JPH11353287A - Scheduling system for process of parallel system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the throughput of the whole system by determining an optimum processor at an I/O request from a process as for the scheduling of processes which are being executed in a parallel computer system consisting of plural processors. SOLUTION: A process generating/ending function 12 allows its processor to generate and end a process. An I/O load monitor function 14 monitors the quantity of I/O requests from processes operating on the processor. A processor selecting function 8 selects an optimum computing processor from an I/O load state management table 10 and sends an answer to the request source. A process moving function 13 moves an arbitrary process 17 which is being executed by the processor to an arbitrary processor 2. In this constitution, the switching overhead of a network at I/O requests of respective parallel processes can be considered, so the network use efficiency can be improved and the throughput of the whole system can be improved.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a process scheduling method in a parallel computer system in which tens to tens of thousands of processors operate in combination.

[0002]

2. Description of the Related Art In a parallel computer system comprising a plurality of processors dedicated to I / O processing and a plurality of processors dedicated to calculation processing, each having an independent memory and connected by a three-dimensional network, an input / output device is used. The only processors that can be connected are processors dedicated to I / O processing. I / O requests from processes running on processors dedicated to computational processing are
This is realized by making an I / O request to an I / O processing dedicated processor to which a target input / output device is connected via a three-dimensional network.

In a three-dimensional network, an X-axis, a Y-axis, and a Z-axis are connected in a grid pattern, and a switch connecting each axis is attached to each intersection, and a communication path of the network is determined by a switching operation of the switch. You. X
Once each in the axial, Y-, and Z-axis directions (total of 3
If the switching operation is performed, it is possible to communicate with the target processor.

[0004] A processor dedicated to calculation processing for making I / O requests and
Depending on the relative position of the processor dedicated to I / O processing, the case where I / O requests can be made only on the X axis, the case where I / O requests can be made on the X axis-Y axis, and the X axis-Y axis-Z axis There are three types where I / O requests can be made using The switching overhead of the network at this time is (X axis) <(X axis−Y axis) <
(X axis-Y axis-Z axis), and the efficiency is highest when an I / O request is made only in a single axis direction (X axis in this case).

However, a plurality of input / output devices are connected to a plurality of processors dedicated to I / O processing, and a running parallel computing process is connected to any input / output device connected to any I / O processing dedicated processor. It cannot be determined whether an I / O request is made until an actual process is executed and an I / O request is made. For this reason, process creation for an operating parallel computing process determines the processor that creates the process, independent of the switching overhead of the three-dimensional network. Also, once the process starts running,
Since the running process is not moved to another processor, the I
/ O requests from the running processor independent of the 3D network switching overhead
An I / O request has been made.

As a technique relating to process scheduling which is an object of the present invention, Japanese Patent Laid-Open No. 5-1217 is disclosed.
Although there is Japanese Patent Publication No. 3 (Japanese Patent Application Publication No. HEI 9-209), this is a technique relating to CPU load.

[0007]

In the above-described conventional method, an I / O request from a parallel computing process operating on a processor dedicated to computation processing to a specific input / output device is connected to the specific input / output device. Process allocation is performed irrespective of the switching overhead of the network until an I / O request is made to a processor dedicated to I / O processing, and once a processor to execute a process is determined, the running process is assigned to another processor. Therefore, if the switching overhead of the network is large, the processing efficiency in the processor part is reduced, and a reduction in throughput is inevitable.

In general, a parallel calculation process synchronizes processes assigned to a plurality of processors and processes them. Therefore, when a local decrease in the throughput occurs on a processor basis, the waiting time for the processing of other processors increases.
This causes a decrease in the throughput of the entire system.

An object of the present invention is to monitor I / O requests from a process executed on each processor dedicated to calculation,
For I / O requests above the threshold, the I / O
The process is rearranged so that the switching overhead of the network in the communication of the / O request becomes appropriate, thereby improving the throughput of the entire system.

