PatentDe  


Dokumentenidentifikation EP0750256 02.10.2003
EP-Veröffentlichungsnummer 0750256
Titel Struktur zur Gruppenzugehörigkeitsverwaltung in einem Mehrfachrechnersystem
Anmelder Data General Corp., Westboro, Mass., US
Erfinder Alfieri, Robert A., Menlo Park, California 94025, US;
Compton, James T., Durham, North Carolina 27705, US;
Huber, Andrew R., Monroeville, Pennsylvania, US;
McGrath, Paul T., Raleigh, North Carolina 27613, US;
Soufi, Khaled S., Durham, North Carolina 27713, US;
Thorstad, Brian J., Cary, North Carolina 27511, US;
Vook, Eric R., Chapel Hill, North Carolina 27514, US
Vertreter derzeit kein Vertreter bestellt
DE-Aktenzeichen 69629630
Vertragsstaaten DE, FR, GB
Sprache des Dokument EN
EP-Anmeldetag 20.06.1996
EP-Aktenzeichen 963045992
EP-Offenlegungsdatum 27.12.1996
EP date of grant 27.08.2003
Veröffentlichungstag im Patentblatt 02.10.2003
IPC-Hauptklasse G06F 9/46
IPC-Nebenklasse G06F 15/16   

Beschreibung[en]

This invention relates generally to multiprocessor systems and, more particularly, to shared-disk cluster systems. More particularly, the invention relates to a framework for joining and disjoining nodes in a multiprocessor cluster system.

A multiprocessor cluster system typically includes multiple nodes, which are interconnected with a private communication interconnect. The cluster system additionally includes a shared cluster resource, such as a virtual hard disk, which is accessible to all of the nodes, which run an operating system supporting coordinated access to the shared resource. Cluster systems have many advantages. They provide high availability to the user because availability does not depend upon all of the nodes being active participants in the cluster. One or more nodes may leave the cluster without necessarily affecting availability. New nodes may be added to the system without requiring that the system be taken down and rebooted. Additionally, nodes may incorporate processor designs that are different from one another, which facilitate expansion of the system. In this manner, the cluster system provides high aggregate performance.

Shared-disk cluster systems have typically been used for database services which require a distributed lock system in order to avoid contamination of data on the shared virtual disk. Membership management in such a cluster system required providing cluster awareness to the distributed lock system. However, such shared-disk cluster systems have been limited because cluster awareness extends to only one layer of subsystem. Particular operating systems have multiple subsystems which are layered in a manner that a higher level subsystem must depend upon the operation of lower level subsystems. Known cluster membership management techniques are not capable of taking such layered subsystems through cluster transitions of nodes joining and leaving the cluster.

Client services are typically distributed among the nodes of the cluster requiring extensive coordination of which node implements which service. This is especially difficult during node transitions of a node joining or leaving the cluster. This is because most services are not aware of the cluster environment. The client services would typically determine on their own the best node to execute on. A recovery mechanism would be required for initiating recovery if the node currently executing the service leaves the cluster. Allowing individual services to implement their own mechanism for this coordination requires detailed modifications to the client services to allow them to run on a cluster system which makes administration of the cluster more burdensome and difficult because inconsistent mechanisms may be used.

United States Patent US-A-5,287,453 discloses a cluster computer system which includes a cluster of interconnected independently-operated computer systems which exchange information through a cluster controller. The members of the cluster exchange a list of all the disks, tapes and printers to be shared cluster-wide. When an operator wants to reference a named disk volume in the system, they mount the disk pack or volume on the disk drive of one of the members of the cluster. The system recognises the volume locally and sends a message to all other members that the named volume has been mounted.

The invention in its various aspects is defined in the independent claims below, to which reference should now be made. Advantageous features are set forth in the appendant claims.

