JPH117430A - Parallel processing system backup method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To inexpensively construct an inter-CPN parallel processing system which has high efficiency and reliability and to operate it by deciding the number of connection mechanisms(CF) in accordance-with the number of CPNs in an inter-processing node(CPN) parallel processing system, sharing hardware resources and assigning an alternate CF. SOLUTION: When a CF supervisory program detects a failure in a CF1 (4), the supervision of the CF1 (4) is suspended and all of copy data that are managed by an CF0 (3) are nullified. A CF state supervisory CPN1 (18) sucks contents of memory M1 (5) through a link cable 19, writes them on memory 2 (6) and moves all of necessary data that are under the management of the CF0 (3) onto a coupling device 2. Processor 1(7) to n(9) are given a lock right to shared memory M3 (16) and processing in each CPN is resumed. In such times, the connection between the alternate CF1(18) and CPN0 (10) to n-1 (12) shares a link cable 21 that connects the CF1 (4) and the CPN0 (10) to n-1 (12).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a system provided with a plurality of CFs and a plurality of CPNs, which is intended to be operated without causing a system down when one CF fails. System, CF monitoring system, and C
This is related to the F transition method.

[0002]

2. Description of the Related Art Coupling devices have been introduced to realize high-speed parallel processing, which is indispensable for processing today's huge amount of information. However, as the size of the system increases, the CF load on the coupling device increases. Further, if the CF goes down, the system is seriously damaged. Therefore, the load and the risk at the time of failure have been dispersed by multiplexing the coupling devices. However, the conventional failure backup method has the following two problems.

1) The cost is doubled in the complete duplex system.

2) In a system in which functions are distributed and a parallel system is reconstructed when a failure occurs, it takes time to recover the system.

[0005]

According to the present invention, the load on the CFs is reduced by multiplexing the coupling devices and constantly monitoring the CFs on each coupling device by the status monitoring CPN sharing hardware with the CFs. It is intended to provide a means for preventing the system from being down due to a CF failure as well as dispersing the data without increasing the cost.

[0006]

In order to solve the above-mentioned problems, according to the present invention, a plurality of coupling devices, each CF and hardware resources on the coupling device are logically divided into Cs.
Alternating C that operates after transition to FPN monitoring CPN and CF functions
F, a CF monitoring program and a CF migration program that operate on the CF state monitoring CPN are prepared.

Here, when a failure of a CF is found, the contents of the local memory of the failed CF are copied to the local memory of another CF by the program, and the CPN is operated as a replacement CF to cause a failure. In the previous state, the operation can be continued without bringing down the system.

[0008]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention, from CF monitoring to CF transition at the time of a CF failure, will be described in detail with reference to the drawings.

FIG. 1 shows one embodiment of the present invention. In FIG. 1, (1) is a first coupling device,
(2) is a second coupling device. On the coupling device, coupling mechanisms CF0 (3) and CF1 (4) and local memory M are respectively provided.
1 (5) and M2 (6). (7) is the first processing device, and (8) and (9) are the second and n-th processing devices, respectively. Each processing device is provided with a central processing node (hereinafter referred to as CPN) and a local memory. CPN0 (10), CPN1 (11), and CPNn-1 (1
2) Local memory M4 (13), M5 (14), M6 (15). These n CPNs are all CPNs for operating the OS. CF0 (3) is connected to each coupling mechanism via a link cable (20) and CF1 (4) is connected to each coupling mechanism via a link cable (21). . These processing devices 1
(7), the processing device 2 (8) and the processing device n (9) are also connected to the shared memory M3 (16). The processing device 1 (7), the processing device 2 (8), and the processing device n (9) can access the shared memory M3 (16) independently of each other. Since a copy of necessary data is stored in the memory, access to the local memory is sufficient as long as there is no change in the shared memory. Since there are a plurality of copies of data, when any one of the data is rewritten, the other data becomes invalid. Further, data on the shared memory cannot be accessed until a lock is obtained for each data, and copy data cannot be rewritten unless the shared memory is first rewritten. This prevents a plurality of copy data from being rewritten at the same time. The information such as whether the copy data is valid is, for example, as shown in FIG. 4, when the copy data in the local memory M4 (13) is rewritten by the CPN 0 (10) on the processing device 1 (7), (10) notifies the CF0 (3) on the coupling device 1 (1) that the data has been rewritten. Then, all the CPNs other than the CPN (10) by the CF0 (3), that is, CPN1 (11), CPNn- The data is written as invalid in the valid flag table 24 of 1 (12). When using the data, the CPN1 (11) and the CPNn-1 (12) check the validity flag table of each processing unit and confirm that the data is invalid. Data cannot be used unless it is copied to (14) or M6 (15). Thus CF
0 (3) and CF1 (4) control the validity of the copy data together with the lock control of the shared memory, and the control data is stored in the local memories M1 (5) and M2 (6), respectively.

