JP2000132432A - Memory dump system for parallel nodes - Google Patents

Abstract

PROBLEM TO BE SOLVED: To decrease the output quantity of a memory dump and a snapshot by recording neither an area whose alteration contents are meaningless or an area whose alteration is obvious by not storing main storage contents irrelevantly to whether or not update is performed. SOLUTION: A dump start indicating means 9 extracts all the main storage contents of a basic node and differences of other nodes from the basic node once a program is loaded and a main storage update detecting means 5 outputs updated pages recorded in an update information hold means 6 as the initial state of the memories of all the nodes when memory dumps are gathered, thereby reducing the memory dump capacity.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a memory dump method for a parallel computer system, and more particularly to a reduction in dump capacity and snapshot capacity when the same program is executed on a plurality of nodes.

[0002]

2. Description of the Related Art The capacity of a main storage device of a computer system has been increased, and the capacity of a memory dump to be collected when a failure has occurred and a snapshot for recording a change history of the main storage device have also been increased. For this reason, it is necessary to increase the capacity of a secondary storage device for storing a memory dump and a snapshot. Conventionally, as described in JP-A-2-85938 or JP-A-7-13805, the amount of memory dump and snapshot output is reduced by recording the changed main memory.

[0003]

The above prior art is directed to one or more central processing units and a single main storage device, and has a main storage device distributed to a plurality of nodes and has the same configuration. No consideration is given to a parallel computer in which the program is executed in parallel, and it is necessary to collect memory dumps and snapshots for the number of nodes.

An object of the present invention is to reduce the amount of memory dump and snapshot output of a parallel computer by utilizing the similarity between nodes.

Another object of the present invention is to provide a means for collecting a record of access to a main storage device by an input / output device which cannot be detected by a central processing unit by storing the contents of a main storage regardless of whether or not the update has been performed. To provide.

Another object of the present invention is to record an area in which the contents of a change have no meaning or an area in which a change is obvious by not storing main memory contents regardless of whether or not the update has been performed. Memory dump,
An object of the present invention is to provide means for reducing the amount of snapshot output.

[0007]

In order to achieve the above object, in the memory dump method of the present invention, a basic node distributes a part of the main storage contents to each node by using broadcast communication, and simultaneously broadcasts the contents. And outputs only the different parts to a secondary storage device such as a disk. At this time, by using a loader that can specify a physical address to be loaded, such as a text or a static data area, the similarity of the memory between the nodes can be improved, and the difference can be reduced.

Further, by recording the update status of the main storage device in each node as a bitmap for each page and outputting the difference at regular intervals, a snapshot with reduced capacity can be realized.

Further, a bit map for recording that the page cannot be detected by the central processing unit, that the main memory is updated by the input / output device, or that the page is always changed but need not be recorded. Is provided, the reproducibility of the snapshot can be improved.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, an embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing an embodiment of the present invention. Nodes 1-1, 1-2, 1-3,... Constitute a parallel computer, and each includes a central processing unit 2, a main storage device 3, a communication unit 4, a main storage update detection unit 5, and a main storage update information storage Means 6, main memory dump necessity information holding means 7, node information holding means 8, and dump processing management means 9.

Each node is connected via a communication path 10 and has a normal one-to-one communication function and a broadcast function for simultaneously transmitting data from one node to each of the other nodes. . Each node does not need to have the secondary storage device, and stores the main storage content by sending data to the node having the secondary storage device via the communication path 10.

Each node is designated in advance as one basic node or a plurality of other nodes, and this information is held in the node information holding means 8. Main memory update detection means 5
Detects the presence / absence of writing for each unit (page) for managing the main storage device, and sets that the main storage update information holding means 6 has been updated when there is writing.

The main memory dump necessity information holding means 7 holds information as to whether or not to output the contents of the main memory corresponding to the dump collection regardless of the main memory update information holding means 6. The update of the main memory content by the input / output device that cannot be detected by the central processing unit can be recorded by setting the input / output area or the like that requires the main memory content.

Each node has a dump processing management means 9,
When the start of dump collection is specified for the system, the dump processing management means 9 of the basic node instructs each node to collect a dump.

FIG. 4 is a flowchart showing the initialization of the dump processing. Each node refers to the node information holding means 8 and determines whether or not the own node is a basic node. (S1)
If the node is a basic node, the main storage contents of the own node are stored in the secondary storage device, and the main storage contents are broadcast to each node for each page. (S2) Nodes other than the basic node receive the main storage contents of the basic node. (S3) The received main storage contents are compared with the corresponding main storage contents of the own node (S4). Different main storage contents are candidates to be output to the secondary storage device.

According to this processing, only the difference between the main storage contents of the basic node and the main storage contents of each node at the designated time can be extracted, and a memory dump with reduced capacity can be provided.

When operating the system after dumping,
When the dump processing management means 9 completes the dump collection processing,
Instruct each node to return to the process that was being performed at the start of dump collection. Each node sets the page information in the main memory update information holding means 6 when the main memory update detecting means 5 detects the subsequent writing to the main memory. When there is a dump collection request after the initial difference collection, the dump processing management means 9 of the basic node instructs each node to store the difference information in the secondary storage device.