[0010]

FIG. 1 is a block diagram of the entire system, and FIG. 2 is a block diagram of the present invention showing details of FIG. It is assumed that n dedicated processors 2
And one system management processor 1 and input / output device 15
Are connected at the intersection of a high-speed three-dimensional network 4 for performing communication between the processors, and each processor is connected to the (X-axis, Y-axis,
The coordinates of the (Z axis) are assigned as the processor number 7. Further, the processor number 7 is a unique number in the system. Here, a case where one processor 3 dedicated to I / O processing is arranged on the X axis will be described. A process 17 operating on the computing processor 2 under a parallel computer having an independent memory for each processor
(Hereinafter referred to as “processor management 11”) operates on each calculation-dedicated processor 2;
In the system management processor 1, a control program (hereinafter, referred to as system management 5) for managing the entire system operates.

The process 17 is generated by the system management 5 requesting the processor management 11 to generate a process. When an I / O request is issued from the running process 17, the processor management 11 communicates the I / O request to the I / O processing dedicated processor 3 and enters the I / O processing dedicated processor 3. The output device control function 18 processes the actual I / O. Each process has a unique process number 20 in the system. Which process is executed by which calculation processor is managed by the system management 5 in a correspondence table between the process number 20 and the processor number 3 (hereinafter referred to as a process management table 9).
When communicating between the processes, it is assumed that the process management table 9 is searched to find the dedicated processor 2 on which the process of the transmission destination is being executed, and communicate with the processor.

In order to realize the present invention, each processor management 11 includes an "I / O load monitoring function 14" for monitoring an I / O request amount of a corresponding processor and a "process migration" for migrating a process to another processor. A function 13 "and a" processor selection function 8 "for selecting an optimal processor 2 dedicated to calculation relating to I / O requests are provided.

In the case of an I / O request from the running process 17, the processor management 11 communicates the I / O request to the processor 3 dedicated to I / O processing to which the input / output device is connected. , A processor 3 dedicated to I / O processing becomes an input / output device
Perform I / O.

When an I / O request exceeding the threshold is made, the processor selection function 8 of the processor management 11 is used to minimize the switching overhead of the network when the I / O request is made. Is selected, the process 17 being executed is moved to the corresponding processor, and the processing is continued by the destination dedicated processor for calculation.

[0015] Thus, I
It is possible to prevent a local decrease in throughput due to the switching overhead of the network in the communication of the / O request.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 2 is a block diagram of one embodiment of the present invention. In FIG. 2, a system management processor 1, each processor 2 and an I / O processing dedicated processor 3 are connected by a high-speed three-dimensional network 4, and all the connected processors can communicate with each other and exchange information. .

The system management 5 of the system management processor 1 has a process management function 6. The process management function 6 is a function for managing a process management table 9 for associating a generated process number with a processor number. When communicating between processes, a process of a communication partner exists through this function. It is possible to communicate only with the process number without being aware of the processor.

Processor management 1 for each dedicated processor
1 is provided with a process generation / termination function 12, a processor selection function 8, a process migration function 13, and an I / O load monitoring function 14 for monitoring the I / O load status of the corresponding dedicated processor 2.

The process generation / termination function 12 is a function for generating and ending a process in its own processor. The I / O load monitoring function 14 is a function of monitoring an I / O request amount from a process running on the own processor. The processor selection function 8 has a function of selecting an optimum calculation processor from the I / O load status management table 10 and responding to the request source. The process migration function 13 is a function for migrating an arbitrary process 17 running on its own processor to an arbitrary processor 2.

FIG. 3 shows a configuration example of the process management table 9. It comprises a process number 21, a processor number 22 in which the process exists, and a process state 23 indicating the operation status of the process. Further, the processor number 22 includes an X coordinate 24, a Y coordinate 25, and a Z coordinate 26.

FIG. 4 shows an example of the configuration of the I / O load status management table 10 of each processor dedicated to calculation. I
It is composed of a process number 31 of the I / O request source, an I / O processing dedicated processor number 32 of the I / O request destination, and a total value 33 of the I / O amount to the I / O processing dedicated processor.

The I / O processing dedicated processor number 32 is composed of an X coordinate 34, a Y coordinate 35, and a Z coordinate 36.

FIG. 5 shows an example of the configuration of the system parameter table 16. An I / O amount threshold value 41 serving as a condition for moving a process, and an activation time interval 42 of a processor selection process 8 for determining whether to move a processor for executing a process when the threshold value is exceeded.
Are given as system parameters at the time of system startup and distributed from the system management processor 1 to each processor.