A preferred embodiment of the invention, described in more detail below with reference to the drawings, provides a method and apparatus for combining particular processors, or nodes, of a multiprocessor system in a cluster that appears substantially as a unified processor to users of the system. Multiple subsystems running on nodes presently in the cluster are notified of transitions of nodes joining and leaving the cluster. This provides a consistent view of active membership in the cluster to the subsystems of the cluster nodes whereby all of the node's subsystems may be taken through the node transitions. This feature is particularly useful with subsystems that are interdependent in levels, with higher level subsystems depending on the operation of lower level subsystems. A particular transition is noticed to the same level subsystem on all nodes. Notification will not proceed to another subsystem level until the noticed subsystem of each node processes that notification and acknowledges that such processing has been completed. When the transition is a node joining the cluster, subsystems are notified beginning with lower level subsystems and proceed in sequence through higher levels of subsystems. When the transition is a node gracefully leaving the cluster, subsystems are notified beginning with higher level subsystems and proceeding in sequence through lower level subsystems. When the transition is a node being ungracefully forced from the cluster by other nodes, subsystems are notified beginning with lower level subsystems and proceeding in sequence through higher level subsystems.

A registration and launch function is provided in which client services, including user applications, are initiated on particular nodes in a cluster in a manner that the cluster appears substantially as a uniform unit to the client services. A node is chosen for each client service and that client service is registered with the node. Nodes presently in the cluster are notified that the particular service is registered with the particular node. In this manner, client services can be transferred to another node if the node on which that service is registered leaves the cluster. The client service may be launched on a node, according to an action parameter included with the service, in response to registering that service with that node. This provides cluster-wide availability to client services because they will not need to explicitly initiate themselves each time they are transferred. Client services may, advantageously, be grouped as a parent service and one or more child services. Grouped client services are registered with the same node and utilize action parameters included with the parent service for all launching activity within the group. The choosing of a node for each client service may include providing a database of choosing factors for the client service and applying the choosing factors to information regarding the availability of the nodes in the cluster. The choosing factors establish rules relating nodes to the client service.

Such registration and launch function is preferably a component of an event manager, which is a subsystem which receives notification of node transitions from the cluster membership manager. The event manager monitors client services registered with a particular node using an event watcher and provides action items which are carried out in response to occurrence of an event, such as a node transition. The event watcher may be enabled in response to registering of a client service and disabled in response to deregistering of the client service.

The preferred embodiment of the invention will now be described in more detail, by way of example, with reference to the drawings, in which:

  • Fig. 1 is a block diagram of a multiprocessor cluster system embodying the invention;
  • Fig. 2 is a state transition diagram of a transition notification framework for one subsystem level;
  • Figs. 3-10 are diagrams of states of subsystems in a two-node cluster illustrating nodes joining the cluster;
  • Figs. 11 and 12 are diagrams of states of subsystems in a two-node cluster illustrating a graceful leave of a node from the cluster;
  • Figs. 13-15 are diagrams of states of subsystems in a two-node cluster illustrating an ungraceful forced leave of a node from the cluster;
  • Fig. 16 is a block diagram illustrating the grouping of client services;
  • Fig. 17 is a diagram similar to Fig. 15 illustrating multiple generations of client service groupings;
  • Fig. 18 is a state transition diagram illustrating the launching of a client service;
  • Fig. 19 is a state transition diagram illustrating the transition states of a client service;
  • Fig. 20 is similar to Fig. 18 illustrating additional transition states; and
  • Fig. 21 is a block diagram of an event manager subsystem.

HARDWARE

Referring now specifically to the drawings, and the illustrative embodiments depicted therein, a multiprocessor cluster system 25 includes multiple nodes 26 and a shared-cluster resource, such as a physical disk 28, which could be made up of multiple physical disk drives (Fig. 1). Each node 26 includes a processor (CPU), physical memory, caches, shared and private bus interfaces, and optional dedicated devices. Each node runs a copy of a UNIX-based operating system, such as DG/UX 5.4 operating system marketed by Data General Corporation of Westboro, Massachusetts, running on any hardware configuration which supports such operation system. An example of such hardware configuration is the AViiON® family marketed by Data General.

Cluster system 25 additionally includes an interconnect 36, which is a dedicated shared-cluster communication media that allows nodes 26 to talk directly to all other nodes in the same cluster, and a shared-cluster I/O bus 32, which allows all nodes to share all devices physically connected to the shared bus, such as disk 28. In the illustrated embodiment, shared bus 32 is a SCSI standard bus.