Further, the CF status monitoring CPN0 (17)
0 (3) and a dedicated CP which shares the hardware and monitors the status of CF0 (3) and CF (4).
N, the CF state monitoring CPN1 (18) is a dedicated CPN that shares hardware with CF1 (4) and monitors the states of CF0 (3) and CF1 (4). CF status monitoring CPN0 (1
7) functions as a replacement CF for CF1 (4) when a failure is detected by the CF status monitoring program, and shares the local memory of CF (3). The CF status monitor CPN1 (18) also functions as a replacement CF for CF0 (3) when a failure is detected by the CF status monitor program.
The local memory of (4) is shared. CF status monitoring C
PN0 (17) and CF status monitoring CPN1 (18) are also externally connected to CF0 (3) and CF1 (4).
The status of the CF is regularly checked by a monitoring program.

Here, if a failure is detected in CF1 (3) by the CF monitoring program running on CF status monitoring CPN1 (18), the CF on CF status monitoring CPN1 (18)
The monitoring program executes the CF monitoring on the CF status monitoring CPN0 (17).
Inform the monitoring program that monitoring of CF1 (3) has been interrupted.
Further, the CF monitoring program on the CF status monitoring CPN1 (18) suspends the monitoring of CF1 (3), and starts the CF migration program. The CF migration program first invalidates all copy data managed by CF0 (3) as shown in FIG. During this time, processing device 1 (7), processing device 2
(8) Since the right to lock the shared memory M3 (16) is not given to the processing device n (9), the processing will not be continued using incorrect data. Next, the content of the memory M1 (5) is stored in CF
Via the link cable (19) used for monitoring, CF
The status monitor CPN1 (18) siphons and writes it to the memory M2 (6), thereby transferring all necessary data under the control of the CF0 (3) to the coupling device (2). If this time, the memory M1 (5)
Even if the data cannot be read due to the failure, the same data as the data in the unreadable memory M1 (5) is transferred to the CF by comparing the shared memory M3 (16) with the copy data on each CPN. The program creates
It is stored in the memory M2 (6). However, in this case, some time is required for the hot standby operation after the failure. Thereafter, by starting a CF program on the CF state monitoring CPN1 (18), the CF1 (1) is changed from CF0 (3).
The transition to 8) is completed. Then, the right to lock the shared memory M3 (16) is given to the processing device 1 (7), the processing device 2 (8), and the processing device n (9), and the processing in each CPN is restarted. At this time, the replacement CF1 (18), CPN0 (10), CPN1 (11), C
The connection between PNn-1 (12) is CF1 (4) and CPN0 (1
0), CPN1 (11) and CPNn-1 (12) share a link cable (21). Thereby, CPN0 (1
0), CPN1 (11), and CPNn-1 (12) are maintained in parallel control, so that the operation can be continued in the state before the failure without causing the system to go down.