Next, the operation of this embodiment will be described with reference to the drawings.

When an I / O request exceeds a threshold value from a process running on a certain processor, a process managed by that processor is moved to another processor. Will be described with reference to FIGS. 2, 4, and 5. FIG.

When an I / O request is issued from a running process, the I / O load monitoring function 14 in the processor management 11 of the dedicated processor 2 causes the I / O load
The I / O request destination of the I / O load status management table 10
The processor number 32 of the processing dedicated processor, the process number 31 of the I / O request source, and the total value 34 of the I / O request amount are updated.

Further, the processor management 11 performs the following processing for each process being executed at every activation time interval 42 of the processor selection processing 8 defined in the system parameter table 16.

Processor selection function 8 for each process number
Start

The processor selection function 8 executes the I / O load status monitoring
The following processing is performed for a process that requests I / O beyond the threshold value of the / O amount. Communication between the I / O processing dedicated processor, which is the I / O request destination of the process, and the three-dimensional network 4 of the computation dedicated processor that executes the process is (X
The processor number of the processor dedicated to calculation, which can be communicated only by the axis), is selected, and the process movement function 13 is called. The process moving function 13 temporarily suspends the running process, and executes the moving process of the running process to the dedicated processor 2 selected by the processor selecting function 8. Thereafter, the processor management 11 of the migration-dedicated computation processor 2 resumes the interrupted processing.

FIG. 6 is a flow chart for explaining the operation of the processor selecting function 8 for selecting a dedicated processor for which a process is to be moved. Hereinafter, the flowchart of FIG. 6 will be described with reference to FIG. 2 and FIG. When the processor selection function 8 is activated, n is used to position the reference pointer n of the I / O load status management table 10 at the first record.
= 1 (step 51) Next, the total value 33 of the I / O request amount in the I / O load status management table 10 is compared with the threshold value 41 of the I / O amount in the system parameter table 16 (step 52). When the total value 33 of the I / O request amount> the threshold value 41 of the I / O amount, the own processor number (X, Y, Z) is set as the selected processor number (X, Y, Z) ( Step 5
4). Next, the Y coordinate of the selected processor number is replaced with the Y coordinate 35 of the I / O processing dedicated processor number, and the Z coordinate of the selected processor number is replaced with the Z coordinate 36 of the I / O processing dedicated processor number (step 55).

The selected processor number (X, Y, Z) is compared with its own processor number (X, Y, Z) (step 56). If they are not equal, the processor transfer function 13 is called to select the selected processor number (X , Y, Z) (step 57). After returning from the processor migration function 8, the reference pointer n is positioned to the next record in order to perform the same processing for the next record in the I / O load status management table 10 (step 58). It is checked whether there is a next record (step 59), and if there is a next record, the same processing is repeated.

If there is no next record, the processor selection function 8 ends. Also, in step 52, when the total value of the I / O request amount 33 <the threshold value 41 of the I / O amount,
If the selected processor number (X, Y, Z) is equal to the own processor number (X, Y, Z) in step 6, the processing of the record is stopped, and step 58 is performed to process the next record. .

FIG. 7 is a conceptual diagram illustrating the operation of the process moving function 13 for moving a process being executed by its own processor to another processor. In FIG. 7, the system management processor will be described as “processor number 000”, the source processor as “processor number 111”, and the destination processor as “processor number 122”.

Processor number 1, which is the source processor
When the process migration function 13 starts the process migration process, a, the system management 5 of the system management processor 1
B to start moving the process. Upon receiving the report, the system management 5 sets the process management function 6
Sets the process state 23 of the process to be moved in the process management table 9 to "pending state", and responds to the process moving function 13 of the source "processor number 111" that the transition of the process to the pending state has been completed. D.

In response, the transfer source “processor number 0
After the process to be moved is stopped, the process movement function 13 of “01” moves to the destination “processor number 12”.
2) process information 19 for the process movement function 13;
Forward process 17 f. The process migration function 13 which has received these information sets the transferred process information 1
9, the process 17, the memory area 18 is restored to the state before the migration, g is released after the process is released from the halt state, and the system management 5 is notified that the migration of the process is completed i.

