JP3951230B2 - Program allocation control method, program, and program allocation control system - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
  The present invention relates to multi-scheduling of jobs using a plurality of processors in a general-purpose or specific computer system.The
[0002]
[Prior art]
A parallel program executed using a plurality of processors (CPU, MPU) is divided into several processing units such as processes and threads, and each operates in parallel on another processor. The processing units operating in parallel on each processor derive one result as a whole while exchanging data with each other.
[0003]
In the parallel program executed in this way, execution is performed in synchronization with each checkpoint. At this time, if the partner processing unit does not reach the check point, data exchange with the partner processing unit cannot be performed, and so-called synchronization wait occurs.
[0004]
When multiple parallel programs are executed in parallel in a parallel computer system, scheduling each processing unit of the parallel program without mutual cooperation between the multiple processors will result in a difference in execution time difference between the processing units executed by each processor. Therefore, there is a problem that the synchronization waiting time becomes large.
[0005]
  To avoid this increase in synchronization waiting time, as a technique for performing a coordinated scheduling between processors, Satoshi Fukuda, “Parallel Operating System” (Corona Inc., issued in 1997) (ISBN4-339-02587-9) ) From page 56 to page 62is there.
[0006]
Here, a process scheduling technique is shown in which parallel processes of a parallel program are executed simultaneously during a time slice, then a global scheduler is called, a process to be executed next is selected, and each processor is notified. .
[0007]
In JP-A-10-74150, in order to execute parallel processes on a plurality of processors, a program allocation table in which the start and stop times of processes for a certain period are determined in advance is notified to each processor. Each processor describes a process scheduling technique in which processes are executed accordingly.
[0008]
However, most parallel programs require sequential processing (sequential processing) such as reading and writing data between parallel operations, especially in programs that read and write large amounts of data. The processing time of the sequential processing may take a long time compared to parallel processing, such as several ms to several seconds.
[0009]
In the above-described prior art, while such sequential processing is performed, processing units other than the processing unit performing sequential processing perform an idle loop that waits for the start of parallel arithmetic processing. In the program for reading and writing the data, the operation rate for the actual calculation of the computer system is lowered.
[0010]
[Problems to be solved by the invention]
The problem to be solved is that, in the conventional technology, while a sequential processing (sequential processing) is performed in a parallel program, processing units other than the processing unit performing this sequential processing are Since an idle loop that waits for the start is performed, for example, in a program that reads or writes a large amount of data, the processing time of one sequential processing may take a long time, and the operating rate for the actual calculation of the computer system decreases. This is the point.
[0011]
An object of the present invention is to solve these problems of the prior art and improve the operating rate of each processor in a computer system having a plurality of processors that execute a plurality of parallel programs and sequential programs in a multiplexed manner.
[0012]
[Means for Solving the Problems]
In order to achieve the above object, according to the present invention, during the sequential processing of parallel programs, an independent processor is allocated only to this sequential processing, and another parallel program or sequential program is allocated to the vacant processor. At the start, an independent processor is assigned simultaneously to all parallel operations of the parallel program. For example, when the sequential processing after the end of the parallel processing is not started for a certain time, an independent processor is assigned only to the sequential processing of the parallel program. As a result, the overhead due to frequent processor assignment changes is reduced at the start of a short time sequential process and the start of a parallel operation process. Also, at the time of periodic schedule switching, it is determined whether the next parallel program to be scheduled is in sequential processing or parallel arithmetic processing. During sequential processing, an independent processor is assigned only to the sequential processing of the parallel program, and parallel processing is performed. At the start of arithmetic processing, independent processors are simultaneously assigned to all parallel arithmetic processing of the parallel program. This reduces overhead due to processor allocation changes that occur in a fragmentary manner when each parallel processing is uneven.
[0013]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
[0014]
FIG. 1 is a block diagram showing a first configuration example of a program allocation control system according to the present invention.
[0015]
1 is a computer system including a CPU (Central Processing Unit), a main memory, a display device, an input device, an external storage device, and the like, and a storage medium such as a CD-ROM via an optical disk drive device or the like. After the program and data recorded in the external storage device are installed in the external storage device, each function for performing the multiprocessor schedule processing according to the present invention is constructed by reading this program from the external storage device into the main memory and processing it by the CPU.
[0016]
In FIG. 1, the processor group 1000 includes four processors 1100 to 1400 including, for example, a CPU and an MPU. As shown in the processor 1100, each of the processors 1100 to 1400 includes a barrier synchronization waiting mechanism 1110 that detects arrival of a synchronization point between the start and end of parallel arithmetic processing.
[0017]
The parallel program scheduler 2000 includes a processor allocated queue 3000, a processor allocation waiting queue 4000, a free processor count 5000 (counter), and a free processor bitmap 5100, and manages a parallel program schedule to be assigned to the processors 1100 to 1400. To do.
[0018]
Each parallel program, as shown in the parallel program 3100, is a table (in the figure and hereinafter referred to as “arithmetic processing”) 3110 for managing an arbitrary number of arithmetic processes, and whether each arithmetic process 3110 is performing an arithmetic operation. An operation flag 3120, an operation processing number 3130 of the parallel program, an operation processing execution number 3140 indicating the number of operation processing actually executing the parallel program, a priority 3150 of the parallel program, and a processor in the parallel program It holds a schedule flag 3160 indicating whether it is assigned or not.
[0019]
  In this example, it is assumed that the number of operation processing executions 3140 indicating the number of operation processes actually executing the parallel program is held, but for example, the number of operation start waits for waiting for the start of the parallel program operation processing, etc. The number of executions of operations that are actually executed in parallel programs can be obtained by simple calculations such asBeInformation may be retained.
[0020]
With such a configuration, in the program allocation control system of this example, when executing a parallel program that divides one continuous operation into a plurality of processing units (processes, threads, etc.) and allocates them to a plurality of processors for processing in parallel. When the start of the sequential calculation process in the parallel program is detected, a processor is assigned to the sequential calculation process, and the assignment of the processor to the parallel calculation process of the parallel program is suspended until the sequential calculation process is completed. .
[0021]
Hereinafter, the schedule processing procedure of the parallel program to each processor by the parallel program scheduler 2000 will be described with reference to FIGS.
[0022]
FIG. 2 is a flowchart showing an example of a parallel program schedule processing procedure by the parallel program scheduler in FIG.