SOFTWARE

Cluster system 25 includes a single membership database 34, which occupies a dedicated shared-cluster virtual disk, which lives on physical disk 28 along with a cluster-cognizant bootstrap 38. Membership database 34 manages persistent node configuration information 40 that is needed to boot, shutdown, or panic a node 26. Such persistent information includes identification of the number of nodes configured with the system, as well as configuration information about each node, indexed by a node identification number. Membership database 34 additionally includes an active membership state database 42, which contains transient information about node states. Such transient information changes dynamically as nodes join the cluster gracefully, gracefully leave the cluster, or are ungracefully forced out of the cluster. A node can have any one of the following states:

  • Inactive - The node is not configured or is not an active member of the cluster.
  • Joining - The node is in the process of joining the cluster, which implies that the node has informed other nodes of its intention to join the cluster gracefully, but not all of the registered subsystems of nodes in the cluster have completed transitions to gracefully include the new node.
  • Joined - The node has fully joined the cluster and all registered subsystems of nodes in the cluster accept the new node as a member of the cluster and have completed their transitions to include the new node.
  • Leaving - The node is in process of leaving the cluster, which implies that the node has informed other nodes of its intention to leave the cluster gracefully, but not all nodes' registered subsystems of nodes in the cluster have completed transitions to gracefully exclude the new node.
  • Forced-Leaving - Other nodes are in the process of forcing this node out of the cluster. Other nodes may force out a node if that node is not functioning properly, such as failing to communicate with other nodes. After the other nodes have completed processing of the forced-leave, which includes running recovery procedures, the other nodes mark this node as inactive. The forced-out node will panic after it has noticed that the other nodes have forced it out. A node panics by halting further processing in order to avoid corrupting shared cluster resources.

A. MEMBERSHIP MANAGER

Cluster system 25 includes a membership manager framework including a transition notification framework 44, which provides notifications to all kernel-level and user-level subsystems that must receive notifications of cluster transitions (Figs. 2-15). The purpose of transition notification framework 44 is to provide cluster-cognizant subsystems, a coherent technique for processing cluster transition information among the nodes. Cluster-cognizant subsystems are subsystems which are registered with a node's cluster membership manager subsystem 46. In the illustrated embodiment, each node 26 includes four kernel-level subsystems, including cluster membership manager subsystem 46, which collectively provide transition notification framework 44, a distributed lock manager (DLM) subsystem 48, a virtual disk manager (VDM) subsystem 50, and a shared file system (SFS) subsystem 52. Each node 26 additionally includes at least one user-level subsystem; namely, an event manager subsystem 54. Such subsystems 46-54 are interdependent upon each other, in levels. In the illustrated embodiment, membership manager 46 is the lowest level subsystem and event manager 54 is the highest level subsystem. However, other higher level subsystems could be provided. A global transition ordering is provided for the subsystems, with lower level subsystems receiving smaller values and higher level subsystems receiving larger values.

Transition notification framework 44 operates as follows. Before a node joins a cluster, interested subsystems of that node register their intention to receive notifications of cluster transitions. A registered subsystem must also supply a thread of control that blocks waiting for transition notifications from membership manager subsystem 46 of that node. During graceful joins and forced leaves of nodes, all nodes' membership manager subsystems 46 coordinate to notify the node's registered subsystems in a bottom-up fashion with respect to the global transition-ordering scheme, as will be illustrated in more detail below. Thus, the membership manager subsystems notify, first, all of the node's subsystems with the lowest order followed by the next highest order, on up to the highest order. Conversely, during graceful leaves of nodes, all nodes' membership manager subsystems 46 coordinate to notify the node's registered subsystems in a top-down fashion, notifying first all nodes' subsystems with the highest order, followed by the next highest order, down to the lowest order. This ordering is so that higher level subsystems' dependencies on lower level subsystems are satisfied. That is, a lower level subsystem first processes a node join transition so that higher level subsystems can be ensured that the subsystems they depend upon, namely, lower level subsystems, are aware of and have completed processing of the join. Conversely, a higher level subsystem must first process a graceful leave so that the lower level subsystems remain operational in the leaving node during the leave transition. An ungraceful leave is processed from the bottom-up to ensure that all error conditions are propagated upward before attempting recovery at the next highest level.