In response to the report, the system management 5 uses the process management function 6 to release the “pending state” of the process state 23 in the process management table 9 and to change the processor number 22 to the “processor number 122” as the transfer destination. , And reply that the release of the suspended state of the process has been completed. In response to this, the destination “processor number 122”
The process migration function 13 of the migration source “processor number 1”
11 ”that reports the completion of the process migration to the process migration function 13 of m. In response to this, the process migration function 13 of the migration source “processor number 111” erases the process information 19 and the process 17. However, if the request cannot be accepted for each request to each processor due to some reason, the movement of the process is stopped and the subsequent processing is continued on the processor.

According to this embodiment, the I / O of each parallel process
Considering the switching overhead of the network in the request, it is possible to improve the network use efficiency of each processor part, and the throughput of the entire system is improved.

[0038]

As described above, according to the process scheduling method of the present invention, in a parallel computer system in which each processor has only an independent memory and does not have a shared memory, the network is controlled by the I / O request of each processor. It is possible to prevent an increase in switching overhead, eliminate a local bottleneck of a specific process, and prevent a decrease in execution performance of a parallel process to be executed.

[Brief description of the drawings]

FIG. 1 is a diagram showing an example of the configuration of the entire system of the present invention.

FIG. 2 is a block diagram showing one embodiment of the present invention.

FIG. 3 is a diagram illustrating an example of configuration information of a process management table 9 according to the embodiment.

FIG. 4 is a diagram illustrating an example of configuration information of an I / O load status management table 10 of each dedicated processor 2 of the present embodiment.

FIG. 5 is a diagram illustrating an example of configuration information of a system parameter table 16 of the system management processor 1 according to the present embodiment.

FIG. 6 is a flowchart illustrating an operation of a processor selection process 8 according to the embodiment.

FIG. 7 is a diagram for explaining an operation in which the process moving function 13 of the present embodiment moves a process being executed by its own processor to another processor.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Processor for system management, 2 ... Processor for calculation only, 3 ... Processor for exclusive use of I / O processing, 4 ... 3D network, 5 ... System management,
6: Process management function, 7: Processor number,
8 Processor selection function 9 Process management table 10 I / O load status management table 11
Processor management, 12: Process creation / termination function, 13: Process movement function, 14: I
/ O load monitoring function, 15 ... I / O device,
16: System parameter table, 17: Process,
18 ... I / O device control function, 19 ...
Process information, 20 ... process number, 2
1, 31: Process number, 22: Processor number, 23: Process state, 24: Processor number (X coordinate), 25: Processor number (Y coordinate), 26: Processor number (Z coordinate), 32: I / O
Processing-dedicated processor number, 33 ... total value of I / O request amount,
34 ... I / O processing dedicated processor number (X coordinate), 35 ...
I / O processing dedicated processor number (Y coordinate), 36 ... I / O processing dedicated processor number (Z coordinate), 41 ... I / O amount threshold, 42 ... Processor selection function activation interval.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Kazuya Higuchi 5030 Totsuka-cho, Totsuka-ku, Yokohama-shi, Kanagawa Prefecture Inside Software Development Division, Hitachi, Ltd. (72) Inventor Toshiaki Sunako 5030 Totsukacho, Totsuka-ku, Yokohama-shi, Kanagawa Prefecture Address: Hitachi, Ltd. Software Development Division (72) Inventor Tetsuro Ota 5030 Totsukacho, Totsuka-ku, Yokohama-shi, Kanagawa Prefecture Hitachi, Ltd. Software Development Division (72) Inventor: Tamotsu Sato Onoe, Naka-ku, Yokohama-shi, Kanagawa 6-81-cho, Hitachi Software Engineering Co., Ltd. In-house (72) Inventor Kenichi Takahashi 6-81 Onoe-cho, Naka-ku, Yokohama-shi, Kanagawa Prefecture In-house Hitachi Software Engineering Co., Ltd. (72) Inventor Tateo Minami Totsuka-ku, Totsuka-ku, Yokohama, Kanagawa 5030-cho Hitachi, Ltd. Inside the Office Software Development Division

Claims (1)

[Claims] 1. A process scheduling method in a parallel system, wherein an optimal processor is determined at the time of an I / O request from a process in scheduling a running process in a parallel computer system including a plurality of processors.