[0023]
The parallel program scheduler 2000 in FIG. 1 periodically executes parallel programs 3100 to 3300 connected to the processor assigned queue 3000 and parallels connected to the processor assignment waiting queue 4000 according to the processor allocation rate determined by the priority 3150 or the like. A schedule switching process for replacing the programs 4100 to 4300 is performed.
[0024]
First, in step S101, it is determined whether or not a parallel program exists in the processor allocation waiting queue 4000. If no parallel program exists in the processor allocation waiting queue 4000 in step S101, the process sleeps until the next schedule timing is reached in step S107.
[0025]
If there is a parallel program in the processor allocation waiting queue 4000 in step S101 described above, the priority 3150 of the parallel program existing in the processor allocation waiting queue 4000 is set to a higher priority according to the processor allocation waiting time in step S102. Change to 3150.
[0026]
Next, in step S103, it is determined whether or not there is a parallel program in the processor assigned queue 3000. If there is a parallel program in the processor assigned queue 3000 in step S103, the priority 3150 of the parallel program existing in the processor assigned queue 3000 is changed to a lower priority 3150 in accordance with the processor assignment time in step S104. change.
[0027]
Next, in step S105, the parallel program of the processor assigned queue 3000 according to the priority 3150 of each parallel program of the processor assigned queue 3000 and the processor assignment waiting queue 4000, which has been changed in step S102 and step S104 described above. And a schedule selection process for performing a process of replacing the processor allocation waiting queue 4000 with the parallel program. Details of the schedule selection processing will be described later with reference to FIG.
[0028]
Next, in step S106, in accordance with the processor-allocated queue 3000 changed in the above-described step S105, a schedule switching process, which is a process for simultaneously assigning each arithmetic process of the parallel program to each processor, is called. Thereafter, in step S107, Sleep until next schedule timing. Details of the schedule switching process in step S106 will be described later with reference to FIG.
[0029]
FIG. 3 is a flowchart showing an example of a schedule selection process procedure in the schedule process of the parallel program to each processor by the parallel program scheduler of FIG.
[0030]
The processing example shown in FIG. 3 shows details of the schedule selection processing in step S105 in FIG. 2. First, in step S201, it is determined whether or not a parallel program exists in the processor assigned queue 3000. .
[0031]
If it is determined in step S201 that no parallel program exists in the processor-allocated queue 3000, step S210 and subsequent steps described later are processed. If a parallel program exists in the processor assigned queue 3000, it is determined in step S202 whether or not a parallel program exists in the processor assignment waiting queue 4000.
[0032]
If there is no parallel program in the processor allocation waiting queue 4000 in step S202, the schedule selection process is terminated in step S214.
[0033]
If there is a parallel program in the processor allocation waiting queue 4000 in step S202 described above, the parallel program (A) having the highest priority 3150 connected to the processor allocation waiting queue 4000 is selected in step S203.
[0034]
Next, in step S204, the number of executions 3140 of the parallel program (A) having the highest priority 3150 selected in step S203 is compared with the number of free processors 5000.
[0035]
Here, by using the number of executions of operation processing 3140 instead of the number of operations 3130 of the parallel program (A), the number of operations that are not actually performing an operation and are waiting for the start of parallel operations Are excluded from comparison with the number of free processors 5000. Details of the number of executions 3140 will be described later with reference to FIG.
[0036]
If the number of executions 3140 of the highest priority parallel program (A) is equal to or less than the number of free processors in step S204 described above, the highest priority parallel program (A) is converted to the processor in step S205. The queue is taken out from the allocation waiting queue 4000 and inserted into the processor allocated queue 3000. Thereafter, the process returns to step S202, and the process is repeated.
[0037]
If the number of executions 3140 of the highest priority parallel program (A) exceeds the number of free processors in step S204 described above, in step S206, the highest priority 3150 connected to the processor assigned queue 3000 is obtained. A low parallel program (B) is selected.
[0038]
In step S207, the priority of the parallel program (A) having the highest priority selected in step S203 described above and the priority of the parallel program (B) having the lowest priority selected in step S206 described above. Make a comparison.
[0039]
In step S207, the priority of the parallel program (A) having the highest priority selected in step S203 is equal to or lower than the priority of the parallel program (B) having the lowest priority selected in step S206. In this case, the process from step S210 described later is processed.
[0040]
In step S207, the priority of the parallel program (A) having the highest priority selected in step S203 is higher than the priority of the parallel program (B) having the lowest priority selected in step S206. In step S208, the parallel program (B) is taken out from the processor assigned queue 3000 and inserted into the processor assignment waiting queue 4000.
[0041]
In step S209, the number of executions 3140 of the parallel program (B) moved to the processor allocation waiting queue 4000 in step S208 is subtracted from the number of free processors 5000.
[0042]
Thereafter, the process returns to step S204, and the processes from step S204 to step S209 are repeated, and the schedule exchange process between the parallel program in the processor assigned queue 3000 and the parallel program in the processor assignment waiting queue 4000 is performed according to the priority.
[0043]
In step S210, that is, the priority of the parallel program (A) having the highest priority selected in step S203 described above in step S207 described above is the parallel program having the lowest priority selected in step S206 described above. If the priority is equal to or lower than the priority of the program (B), the parallel program (C) having the highest priority and connected to the processor allocation waiting queue 4000 and having the number of executions 3140 of the number of free processors 5000 or less is selected.
[0044]
In step S210, if there is no parallel program (C) having the highest priority among the queues assigned to the processor allocation waiting queue 4000 and the number of executions of operation processing 3140 equal to or less than 5000 free processors, step S211 is performed. As described above, the process proceeds to step S214 described above, and the schedule selection process ends.
[0045]
In step S211, if there is a parallel program (C) having the highest priority among the queues connected to the processor allocation waiting queue 4000 and the number of executions of operation processing 3140 being less than or equal to 5000 free processors, in step S212, The parallel program (C) is taken out from the processor allocation waiting queue 4000 and inserted into the processor allocated queue 3000.
[0046]
In step S213, the number of executions 3140 of the parallel program (C) moved to the processor-allocated queue 3000 in step S212 is subtracted from the number of free processors 5000. Thereafter, the process returns to step S210, and the schedule selection process for the free processor is repeated.
[0047]
FIG. 4 is a flowchart showing an example of a schedule switching process procedure in the parallel program schedule process for each processor by the parallel program scheduler of FIG.