FIG. 5 is a flowchart showing an example of the difference information collecting process. Each node refers to the main memory update information holding means 6 to determine whether or not the main memory content has been updated (S8).
The updated main storage content is a candidate to be output as a secondary difference. When the output is completed, each node clears the main memory update information holding means 6 and continues to execute the processing that was being performed when the dump collection was started.

Thus, the contents of the main memory of each node at each designated time can be extracted as an n-th difference from the initial difference, and a snapshot with reduced capacity can be provided.

By specifying an interval as a trigger for repeatedly collecting a dump, a snapshot whose capacity has been reduced can be collected at regular intervals.

At this time, in order to prevent the total capacity of the snapshot from increasing indefinitely, the dump processing management means 9 of the basic node monitors the capacity of the initial difference and all n-order differences, and when the capacity becomes the designated capacity. Can delete the difference by reinitializing the dump processing, so that the dump capacity does not become larger than the limit value.

When the number of pages to be updated is larger than the interval at which snapshots are taken, dump processing management is performed so that dump collection is started when the number of updated main storage pages exceeds a specified value. When the means 9 is set, the dump processing management means 9 of each node detects that the number of updated pages held by the main memory update information holding means 6 has become equal to or larger than the specified number of pages, and performs the dump processing management of the basic node. It can communicate with the means 9 and request the start of dump collection.

The dump processing management means 9 of the basic node instructs each node to start dumping in the same manner as when the interval has been reached, so that after a large number of main memory changes have occurred, a snapshot of the main memory is taken. Is created, even if a system failure occurs before a dump is collected, it is possible to minimize the loss of calculation results.

FIG. 2 shows a program loading state in each node. By loading the text 11 and the static data 12 in the same area at each node, the difference is reduced, and the dump output capacity can be reduced.

FIG. 3 shows a case where the main memory dump necessity information holding means 7 is set. Like the input / output area 14,
If writing by an input / output device that cannot be detected by the central processing unit cannot be detected by the main memory update detection unit 5,
This area is output irrespective of the information in the main memory update information holding means 6 (YES in S6 in FIG. 5).

As in the timer area 15, the central processing unit always updates the data by the main memory update detecting means 5. However, if the area does not need to be dumped, the information is stored in the main memory update information holding means 6. Output is suppressed (YES side of S5 in FIG. 5).

This makes it possible to reliably collect a memory dump in an area where an update cannot be detected without adding a device, and it is possible to reduce a dump capacity by not collecting unnecessary data. Becomes

[0028]

According to the present invention, it is possible to reduce the capacity required for taking a memory dump and a snapshot in a parallel computer having a large number of nodes. And the like, and the data collection time is shortened, so that the availability of the system can be improved.

Further, it is possible to record a change in the main memory by the input / output device, which cannot be detected by the central processing unit, without adding a device, thereby realizing a snapshot with high reproducibility.

[Brief description of the drawings]

FIG. 1 is a block diagram of a parallel computer according to an embodiment of the present invention.

FIG. 2 is a view for explaining means for designating a loading position of a program so as to be the same on each node.

FIG. 3 is a diagram illustrating a unit that specifies dump collection regardless of update information.

FIG. 4 is a flowchart for collecting an initial difference.

FIG. 5 is a flowchart for collecting an n-th difference.

[Explanation of symbols]

1-1, 1-2, 1-3 ... nodes, 2 ... central processing unit, 3 ... main memory content comparison means, 4 ... communication means, 5
... Main memory update detecting means, 6 ... Update information holding means, 7
... Dump necessity information holding means, 8 ... Node information holding means,
9: dump start instructing means, 10: communication path, 11: program text, 12: program static data, 13: program dynamic data, 14: area where the input / output device writes, 15: dump Area that does not need to be collected, 16: Main storage device, 17
... Main memory contents storage means, 18 ... Secondary storage device, 19 ... Loading position designation means.

Claims (6)

[Claims] 1. A means for comparing main memory contents between nodes in a parallel computer having at least one central processing unit and a plurality of nodes each comprising a main memory, and a unit for managing the main memory of each node A means for knowing whether or not there is an update for each node, a means for recording, and a means for storing specific main storage contents of each node in a secondary storage device. A) memory dump method for parallel nodes, wherein the contents of all main memories and the difference (initial difference) between the basic node and the main memory of each node are stored in a secondary storage device. 2. The main storage content storage means according to claim 1, wherein the main storage content (secondary difference) of each node updated after storing the initial difference and the main storage content updated after storing the secondary difference are stored. A memory dump method for a parallel node, characterized in that a plurality of differences are stored as a tertiary difference and a main storage content updated after storing the tertiary difference as a quaternary difference to record main storage update histories of all nodes. 3. The dump system according to claim 1, wherein the basic node broadcasts the contents of the main storage to each node,
A memory dump method for parallel nodes, wherein each node compares the contents of its own main storage and the contents of its broadcast main storage all at once. 4. The parallel node memory according to claim 1, further comprising means for designating that a main memory content is always stored in a specific area irrespective of whether or not the data is updated. Dump method. 5. The parallel node memory according to claim 1, further comprising means for designating that the contents of the main memory are not always stored in a specific area regardless of whether or not the data is updated. Dump method. 6. A memory dump method for a parallel node according to claim 1, further comprising means for designating a position on a main storage device when loading an execution program.