FIG. 2 shows an embodiment other than FIG. 2 differs from FIG. 1 in the connection method between the coupling devices. The operation in the normal state is as described in FIG. By performing the TCMP coupling (20) between the coupling devices, the hot standby operation during the CF failure can be performed.
Since the memory contents of the other party can be read directly without copying the contents of the memory which requires the most processing time, there is an advantage that the CF function can be switched smoothly.

FIG. 3 shows an embodiment other than FIGS. By connecting all coupling devices, processing devices, and shared memories via a bus (23), CF and CPN are connected.
Only the link cable connecting between them becomes unnecessary, and the normal operation and the operation of the CF transfer program are as described in FIG.

[0014]

As described above, according to the present invention,
In an inter-CPN parallel processing system to which many CPNs are connected, the number of CFs is determined in accordance with the number of CPNs, and hardware resources are shared to assign replacement CFs. High CP
It is possible to construct and operate the N-to-N parallel processing system without increasing the cost.

[Brief description of the drawings]

FIG. 1 shows an example of a system configuration diagram to which the present invention is applied.

FIG. 2 is a diagram showing an example other than FIG. 1 to which the present invention is applied.

FIG. 3 is a diagram showing an example other than FIGS.

FIG. 4 is a diagram illustrating a parallel control method of a CF.

FIG. 5 is a diagram illustrating a CF replacement method according to the present invention.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... 1st coupling apparatus, 2 ... 2nd coupling apparatus, 3 ... 1st coupling mechanism (CF0), 4 ... 2nd coupling mechanism (CF1), 5
... Local memory (M1) of CF0, 6 ... Local memory (M2) of CF1, 7 ... First processing device, 8 ... Second processing device, 9 ... Nth processing device, 10 ... First central processing Node (CPN0), 11: second central processing node
(CPN1), 12 ... n-th central processing node (CPNn-1), 13 ...
CPN0 local memory (M4), 14... CPN1 local memory (M5), 15... CPNn-1 local memory
(M6), 16: Shared memory (M3), 17: First CF status monitor CPN0 (alternate CF0), 18: Second CF status monitor CPN1 (alternate CF1), 19: CF status monitor CP
External link cables connecting N and CF, 20 ... C
Link cable group connecting F0 and CPN0, CPN1, CPNn, 21 ... CF1 and CPN0, CPN1, C
Link cable group connecting between PNn, 22 ... Cable connecting tightly-coupled multiprocessor between coupling devices, 23 ...
A bus connecting each processing device, coupling device, shared memory, 2
4: Valid flag table for copy data

Claims (5)

[Claims] 1. A processing node (CPN) on a processing unit in several parallel processing systems as shown in FIG.
Is coupled to a coupling device.
In the inter-parallel processing system, the multiplexing of the coupling devices improves the parallel processing performance of the system. CPN), the CPN constantly monitors the state of each CF, and when a failure is found in a certain CF, the CPN itself functions as a replacement CF to prevent the entire system from going down. A CF status monitoring type inter-CPN parallel processing system that enables quick switching operation (hot standby). 2. In the CF status monitoring type inter-CPN parallel processing system according to claim 1, the coupling devices sharing hardware resources with the CF and monitoring each CF are connected as shown in FIG. A CF state monitoring CPN connected to a tightly-coupled multiprocessor (TCMP) so that a local memory content other than the subordinate memory can be read when a CF failure is detected, and enabling a high-speed transition of the CF function. 3. A system as shown in FIG. 3, wherein a CPN on each processing unit and a CF on a coupling unit are connected by a bus connection. A state monitoring type parallel processing system between CPNs and a CPN dedicated to CF state monitoring. 4. A CF status monitor on the CF status monitoring CPN according to claim 1, wherein the CF status is detected when a CF fault is detected.
A CF state monitoring program that performs function transfer processing. 5. A CF transition program, wherein the CF state monitoring program according to claim 4 performs transition processing to a replacement CF when a CF failure is detected.