[0048]
The processing example shown in FIG. 4 shows details of the schedule switching process in step S106 in FIG. 2. First, in step S301, the first parallel program (D) in the processor allocation waiting queue 4000 is selected.
[0049]
Next, in step S302, it is determined whether or not the parallel program (D) that is the selection target in the above-described step S301 exists. If the parallel program (D) does not exist, the processing after step S307 described later is performed. I do.
[0050]
If the parallel program (D) exists in step S302 described above, in step S303, the schedule flag 3160 of the parallel program (D) is “TRUE” indicating that the processor is allocated, or the processor assigns it. It is determined whether it is “FALSE” indicating that it has not been performed.
[0051]
If it is determined in step S303 that the schedule flag 3160 of the parallel program (D) is “FALSE”, the next parallel program (D) of 4000 in the processor allocation waiting queue is selected in step S306.
[0052]
If it is determined in step S303 that the schedule flag 3160 of the parallel program (D) is “TRUE”, in step S304, the arithmetic processing to which the processor of the parallel program (D) is assigned is stopped, and the arithmetic processing is started. Deallocate the processor.
[0053]
After that, in step S305, the schedule flag 3160 of the parallel program (D) is changed to “FALSE” indicating that no processor is allocated. In step S306, the next parallel program (D) in the processor allocation waiting queue 4000 is changed. Select.
[0054]
By repeating steps S302 to S306 in this manner, the processor assignment to all the arithmetic processes of all the parallel programs in the processor assignment waiting queue 4000 is released and stopped, and then in step S307, the processor assigned queue 3000 is released. The first parallel program (E) is selected.
[0055]
In the next step S308, it is determined whether or not the parallel program (E) selected in step S307 exists. If the parallel program (E) does not exist, the schedule switching process is terminated in step S313.
[0056]
If the parallel program (E) exists, in step S309, the schedule flag 3160 of the parallel program (E) is “TRUE” indicating that a processor is assigned, or a processor is assigned. It is determined whether it is “FALSE” indicating that it is not.
[0057]
If the determination result in step S309 is “TRUE” indicating that the schedule flag 3160 of the parallel program (E) is assigned to a processor, the next parallel program in the processor assigned queue 3000 is obtained in step S312. Select (E).
[0058]
If the schedule flag 3160 of the parallel program (E) is “FALSE” indicating that no processor is assigned in step S309 described above, in step S310, the arithmetic processing of the parallel program (E) is performed. Then, processor allocation processing is performed. The details of the processor assignment process will be described later with reference to FIG.
[0059]
Next, in step S311, the schedule flag 3160 of the parallel program (D) is changed to “TRUE” indicating that the processor is allocated. Thereafter, in step S312, the next parallel program (E) in the processor-allocated queue 3000 is selected, and the process returns to step S308 described above.
[0060]
Then, by performing the processing from step S308 to step S312 for the next parallel program (E) in the processor assigned queue 3000 selected in step S312, all the parallel programs in the processor assigned queue are all processed. An independent processor is assigned to the operation processing that is being executed.
[0061]
FIG. 5 is a flowchart showing an example of a processor assignment processing procedure in the parallel program schedule processing to each processor by the parallel program scheduler of FIG.
[0062]
The processing example shown in FIG. 5 shows the details of the processor assignment processing in step S310 in FIG. 4. First, in step S401, the first arithmetic processing (T) of the parallel program (E) to which the processor is assigned is executed. select.
[0063]
In step S402, it is determined whether or not the calculation process (T) selected in step S401 exists. If the calculation process (T) does not exist, the processor allocation process ends in step S411.
[0064]
Also, if there is a calculation process (T) in step S402, in step S403, the calculation flag 3120 of the calculation process (T) is “TRUE” indicating that the calculation process is being executed, or the calculation process is executed. It is determined whether it is “FALSE” indicating waiting. Details of the calculation flag 3120 will be described later with reference to FIG.
[0065]
If the result of the determination in step S403 is “FALSE” indicating that the operation flag 3120 of the operation process (T) is waiting for execution of the operation process, the processor is not allocated and the parallel program ( The next calculation process (T) of E) is selected.
[0066]
In addition, when the result of the determination in step S403 is “TRUE” indicating that the calculation flag 3120 of the calculation process (T) is being executed, in step S404, from the free processor bitmap 5100, One free processor (P) is selected. Then, the bit of the selected processor (P) is cleared in step S405.
[0067]
Next, in step S406, it is determined whether or not there is a parallel program waiting for processor allocation in the processor allocation waiting queue 4000.
[0068]
If there is a parallel program waiting for processor allocation, in step S407, the barrier synchronization waiting mechanism 1110 of the processor (P) is set to trap the issuance of a barrier synchronization waiting instruction. The details of the trap processing when the barrier synchronization wait instruction is issued will be described later with reference to FIG.
[0069]
If there is no parallel program waiting for processor assignment in the processor assignment wait queue 4000 in step S406, it is determined in step S408 whether or not the number of operation processes 3130 of the parallel program (E) is equal to the number of execution processes 3140. To do.
[0070]
If it is determined in step S408 that the number of arithmetic processes 3130 of the parallel program (E) is not equal to the number of arithmetic processes executed 3140, the process of step S407 is performed. That is, the barrier synchronization waiting mechanism 1110 of the processor (P) is set to trap the issuance of a barrier synchronization waiting instruction.
[0071]
Further, in the determination result in step S408 described above, when the number of arithmetic processings 3130 of the parallel program (E) is equal to the number of arithmetic processing executions 3140, the arithmetic processing (T) is dispatched to the processor (P) in step S409. (Assign).
[0072]
In step S410, the next calculation process (T) of the parallel program (E) is selected, the process returns to step S402, and the processes from step S402 to step S410 described above are repeated. In this way, the processor allocation process to the parallel program (E) is performed.
[0073]
Next, the parallel program control operation by the program allocation control system of this example will be described with reference to FIG.
[0074]
FIG. 6 is an explanatory diagram showing an example of the control operation of the parallel program in the program allocation control system of FIG.
[0075]
Arithmetic processing 3121 to 3124 of the parallel program 3100 are independently assigned to the four processors from the processors 1100 to 1400, and sequential processing and parallel arithmetic processing are alternately repeated in the flow from time T501 to T503. Running.