Each node's membership manager subsystem will not proceed with notification to the next-in-line subsystem until each node's currently-in-line subsystem acknowledges its completion of processing for the transition. However, each node's membership manager subsystem may notify a registered subsystem to process multiple transitions for different nodes at the same time. Each of these transitions may be of a different type. This improves performance in situations where many nodes are undergoing transitions contemporaneously, such as when many nodes boot after a power failure that has powered down the entire cluster. However, each node's membership manager will not notify subsystems out-of-order for a particular transitional node. As a result, multiple transitions for different nodes may be processed at different subsystem levels at the same time, but the cluster membership manager framework will ensure proper subsystem ordering for each transitional node.

For examples of use of cluster subsystems to participate in graceful joins, graceful leaves, and ungraceful forced leaves utilizing transition notification framework 44, reference is made to Figs. 2-15, which illustrate a cluster system having potentially a two-node cluster. The examples illustrated in Figs. 3-15 may be generalized to three or more nodes with each node transition sequencing through each subsystem, one at a time, across all nodes. Each membership manager would not propagate the same node transition to the next highest subsystem until all nodes at the current level have acknowledged their completion of transition processing for the new node.

In the state illustrated in Fig. 3, the cluster contains no active members. The system administrator powers node N0 and begins its boot. Node N1 is left powered down. In order to gracefully join the cluster as the first active member, node N0 opens the cluster membership database 34 and retrieves its configuration information. Node N0 initializes its kernel subsystems 46-52, which each register themselves for transition notification (Fig. 4). The subsystems spawn a thread that makes a kernel call which blocks because no cluster transitions have occurred at this time. Node N0's INIT subsystem (not shown) initiates node N0's graceful join through the highest currently registered subsystem. Membership manager subsystem 46 of node N0 forms the cluster and marks node N0's active state as joining. Membership manager 46 of node N0 notifies its DLM subsystem 48 that node N0 is joining the cluster (Fig. 5). The thread of DLM subsystem 48 of node N0 is awakened, notices node N0's new joining state, hands the joined processing off to a different DLM thread, then completes its join. Node N0's DLM subsystem marks node N0's state as joined and informs membership manager 46 of node NO that DLM has completed its joined processing for node N0. The same process is repeated for VDM subsystem 50 and SFS subsystem 52 of node N0 (Fig. 6).

After having joined the cluster at the kernel level, node N0 proceeds to user space (Figs. 7a-7c). Node N0's INIT subsystem (not shown) spawns event manager subsystem 54, which spawns a thread which returns immediately because the event manager 54 of node N0 has not yet processed node N0's graceful join. After having processed node N0's graceful join, node N0's event manager 54 marks node N0's state as joined and informs node N0's membership manager 46 that it has processed node N0's graceful join.

In Fig. 8, the administrator powers and boots node N1 which causes nodes N0 and N1 to perform a graceful join of node N1. Node N1 opens the membership database 34, retrieves it's configuration information and initializes its kernel subsystems 46-52, which register for transition notifications. The membership manager of joining nodes must negotiate with the cluster master node in order to join the cluster. When there are multiple nodes in the cluster, one node becomes the master node utilizing Decker's algorithm, which is known in the art The master node writes its heartbeat in a particular area of membership database 34. Joining nodes will examine such area for the heartbeat in order to identify the master node. Membership manager subsystem 46 of node N1 negotiates with the membership manager of node N0, which must be the master node because it is the only node in the cluster, in order to join the cluster. The membership managers of nodes N0 and N1 mark the state of node N1 as joining the cluster. The membership managers of nodes N0 and N1 notify their respective DLM subsystems 48 that node N1 is joining the cluster. Both DLM subsystems wake up from their calls to begin processing node N1's graceful join. After both DLM subsystems have coordinated in processing node N1's graceful join, the DLM subsystems mark node N1's state as joined and acknowledge to the membership manager. After having received both DLMs' acknowledgements, the membership managers of nodes N0 and N1 notify the respective VDM subsystems 50 that node N1 is joining the cluster (Fig. 9). After both VDM subsystems have processed node N1's join, both subsystems mark node N1's state as joined and acknowledge the same to their respective membership managers.

