If analysis target profile information is not extracted from the compressed profile information, NULL is specified for ANALYZER.

  These parameters are available by default by the compiler 12. Also, by specifying options in the compiler 12, these parameters can be replaced with user-specified processing.

  Further, the compiler 12 designates the profile information processing position by “profile_dump (REGION_NAME)” (step C5). The function profile_dump is a library function that performs the following processes (1) and (2) at the time of execution.

  Process (1): Collect profile information of the target range of the range name REGION_NAME distributed to the respective computation nodes 3-0 to 3-N. For this purpose, the computation nodes 3-0 to 3-N communicate with each other. For example, in FIG. 11, the computation node 3-0 collects profile information of the target range of the range name R0 distributed to the computation nodes 3-1 to 3-3. As described above, the code including the function profile_dump functions as an aggregation code for aggregating profile information in the target range specified by the SPMD program 11.

  Process (2): The profile information collected by each computation node 3-N is compressed and integrated using the function specified by the parameter of profile_start. For example, in FIG. 11, the computation node 3-0 collects profile information of the target range of the range name R0 distributed to the computation nodes 3-1 to 3-3, compresses the collected profile information, and Integrate. The profile information after compression and integration is as shown in FIG.

  Further, the compiler 12 may specify a list of range names (that is, a plurality of range names) as the range name REGION_NAME. In this case, the above processes (1) and (2) are executed for a plurality of target ranges. When a plurality of range names are specified, since the computation nodes 3-0 to 3-3 compress and integrate the profile information of the target ranges of the range names R0 to R3, respectively, compression and integration of the plurality of profile information are performed in parallel. can do. Therefore, as compared with the conventional method of collecting and processing all profile information in the front-end node, it is possible to reduce the computing node resources (that is, the amount of memory and disk used in each computing node) to be used. it can. In addition, the processing speed is improved by parallel processing of a plurality of calculation nodes.

  Next, the parameter COMPRESSOR in FIG. 15 will be described.

  For example, when the four calculation nodes 3-0 to 3-3 execute the executable file 13 as shown in FIG. 16, the calculation node 3-0 acquires the profile information of Table 1 described above. When profile information is analyzed, the detailed difference in execution time is not important, so the computation node 3-0 does not store the actual execution time as shown in Table 1, but instead uses the parameter COMPRESSOR (here function Save the execution time rounded by ROUND10). Table 2 shows the result of the calculation node 3-0 rounding the execution time according to the parameter COMPRESSOR.

表２に示すように、同じ実行時間のプロファイル情報を有する計算ノードが多数存在する可能性が高まる。

As shown in Table 2, there is a high possibility that there are a large number of computation nodes having the same execution time profile information.

  In step C4, the compiler 12 determines whether or not communication processing is included in the target range, and sets a different parameter COMPRESSOR (round function) to the parameter COMPRESSOR of profile_start inserted in step C3 according to the determination result. Can be set. For example, in the target range including the communication processing in FIG. 9 (basic blocks 432 and 433 and the basic block 43 including the call to the function H), the compiler 12 sets the parameter to round the execution time with the function ROUND10000 instead of the function ROUND10. You can change COMPRESSOR. In this case, in step C4, the compiler 12 sets the function ROUND10000 in the parameter COMPRESSOR of the profile_start when the communication process is included in the target range, and the profile_start parameter when the communication process is not included in the target range. Set the function ROUND10 to COMPRESSOR. Thereby, the execution time in the profile information of the target range including the communication process is a multiple of 10,000, and the execution time in the profile information of the target range not including the communication process is a multiple of 10.

  As a result, when referring to the aggregated profile information, the profile information analyst or the analysis apparatus can distinguish the target range including the communication process from the target range not including the communication process at a glance. In particular, the features of the SPMD program 11 can be grasped. The function for rounding the execution time t is not limited to ROUND10 and ROUND10000, but may be ROUND100, ROUND1000, or ROUND100000.

  Next, the parameter SELECTOR in FIG. 15 will be described.