[0076]
The arithmetic processing 3121 of the parallel program 3100 is a representative arithmetic processing that performs sequential processing in addition to the parallel arithmetic processing, and the arithmetic processing 3122 to 3124 is arithmetic processing that performs only parallel arithmetic processing.
[0077]
First, at time T501, the arithmetic processing 3121 starts sequential processing, and the arithmetic processing 3122 to 3124 issue a barrier synchronization waiting instruction to wait for establishment of barrier synchronization indicating that the next parallel arithmetic processing is started.
[0078]
At this time, when the barrier synchronization waiting instruction trap is set by the processing in step S407 described with reference to FIG. 5, the barrier synchronization waiting trap processing described later with reference to FIG. 7 operates.
[0079]
Next, at time T502, the arithmetic processing 3121 completes the sequential processing and issues a barrier synchronization instruction to perform parallel arithmetic processing.
[0080]
By issuing the barrier synchronization command of this arithmetic processing 3121, the barrier synchronization waiting of each of the arithmetic processings 3122 to 3124 that issued the barrier synchronization at the above-described time T501 is canceled, and the parallel arithmetic processing is performed in each of the arithmetic processings 3121 to 3124. Start.
[0081]
Then, at time T503, the parallel calculation processing of each calculation processing 3121 to 3124 is completed, the calculation processing 3121 starts sequential processing, and the remaining calculation processing 3122 to 3124 issue a barrier synchronization instruction to perform the next parallel processing. It waits for the establishment of barrier synchronization indicating that the arithmetic processing has started.
[0082]
Hereinafter, the barrier synchronization waiting trap process will be described with reference to FIG.
[0083]
FIG. 7 is a flowchart showing an example of the barrier synchronization wait trap processing procedure in the parallel program schedule processing to each processor by the parallel program scheduler of FIG.
[0084]
The barrier synchronization waiting trap processing 600 shown in FIG. 7 is performed when each arithmetic processing 3122 to 3124 issues a barrier synchronization waiting instruction at the time T501 in FIG. FIG. 5 is a process for a barrier synchronization wait exception when it is set to trap the issuance of a barrier synchronization wait instruction in step S407 in FIG.
[0085]
First, in step S601, the calculation flag 3120 corresponding to the calculation process 3110 assigned to the processor in which the barrier synchronization waiting trap has occurred is changed to “FALSE”.
[0086]
Next, in step S602, 1 is subtracted from the operation processing execution count 3140 of the parallel program 3100 to which the operation processing 3110 assigned to the processor in which the barrier synchronization waiting trap has occurred. In step S603, the assignment of the arithmetic processing 3110 to the processor in which the barrier synchronization waiting trap has occurred is released.
[0087]
In step S604, it is determined whether the number of executions 3140 of the parallel program 3100 subtracted in step S602 described above is greater than “0”. When the arithmetic processing execution number 3140 is larger than “0”, steps from Step S607 described later are performed.
[0088]
If the operation processing execution number 3140 is “0”, the operation processing execution number 3140 of the parallel program 3100 to which the operation processing 3110 in which the trap has occurred belongs to the value of the operation processing number 3130 in step S605.
[0089]
Further, in step S606, the operation flags 3120 of all the operation processes of the parallel program 3100 to which the operation process 3110 in which the trap has occurred belong are changed to “TRUE”.
[0090]
Then, in step S607, it is determined whether a similar barrier synchronization waiting trap has occurred in another processor and is suspended. If there is a pending barrier synchronization waiting trap, the processing from step S601 to step S607 described above is repeated.
[0091]
If there is no pending barrier synchronization waiting trap, the parallel program scheduler process described above with reference to FIGS. 2 to 5 is performed in step S608, and the barrier synchronization waiting trap process is terminated in step S609.
[0092]
Next, the operation control of the parallel program by the parallel program scheduler of this example will be described with reference to FIG. 8, particularly in the case where a plurality of parallel programs are executed in multiple.
[0093]
That is, in the program allocation control system of FIG. 1, when a parallel program that divides one continuous operation into a plurality of processing units and processes them in parallel is executed by a plurality of processors, When the start of the sequential calculation process is detected, a processor is assigned to the sequential calculation process. However, for the parallel calculation process of the parallel program A, the processor assignment is canceled, and the processor thus unassigned is This is assigned to parallel processing of another parallel program B.
[0094]
FIG. 8 is an explanatory diagram showing an example of operation control of a plurality of parallel programs executed in the program allocation control system of FIG.
[0095]
In this example, it is assumed that two parallel programs of parallel program A and parallel program B are operating in a multiplex manner with four processors 1100 to 1400. In addition, it is assumed that the parallel program A is a four parallel program having four arithmetic processes, and the parallel program B is a three parallel program having three arithmetic processes.
[0096]
First, at time T701, the parallel program A is scheduled, and each of the arithmetic processing A3121 to A3124 of the parallel program A is assigned to the processors 1100 to 1400, and the parallel arithmetic processing is executed. At this time, the parallel program B is connected to the processor allocation waiting queue 4000 and is waiting for the next processor allocation.
[0097]
Next, at time T702, the parallel processing of the parallel program A is completed, the arithmetic processing A 3121 starts sequential processing, and the arithmetic processing A 3122 to A3124 of the parallel program A waits for barrier synchronization for the next parallel processing. Issue an instruction.
[0098]
By issuing this barrier synchronization wait instruction, a barrier synchronization wait trap occurs, and the barrier synchronization wait trap processing 600 described with reference to FIG. 7 assigns the arithmetic processing A 3122 to A 3124 of the parallel program A to the processors 1200 to 1400. By the operation of the parallel program scheduler 2000 released with reference to FIGS. 2 to 5, the arithmetic processing B 3121 to B 3123 of the parallel program B are assigned to the processors 1200 to 1400 and execute the parallel arithmetic processing of the parallel program B. .
[0099]
Further, at time T703, the sequential processing of the parallel program A is completed, and the arithmetic processing A 3121 issues a barrier synchronization wait instruction. By issuing this barrier synchronization wait instruction, a barrier synchronization wait trap is generated, and by the barrier synchronization wait trap processing 600 described with reference to FIG. 7, the assignment of the arithmetic processing A 3121 to the processor 1100 is once canceled. The operation of the parallel program scheduler 2000 described with reference to FIG. 5 determines the schedule of the parallel program A according to the priority in consideration of the processor usage and the like. The arithmetic processing A 3121 to A 3124 of the parallel program A are transferred to the processors 1100 to 1400. The parallel operation processing of the parallel program A is executed.