After having joined the cluster at the kernel level, node N1 proceeds to user space with its INIT subsystem (not shown) spawning event manager 54. Node N1's event manager registers itself with the membership managers. Node N1's event manager spawns a thread that makes a kernel call which is returned immediately because node N1's event manager must process node N1's graceful join. Node N0's event manager wakes up to process node N1's graceful join. After having coordinated to process node N1's graceful join, both event managers 54 mark node N1's state as joined and acknowledge the graceful join to their respective membership managers. Node N1 is joined as illustrated in Fig. 10.

A node may initiate a graceful leave while the node is still in the joining state. However, a joining subsystem will not convert the joining states directly to a leaving or an inactive state. The joining subsystem must complete and acknowledge the joined transition. The membership manager will only reverse the joining state to the leaving state between notifications to registered subsystem levels.

A processing, by transition notification framework 44, of a graceful leave of a node, such as what would occur during a shutdown of a node, is illustrated by reference to Figs. 11 and 12. Node N0 initiates its shutdown by making the appropriate call to initiate a graceful leave. The membership manager subsystems of nodes N0 and N1 mark node N0 as leaving. The membership managers of nodes N0 and N1 wake up both event managers 54 with node N0's transition. Both event managers note node N0's state as leaving and begin their coordinated processing of node N0's graceful leave. As will be described in more detail below, the processing of node N0's graceful leave by both event managers may involve a considerable amount of application level shutdown, after which both event managers mark node N0 as inactive and notify their respective membership managers. Node N0's membership manager automatically de-registers node N0's event manager for transition notification whereby node N0's event manager will receive no further notifications. Next, the membership manager of nodes N0 and N1 perform the same iteration with the SFS subsystems 52 of both nodes, then with both VDM subsystems 50, and then DLM subsystems 48. Finally, the membership manager of both nodes mark node N0 as inactive, which also is the end of node N0's graceful leave. Node N0 performs kernel level shutdown processing and returns to the boot command line.

It should be noted that a node may not initiate a graceful join while other nodes are processing the node's graceful leave. In practice, this situation can occur when the leaving node has died abnormally and re-boots before other nodes have had a chance to notice that the leaving node has died. As soon as the other nodes notice that the leaving node has actually died, the other nodes will force the dead node out of the cluster, aborting their graceful leave processing. The other nodes will subsequently accept the new node's graceful join request.

An ungraceful, forced leave is an abnormal situation; for example, when a node is no longer capable of communicating with the other cluster nodes. Once the forced out node notices that the other nodes have forced that node out of the cluster, the forced-out node panics. Transition notification framework 44 ensures that the forced out node does not corrupt any shared-cluster resources. When a registered subsystem is in the middle of processing a graceful join or leave of the forced-out node, each node's membership manager could re-notify the processing subsystem to abort its graceful processing for the node and begin recovering processing.

An example of a forced leave is illustrated with respect to Figs. 13-15 which begins with node N1 being joined to the cluster and node N0 booting and joining the cluster gracefully. If, by way of example, the membership managers of nodes N0 and N1 have processed node N0's graceful join up through the VDM subsystems 50, but have not received an acknowledge from the SFS subsystems 52 of their completion of node N0's graceful join processing because the SFS subsystems of nodes N0 and N1 cannot communicate due to an interconnect failure on node N0, node N0's graceful join is noted as in a hung state (Fig. 13). The membership manager of node N1 notices that node N1 can no longer communicate with node N0. Node N1 forces node N0 out of the cluster by marking node N0's state as forced-leaving. The membership manager of node N0 notices that node N1 has forced out node N0 and panics immediately (Fig. 14). The membership manager of node N1 initiates forced-leave processing starting from the lowest level subsystem and proceeding up to the highest registered subsystem. The DLM subsystem 48 of node N1 marks node N0 as forced-leaving, notices the abrupt transition from joined and begins recovery processing, as illustrated in Fig. 14. After the DLM subsystem of node N1 acknowledges its completion of recovery processing for the forced-leave of node N0 by marking node N0's state as inactive, the membership manager of node N1 performs the same iteration with respect to VDM subsystem 50.