  For example, in FIG. 9, the target range (basic block 421) of the range name R4 included in the function G may be executed a plurality of times by the loop of the main function. For this reason, position identification information is used to identify the number of rotations of the loop (that is, the number of executions of the target range). Since the target range of the range name R4 may be executed from the loop of the main function (hereinafter referred to as L0) by the number of rotations specified by the loop index i, the position identification is used to identify the number of rotations of the loop. Information is represented by (main, L0, i). For example, Table 3 shows profile information when the loop of the main function rotates three times at the calculation nodes 3-0 to 3-3.

ループの回転数が多い場合は、プロファイル情報のデータ量が膨大になるため、すべてのプロファイル情報を保存するのは現実的ではない。このような場合に、コンパイラ１２は、ステップＣ３で挿入したprofile_startのパラメータSELECTORに取捨選択関数として関数PARTを設定する。関数PARTは、位置識別情報をプロファイル情報の取捨選択の判定に利用する。また、例えば、最初の２回のループのプロファイル情報を取得する場合は、コンパイラ１２は、図１７に示すような、関数PARTを含むコードをＳＰＭＤプログラム１１に設定する。

When the number of rotations of the loop is large, the amount of profile information data is enormous, so it is not realistic to save all the profile information. In such a case, the compiler 12 sets the function PART as a selection function in the parameter SELECTOR of profile_start inserted in step C3. The function PART uses position identification information for determination of selection of profile information. For example, when acquiring the profile information of the first two loops, the compiler 12 sets a code including the function PART as shown in FIG.

  In FIG. 17, the position information P has a list format, and P [x] takes out the x-th element. In FIG. 17, when the function PART set in the parameter SELECTOR returns 1, the execution time of the target range is measured, and when it returns 0, the execution time of the target range is not measured. When the compiler 12 sets the function PART shown in FIG. 17 to profile_start of the target range of the range name R4, the calculation nodes 3-0 to 3-3 acquire the profile information shown in Table 4 (step E4).

最終的に、フロントエンド２は表５に示すプロファイル情報を取得する（ステップＥ７）。

Finally, the front end 2 acquires the profile information shown in Table 5 (step E7).

取捨選択関数としての関数PARTは、各計算ノード３−０〜３−ＮがステップＥ３で実行可能プログラム１３の実行時間の測定を開始するときに、各計算ノード３−０〜３−Ｎが指定されたループの回転数に対応する対象範囲の実行時間の取得を開始するかどうかを判断するために利用される。

The function PART as the selection function is designated by each calculation node 3-0 to 3-N when each calculation node 3-0 to 3-N starts measuring the execution time of the executable program 13 in step E3. This is used to determine whether or not to start the execution time of the target range corresponding to the number of rotations of the loop that has been made.

  Next, the parameter ANALYZER in FIG. 15 will be described.

  Consider a case in which the front end 2 acquires not the profile information shown in Table 5 but profile information in which the number of rotations of the loop is 0 or 1 and the execution time of the target range of the range name R4 is the longest. In this case, the compiler 12 sets the function CHOOSE_MAX as an analysis function in the parameter ANALYZER of profile_start inserted in step C3, and sets the code including the function CHOOSE_MAX as shown in FIG. In the function CHOOSE_MAX, the profile information whose execution time is 10 is deleted from the data shown in Table 4, so that the profile information shown in Table 6 is finally created.

分析関数としての関数CHOOSE_MAXは、各計算ノード３−０〜３−ＮがステップＥ５で対象範囲のプロファイル情報を統合する場合に利用することができる。

The function CHOOSE_MAX as the analysis function can be used when each calculation node 3-0 to 3-N integrates the profile information of the target range in step E5.

  The sorting function is used in step E3, and the analysis function is used in step E5. With these functions, the profile information acquired by each of the calculation nodes 3-0 to 3-N is reduced. When each calculation node 3-0 to 3-N uses a selection function or an analysis function, the profile is compared with the method in which the front-end node 2 extracts only specific profile information after aggregating all the profile information. It is possible to reduce the amount of temporary memory and disk space required for acquiring information.