[0100]
At time T704, the parallel program scheduler 2000 starts a periodic operation as described with reference to FIG. 2, determines the schedule of the parallel program B according to the priority in consideration of the processor usage amount, and the like. B arithmetic processing B3121 to B3123 are assigned to the processors 1200 to 1400, and the parallel arithmetic processing of the parallel program B is executed.
[0101]
Next, another embodiment of the program allocation control system according to the present invention will be described with reference to FIGS.
[0102]
FIG. 9 is a block diagram showing a second configuration example of the program allocation control system according to the present invention.
[0103]
In the program allocation control system shown in FIG. 9, a barrier synchronization waiting timer 1120 is provided in each of the processors 1100 to 1400 in the program allocation control system shown in FIG.
[0104]
The multiprocessor scheduling method in this example includes a processor allocation processing procedure described in FIG. 10 in addition to the processing procedure of the parallel program scheduler described with reference to FIGS.
[0105]
Hereinafter, the processor assignment processing procedure in the program assignment control system having the configuration shown in FIG. 9 will be described with reference to FIG.
[0106]
FIG. 10 is a flowchart showing an example of a processor assignment processing procedure in the parallel program schedule processing to each processor by the parallel program scheduler in the program assignment control system of FIG.
[0107]
First, in step S801, the first arithmetic process (T) of the parallel program (E) to which the processor is assigned is selected, and in step S802, it is determined whether or not the selected arithmetic process (T) exists.
[0108]
If the arithmetic process (T) does not exist, the processor assignment process is terminated in step S811, and if the arithmetic process (T) exists, the arithmetic flag 3120 of the arithmetic process (T) is being executed in step S803. It is determined whether it is “TRUE” indicating that it is “FALSE” or “FALSE” indicating that it is waiting for execution of arithmetic processing.
[0109]
When the calculation flag 3120 of this calculation process (T) is “FALSE” indicating that the calculation process is waiting to be executed, the processor is not allocated and the next calculation process (E) of the parallel program (E) is performed (step S810). T) is selected, and if the calculation flag 3120 of the calculation process (T) is “TRUE” indicating that the calculation process is being executed, the processor in the empty state is determined from the empty processor bitmap 5100 in step S804. Select one (P).
[0110]
In step S805, the processor (P) bit selected in the above-described step S804 of the free processor bitmap 5100 is cleared. Further, in step S806, there is a parallel program waiting for processor assignment in the processor assignment wait queue 4000. Determine whether or not.
[0111]
If there is a parallel program waiting for processor allocation in the processor allocation waiting queue 4000, a barrier synchronization waiting timeout is set for the barrier synchronization waiting timer 1120 of the processor (P) in step S807. When a timeout for waiting for barrier synchronization occurs, the barrier synchronization waiting trap processing described with reference to FIG. 7 in the system having the configuration shown in FIG. 1 is performed.
[0112]
If there is no parallel program waiting for processor allocation in the processor allocation waiting queue 4000 in step S806 described above, it is determined in step S808 whether the number of arithmetic processings 3130 of the parallel program (E) is equal to the number of arithmetic processing executions 3140. To do.
[0113]
When the number of arithmetic processings 3130 of the parallel program (E) is not equal to the number of arithmetic processing executions 3140, in step S807, a barrier synchronization waiting timeout is set for the barrier synchronization waiting timer 1120 of the processor (P).
[0114]
Thereafter, or when the number of arithmetic processings 3130 of the parallel program (E) is equal to the number of arithmetic processing executions 3140, the arithmetic processing (T) is dispatched (assigned) to the processor (P) in step S809.
[0115]
In step S810, the next calculation process (T) of the parallel program (E) is selected, the process returns to step S802, and the processing from step S802 to step S810 is repeated to assign the processor to the parallel program (E). Process.
[0116]
Next, with reference to FIG. 11, regarding the operation control procedure of the parallel program by the parallel program scheduler of the present example, particularly when a plurality of parallel programs are executed in a multiplexed manner, An operation of assigning processors to sequential arithmetic processing and releasing / holding processor assignment to parallel arithmetic processing after a predetermined time has elapsed will be described.
[0117]
FIG. 11 is an explanatory diagram showing an example of operation control of a plurality of parallel programs executed in a multiple manner in the program allocation control system of FIG.
[0118]
Here, it is assumed that two parallel programs of parallel program A and parallel program B are operating in a multiplex manner with four processors 1100 to 1400. In addition, it is assumed that the parallel program A is a four parallel program having four arithmetic processes, and the parallel program B is a three parallel program having three arithmetic processes.
[0119]
First, at time T901, the parallel program A is scheduled, and each of the arithmetic processes A3121 to A3124 of the parallel program A is assigned to the processors 1100 to 1400 to execute the parallel arithmetic process. At this time, the parallel program B is connected to the processor allocation waiting queue 4000 and is waiting for the next processor allocation.
[0120]
At the next time T902, the parallel arithmetic processing of the parallel program A is completed, the arithmetic processing A 3121 starts sequential processing, and the arithmetic processing A 3122 to A3124 of the parallel program A are barrier synchronization waiting instructions for the next parallel arithmetic processing. Issue.
[0121]
By issuing this barrier synchronization wait instruction, the operation of the barrier synchronization wait timer 1120 is started. The barrier synchronization waiting timer 1120 generates a barrier synchronization waiting trap when the timer becomes “0” after being subtracted with the passage of time.
[0122]
At time T903, the sequential processing of the parallel program A is completed, and the arithmetic processing A 3121 issues a barrier synchronization wait instruction. As a result, barrier synchronization is established by issuing a barrier synchronization wait in all the arithmetic processes A3121 to A3124, and parallel arithmetic processes are started all at once.
[0123]
Furthermore, at time T904, the parallel operation processing of the parallel program A is completed, the operation processing A 3121 starts sequential processing, and the operation processing A 3122 to A 3124 issue a barrier synchronization wait instruction for the next parallel operation processing. By issuing this barrier synchronization wait instruction, the operation of the barrier synchronization wait timer 1120 is started.
[0124]
At time 905, the barrier synchronization wait timer is set to “0” in each of the processors 1200 to 1400 where the arithmetic processing A 3122 to A 3124 of the parallel program A is executing the barrier synchronization instruction, and a barrier synchronization wait trap is generated.
[0125]
By this barrier synchronization waiting trap processing 600, the assignment of the arithmetic processing A3122 to A3124 of the parallel program A to the processors 1200 to 1400 is canceled, and the parallel program scheduler 2000 causes the arithmetic processing B3121 to B3123 of the parallel program B to be changed to the processors 1200 to 1200. 1400 is assigned to execute parallel arithmetic processing of the parallel program B.
[0126]
Next, another (third) embodiment of the program allocation control system according to the present invention will be described with reference to FIGS.
[0127]
FIG. 12 is a block diagram showing a third configuration example of the program allocation control system according to the present invention.
[0128]
In the program allocation control system shown in FIG. 12, a barrier synchronization status 1130 is provided for each of the processors 1100 to 1400 in the program allocation control system shown in FIG.
[0129]
The multiprocessor scheduling method in this example includes the schedule switching processing procedure described with reference to FIGS. 13 to 15 in addition to the processing procedure of the parallel program scheduler described with reference to FIGS. 2 and 3.
[0130]
Hereinafter, the processor allocation processing procedure in the program allocation control system having the configuration shown in FIG. 2 will be described with reference to FIGS.
[0131]
FIG. 13 is a flowchart showing an example of a schedule switching process procedure in the parallel program schedule process for each processor by the parallel program scheduler in the program allocation control system of FIG.
[0132]
First, in step S1001, the top parallel program (D) of the processor allocation waiting queue 4000 is selected. In step S1002, whether or not the top parallel program (D) of the processor allocation waiting queue 4000 to be selected exists is determined. to decide.
[0133]
When the parallel program (D) does not exist, the processing after step S1007 described later is processed. When the parallel program (D) exists, the schedule flag 3160 of the parallel program (D) is assigned to the processor in step S1003. Whether it is “TRUE” indicating that the processor is assigned or “FALSE” indicating that the processor is not allocated.
[0134]
When the schedule flag 3160 is “FALSE”, the next parallel program (D) of 4000 in the processor allocation waiting queue is selected in step S1006, and when the schedule flag 3160 is “TRUE”, step S1004 is selected. In FIG. 4, processor assignment cancellation processing is performed for the arithmetic processing of the parallel program (D). Details of the processor assignment cancellation processing will be described later with reference to FIG.
[0135]
In the next step S1005, the schedule flag 3160 of the parallel program (D) is changed to “FALSE” indicating that the processor is not assigned, and in step S1006, the next parallel program (in the processor assignment waiting queue 4000) ( D) is selected.
[0136]
Then, by repeating the processes in steps S1002 to S1006 described above, the processor assignment to all the arithmetic processes of all the parallel programs in the processor assignment waiting queue 4000 is released and stopped, and then in step S1007. The first parallel program (E) in the processor-allocated queue 3000 is selected.
[0137]
Next, in step S1008, it is determined whether or not the first parallel program (E) in the processor-allocated queue 3000 described in step S1007 described above exists.
[0138]
If the parallel program (E) does not exist, the schedule switching process is terminated in step S1013. If the parallel program (E) exists, the schedule flag 3160 of the parallel program (E) is assigned to the processor in step S1009. Whether it is “TRUE” indicating that the processor is assigned or “FALSE” indicating that the processor is not allocated.
[0139]
If the schedule flag 3160 of the parallel program (E) is “TRUE”, indicating that the processor is allocated, the next parallel program (E) of the processor allocated queue 3000 is selected in step S1012, and from step S1008 Return to the process.
[0140]
If the schedule flag 3160 of the parallel program (E) is “FALSE” indicating that no processor is allocated, in step S1010, processor allocation processing is performed for the arithmetic processing of the parallel program (E). The details of the processor assignment process will be described later with reference to FIG.
[0141]
Thereafter, in step S1011, the schedule flag 3160 of the parallel program (D) is changed to “TRUE” indicating that the processor is allocated, and in the next step S1012, the next parallel program ( E) is selected.
[0142]
Thereafter, returning to the processing in step S1008, the processing from step S1009 to step S1012 is repeated for the next parallel program (E) in the processor-allocated queue 3000 selected in step S1012, and the processor has already been allocated. An independent processor is assigned to an operation process that is executing all operations of all parallel programs in the queue and operated.
[0143]
Next, with reference to FIG. 14, the processor assignment deallocation processing procedure in step S <b> 1004 in FIG. 13 will be described.
[0144]
FIG. 14 is a flowchart showing an example of a processor allocation deallocation process procedure in the parallel program schedule process for each processor by the parallel program scheduler of FIG.
[0145]
First, in step S1101, the first arithmetic processing (T) of the parallel program (D) whose processor assignment is to be canceled is selected.
[0146]
In step S1102, it is determined whether or not the calculation process (T) selected in step S1101 exists. If the calculation process (T) does not exist, the processor deallocation process ends in step S1109.
[0147]
Also, in the above-described step S1102, if the calculation process (T) exists, in step S1103, whether the calculation flag 3120 of the calculation process (T) is “TRUE” indicating that the calculation process is being executed, It is determined whether it is “FALSE” indicating that the process is waiting to be executed.
[0148]
When the calculation flag 3120 of the calculation process (T) is “FALSE”, indicating that the calculation process is waiting to be executed, the processing after step S1108 described later is performed without deallocating the processor.
[0149]
When the calculation flag 3120 of the calculation process (T) is “TRUE” indicating that the calculation process is being executed, the barrier synchronization status 1130 of the processor to which the calculation process (T) is assigned is referred to in step S1104. Then, it is determined whether the calculation process (T) is waiting for barrier synchronization.
[0150]
If the barrier synchronization status 1130, that is, the calculation process (T) is waiting for the barrier synchronization, in step S1105, the calculation flag 3120 of the calculation process (T) is changed to “FALSE” indicating that the calculation process is waiting to be executed. After that, or when the barrier synchronization status 1130, that is, the calculation process (T) is not waiting for the barrier synchronization, the processor dispatch to the calculation process (T) is canceled in step S1106.
[0151]
In step S1107, the bit corresponding to the processor to which the calculation process (T) of the free processor bitmap 5100 is assigned is set. Further, in step S1108, the next calculation process (T) of the parallel program (D) is selected, and the process returns to step S1102.
[0152]
By repeating the processing from step S1102 to step S1208, processor assignment cancellation processing to the parallel program (D) is performed.