  Regarding the above embodiment, the following additional notes are disclosed.

  (Supplementary Note 1) Profile information corresponding to a specified range of the parallel calculation program acquired by each of the parallel calculation devices is assigned to the parallel calculation device assigned to each specified range of the parallel calculation program. An insertion step of inserting into the parallel calculation program an aggregate code to be aggregated and a compressed integrated code for compressing and integrating the aggregated profile information, and compiling to compile the parallel computation program into which the aggregate code and the compressed code are inserted A compiling method characterized in that the process is executed. According to such a configuration, when each parallel computing device executes the parallel computing program as an executable file, each parallel computing device compresses and integrates the acquired plurality of profile information. It is possible to reduce the amount of storage medium used necessary for collecting data.

  (Supplementary Note 2) In the insertion step, at least one of a selection code for selecting each profile information or an analysis code for extracting profile information to be analyzed from the compressed and integrated profile information is used as a parallel calculation program. The compiling method according to Supplementary Note 1, wherein the compiling method is inserted. According to such a configuration, the profile information acquired by each parallel computing device is reduced, so that it is possible to reduce the usage amount of the storage medium necessary for collecting profile information as compared with the related art.

  (Additional remark 3) The said compression integration code rounds the information of the execution time contained in the said profile information by a different number of digits according to whether the designated range of the said parallel calculation program includes communication processing. 3. The compiling method according to appendix 1 or 2, which is a feature. According to such a configuration, the profile information analyst or the analysis device can clearly distinguish the range including the communication processing from the range not including the communication processing, and thus can efficiently grasp the characteristics of the parallel calculation program. it can.

  (Supplementary Note 4) A function call graph of the entire parallel computing program and a control flow graph of each function are created, and at least one of a function unit or a basic block unit included in the function call graph and the control flow graph, the profile 4. The compiling method according to any one of appendices 1 to 3, further comprising a specifying step of specifying a range for acquiring information. According to such a configuration, it is possible to designate a range in which profile information is automatically acquired when each parallel computing device executes a parallel computing program.

  (Supplementary Note 5) Profile information corresponding to the designated range of the parallel computing program acquired by each of the parallel computing devices is assigned to the parallel computing device assigned to each designated range of the parallel computing program. An insertion step of inserting into the parallel calculation program an aggregate code to be aggregated and a compressed integrated code for compressing and integrating the aggregated profile information, and compiling to compile the parallel computation program into which the aggregate code and the compressed code are inserted A program characterized by causing a process to be executed. The same effect as the above supplementary note 1 is achieved.

  1 parallel computer, 2 front-end node, 3-1 to 3-N computation node, 11 SPMD program, 12 compiler, 13 executable program, 21 CPU, 22 memory, 23 hard disk drive (HDD), 24 communication interface, 25 input Part, 26 output part

Claims (4)

On the computer,
An aggregation code for collecting profile information corresponding to a designated range of a parallel computing program acquired by each of the parallel computing devices to a parallel computing device assigned for each designated range of the parallel computing program, and the aggregation An insertion step of inserting a compression integration code for compressing and integrating the profile information obtained into the parallel calculation program;
And a compiling step of compiling the parallel calculation program into which the aggregated code and the compressed code are inserted.   The inserting step further includes inserting at least one of a selection code for selecting each profile information or an analysis code for extracting analysis target profile information from the compressed and integrated profile information into the parallel calculation program. The compiling method according to claim 1, wherein:   The compression integration code rounds the execution time information included in the profile information with a different number of digits according to whether or not a specified range of the parallel calculation program includes communication processing. Item 3. The compiling method according to item 1 or 2. On the computer,
An aggregation code for collecting profile information corresponding to a designated range of a parallel computing program acquired by each of the parallel computing devices to a parallel computing device assigned for each designated range of the parallel computing program, and the aggregation An insertion step of inserting a compression integration code for compressing and integrating the profile information obtained into the parallel calculation program;
And a compiling step of compiling the parallel calculation program into which the aggregated code and the compressed code are inserted.

Images