[0153]
Next, the processor allocation processing procedure in step S1010 in FIG. 13 will be described with reference to FIG.
[0154]
FIG. 15 is a flowchart showing an example of a processor assignment processing procedure in the parallel program schedule processing to each processor by the parallel program scheduler of FIG.
[0155]
First, in step S1201, the first arithmetic processing (T) of the parallel program (E) to which the processor is assigned is selected, and then in step S1202, it is determined whether or not the arithmetic processing (T) selected here exists. To do. If the calculation process (T) does not exist, the processor allocation process ends in step S1210.
[0156]
If the calculation process (T) exists, in step S1203, the calculation flag 3120 of the calculation process (T) is “TRUE” indicating that the calculation process is being executed, or the calculation process is waiting to be executed. Whether it is “FALSE” indicating “” is determined.
[0157]
When the calculation flag 3120 of the calculation process (T) is “FALSE” and the calculation process (T) indicates that the calculation process is waiting to be executed, the processes after step S1209 described later are performed without assigning processors.
[0158]
If the calculation flag 3120 of the calculation process (T) is “TRUE” and the calculation process (T) indicates that the calculation process is being executed, in step S1204, the empty processor bitmap 5100 indicates that the processor ( One P) is selected, and in step S1205, the bit of the processor (P) selected in the above-described step S1204 of the empty processor bitmap 5100 is cleared.
[0159]
In step S1206, it is determined whether or not the number of arithmetic processes 3130 of the parallel program (E) is equal to the number of arithmetic processes executed 3140. If the number of arithmetic processings 3130 of the parallel program (E) is not equal to the number of arithmetic processing executions 3140, in step S1207, a barrier synchronization waiting timeout is set for the barrier synchronization waiting mechanism 1110 of the processor (P). When a timeout for waiting for barrier synchronization occurs, the barrier synchronization waiting trap processing described with reference to FIG. 7 is performed.
[0160]
Thereafter, or when the number of operations 3130 of the parallel program (E) is equal to the number of operations executed 3140 in step S1206, the operation (T) is dispatched to the processor (P) in step S1208.
[0161]
In step S1209, the next arithmetic processing (T) of the parallel program (E) is selected, the processing returns to step S1202, and the processing from step S1202 to step 1209 is repeated to assign the processor to the parallel program (E). Process.
[0162]
Next, with reference to FIG. 16, the parallel program operation control procedure by the parallel program scheduler of this example, particularly when a plurality of parallel programs are executed in a multiplexed manner, to the processor of the parallel program at a predetermined cycle. If the sequential calculation process is being executed in the parallel program before the change that was determined to be changed in the next cycle, the operation of assigning this sequential calculation process to the free processor in the next cycle is performed. explain.
[0163]
FIG. 16 is an explanatory diagram showing an example of operation control of a plurality of parallel programs executed in a multiple manner in the program allocation control system of FIG.
[0164]
Here, it is assumed that two parallel programs of parallel program A and parallel program B are operating in a multiplex manner with four processors 1100 to 1400.
[0165]
In addition, it is assumed that the parallel program A is a four parallel program having four arithmetic processes, and the parallel program B is a three parallel program having three arithmetic processes.
[0166]
First, at time T1301, the parallel program A is scheduled, and the arithmetic processes A3121 to A3124 of the parallel program A are assigned to the processors 1100 to 1400 to execute the parallel arithmetic process.
[0167]
At this time, the parallel program B is connected to the processor allocation waiting queue 4000 and is waiting for the next processor allocation.
[0168]
At the next time T1302, the parallel arithmetic processing of the parallel program A is completed, the arithmetic processing A 3121 starts sequential processing, and the arithmetic processing A 3122 to A3124 of the parallel program A are barrier synchronization waiting instructions for the next parallel arithmetic processing. Is issued.
[0169]
Then, at time T1303, a periodic operation of the parallel program scheduler occurs, the assignment of the processors 1100 to 1400 to the parallel program A is released according to the priority in consideration of the processor usage and the like, and the processor 1200 is assigned to the parallel program B. ˜1400 is allocated.
[0170]
In this processor deallocating process to the parallel program A, the arithmetic processes A 3122 to A 3124 are waiting for barrier synchronization and can be executed by only one processor to the arithmetic process A 3121, so that the arithmetic process A 3121 is transferred to the empty processor 1100. Assigned.
[0171]
As described above with reference to FIGS. 1 to 16, in this example, when executing a parallel program in which one continuous operation is divided into a plurality of processing units and assigned to a plurality of processors for parallel processing, When the start of the sequential calculation process in the program is detected, a processor is assigned to the sequential calculation process, but the assignment of the processor to the parallel calculation process of the parallel program is suspended until the sequential calculation process is completed.
[0172]
Alternatively, when the start of the sequential operation processing in the parallel program A being processed is detected and a processor is assigned to the sequential operation processing, the processor assignment to the parallel operation processing of the parallel program A is released, and the processor that has released this assignment Are assigned to parallel processing of other parallel programs.
[0173]
In addition, after detecting the start of the sequential calculation process, it is detected that a predetermined time has elapsed, and then, for example, processor allocation to the sequential calculation process and processor allocation release to the parallel calculation process are executed. Alternatively, it is determined whether or not the allocation of the parallel program to the processor needs to be changed at a predetermined cycle, and if the sequential calculation processing is being executed in the parallel program before the change determined to be changed at the next cycle, Sequential arithmetic processing is assigned to an empty processor in the next cycle.
[0174]
Thus, in this example, during sequential processing, an independent processor is allocated only to the sequential processing of the parallel program, and another parallel program or sequential program is allocated to the free processor. By simultaneously assigning independent processors to all the parallel computation processes of a program, the processor utilization rate for actual computations can be improved.
[0175]
Also, it has a function to detect that the waiting time for the start of the parallel processing has reached a certain time, and if the sequential processing after the end of the parallel processing is not started for a certain time, the parallel processing By assigning independent processors only to the sequential processing of programs, overhead due to frequent processor allocation changes can be reduced at the start of short-time sequential processing and parallel arithmetic processing.
[0176]
In addition, it has a function to statically detect that it is waiting for the start of parallel computation. When switching schedules, it determines whether the next parallel program to be scheduled is in sequential processing or parallel computation processing, and is in sequential processing. Assigns independent processors only to sequential processing of parallel programs, and at the start of parallel computing processing, assigns independent processors to all parallel computing processing of parallel programs at the same time, so that each parallel computing processing is uneven. In such a case, overhead due to a change in processor allocation that occurs in a fractional manner can be reduced.
[0177]
As a result, in a computer system using a plurality of processors that execute a plurality of parallel programs and sequential programs in a multiplexed manner, it is possible to improve the operating rate of the processor for actual calculations.
[0178]
In addition, this invention is not limited to the example demonstrated using FIGS. 1-16, and can be variously changed in the range which does not deviate from the summary. For example, in this example, the number of processors is described as four, but the number of processors is not limited to four.
[0179]
In this example, the processing unit allocated to a plurality of processors (CPU, MPU) may be a process, a thread, or each parallel program unit.
[0180]
Also, as a computer configuration example, a computer configuration without a keyboard or optical disk drive may be used. In this example, an optical disk is used as a recording medium. However, an FD (Flexible Disk) or the like may be used as a recording medium. As for the program installation, the program may be downloaded and installed via a network via a communication device.
[0181]
【The invention's effect】
According to the present invention, in a computer system using a plurality of processors that execute a plurality of parallel programs and sequential programs in a multiplexed manner, it is possible to improve the operating rate of the processor for actual computation.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a first configuration example of a program allocation control system according to the present invention.
FIG. 2 is a flowchart showing an example of a parallel program schedule processing procedure by the parallel program scheduler in FIG. 1;
FIG. 3 is a flowchart showing an example of a schedule selection process procedure in a parallel program schedule process for each processor by the parallel program scheduler of FIG. 1;
4 is a flowchart showing an example of a schedule switching process procedure in a parallel program schedule process for each processor by the parallel program scheduler of FIG. 1;
5 is a flowchart showing an example of a processor assignment process procedure in a parallel program schedule process for each processor by the parallel program scheduler of FIG. 1;
6 is an explanatory diagram showing an example of a control operation of a parallel program in the program allocation control system of FIG. 1. FIG.
7 is a flowchart showing an example of a barrier synchronization wait trap processing procedure in a parallel program schedule processing to each processor by the parallel program scheduler of FIG. 1;
8 is an explanatory diagram showing an example of operation control of a plurality of parallel programs executed in a multiplexed manner in the program allocation control system of FIG. 1. FIG.
FIG. 9 is a block diagram showing a second configuration example of the program allocation control system according to the present invention.
10 is a flowchart showing an example of a processor assignment process procedure in a parallel program schedule process for each processor by a parallel program scheduler in the program assignment control system of FIG. 9;
11 is an explanatory diagram showing an example of operation control of a plurality of parallel programs executed in a multiplexed manner in the program allocation control system of FIG. 9;
FIG. 12 is a block diagram showing a third configuration example of the program allocation control system according to the present invention.
13 is a flowchart showing an example of a schedule switching process procedure in a parallel program schedule process for each processor by a parallel program scheduler in the program allocation control system of FIG. 12;
14 is a flowchart showing an example of a processor deallocation process procedure in a parallel program schedule process for each processor by the parallel program scheduler of FIG. 12;
15 is a flowchart showing an example of a processor assignment processing procedure in the parallel program schedule processing to each processor by the parallel program scheduler of FIG. 12;
16 is an explanatory diagram showing an example of operation control of a plurality of parallel programs executed in a multiple manner in the program allocation control system of FIG. 12;
[Explanation of symbols]
1000: Processor group, 1100 to 1400: Processor, 1110: Barrier synchronization wait mechanism, 1120: Barrier synchronization wait timer, 1130: Barrier synchronization status, 2000: Parallel program scheduler, 3000: Processor assigned queue, 3100-3300, 4100 4300: Parallel program, 3110: Operation processing (management table), 3120: Operation flag, 3130: Number of operation processing, 3140: Number of execution of operation processing, 3150: Priority, 3160: Schedule flag, 4000: Processor allocation waiting queue, 5000 : Number of free processors (counter), 5100: Free processor bitmap.

Claims (6)

A plurality of parallel programs to be processed in parallel by dividing one consecutive operations on a plurality of processing units, respectively, a program allocation control method of a computer system that executes the multiplexed plurality of processors,
A first step of detecting completion of parallel arithmetic processing in the parallel program A being processed and start of sequential arithmetic processing by the first processor ;
Upon detecting the start of the completion and the sequential processing of the parallel processing in the first step, a second step of releasing the allocation of processors other than the first processor for parallel processing of the parallel program A When,
A processor cancels the allocation, assignment to parallel processing of other parallel program B, that perform parallel processing of said parallel program B in parallel with the sequential processing of the parallel program A according to the first processor A program allocation control method comprising: a third step. The program allocation control method according to claim 1,
As a parallel program B for assigning a processor whose deallocation has been canceled,
Among the parallel programs other than the parallel program A, the number of operation processing executions indicating the number of operation processes actually executing the parallel program is equal to or less than the number of free processors, and the pre-assigned allocation priority is the highest A program allocation control method comprising selecting a parallel program. A program allocation control method according to claim 1 or 2 , wherein:
After detecting the start of the sequential calculation process in the first step, detecting that the sequential calculation process has passed a predetermined time;
A program allocation control method comprising: performing the processing in the second step and the processing in the third step after the detection in the step. A program allocation control method according to any one of claims 1 to 3, comprising:
Determining whether it is necessary to change the assignment of the parallel program to the processor at a predetermined cycle; and
A program assignment control, comprising: a step of assigning the sequential calculation process to an empty processor in the next period if the sequential calculation process is being executed in the parallel program before the change determined to be required in the next period. Method.   The program for making a computer perform each step in the program allocation control method in any one of Claims 1-4. A plurality of parallel programs to be processed in parallel by dividing one consecutive operations on a plurality of processing units, respectively, a program assignment control system of the computer system to run multiple in multiple processors,
First means for detecting completion of parallel arithmetic processing in the parallel program A being processed and start of sequential arithmetic processing by the first processor ;
Upon detecting the start of the completion and the sequential processing of the parallel processing by the first means, to deallocate the processors other than the first processor for parallel processing of the parallel program A, releasing the allocation was processor, assigned to the parallel processing of other parallel programs B, said parallel program second means parallel processing to run in parallel with the sequential processing of the parallel program a according to the first processor B And a program allocation control system.

Images