Node N0 may not re-join the cluster gracefully until all of node N1's subsystems have completed their processing of node N0's forced leave. When the membership manager of node N1 has finally caught up with SFS subsystem 52 of node N1, this subsystem will abort its processing of node N0's original graceful join and will perform recovery processing followed by an acknowledgement of its completion of the forced leave processing for node N0 by marking node N0's state as inactive. The membership manager of node N1 would normally continue iterating the forced-leave notification through the highest registered subsystem, event manager 54. However, because SFS subsystem 52 was the highest subsystem to be notified of node N1's graceful join attempt, forced-leave processing will progress only through the SFS subsystem. After processing the forced-leave notification through the highest appropriate subsystem, in this case the SFS subsystem, the membership manager of node N1 marks the state of node N0 as inactive (Fig. 15).

Table 1 illustrates, for a given transition node, the types of notifications that the membership manager will send to a registered subsystem and the corresponding acknowledgements that the membership manager expects to receive of the registered subsystem after the registered subsystem has completed its processing of the transition. Table 1 also lists the re-notifications that the membership manager may send to the registered subsystem while the registered subsystem is still processing the original notification for the transitional node.

Some registered subsystems may need to perform a two-or-more-phase commitment operation for one of the particular node transitions. In order to provide such multiple phase commitment, membership manager 46 provides barrier synchronization. Each registered subsystem may specify a number of barriers the subsystem wants for each type of node transition. The membership managers then provide notifications that are barrier-synchronized with subsystem levels. All nodes at a given subsystem level must acknowledge its completion of the transition processing for the particular barrier before the membership manager will proceed to the next-in-order subsystem. For example, if the DLM subsystem asks for two "joining" barriers during a joining transition, all DLM subsystems must acknowledge joining barrier 0 before they will be notified of barrier 1. After they acknowledge barrier 1, the joining transition will propagate to the VDM subsystem, which may have a different number of barriers. Also, all subsystems at a particular level must register with the same number of barriers for each type of transition.

A state-diagram for transition notification network 44 is illustrated in Fig. 2 to illustrate the types of notifications that an individual registered subsystem may receive for a node transition. As may be observed in Fig. 2, transitioning may proceed through multiple barriers for each transition type. For clarity, only one barrier is illustrated for forced-leaving. However, multiple barriers are allowed. Dashed lines represent transition notifications. Solid lines represent acknowledgements from the individual subsystem on a single node or across all nodes. MEMBERSHIP MANAGER NOTIFICATIONS OF TRANSITIONS FOR A PARTICULAR NODE TO A REGISTERED SUBSYSTEM AND EXPECTED ACKNOWLEDGEMENTS FROM THE REGISTERED SUBSYSTEM Sent Notification

(from MM to subsystem)
New State After All Nodes Have Acknowledged Allowed Re-Notifications (from MM to subsystem) Notes
Joining Joined Forced-Leaving Straightforward, but this subsystem must keep a lookout for the Forced -Leaving re-notification and abort its graceful join in a timely fashion. Leaving Inactive Forced-Leaving Straightforward, but this subsystem must keep a lookout for the Forced-Leaving re-notification and abort its graceful leave in a timely fashion. Forced-Leaving Inactive None Causes this subsystem to abort any graceful join or leave processing for the node. Inactive Inactive Joining MM should never send only an Inactive notification to this subsystem. MM sends this state along with real transition notifications for other nodes. Joined Joined Forced-Leaving MM should never send only a Joined notification to this subsystem. MM sends this state along with real transition notifications for other nodes.

B. EVENT MANAGER

Event manager subsystem 54 is a user space subsystem which provides cluster-wide availability to client services. This latter function is performed by a registration and launch service 56 (Figs. 16-21) which is a component of event manager subsystem 54 (Fig. 21). Event manager subsystem 54 includes an event manager daemon 58 having multiple watchers 60a-60g which monitor for particular conditions. If a watcher detects a problem, the event manager subsystem 54 will resolve the problem via action functions 62. Registration and launch service 56 may be considered a watcher of event manager daemon 58, but performs additional useful functions as will be explained in more detail below.

A client service is any computing activity, including user applications, which is performed on one node or on more than one node in a cluster. One difficulty is determining which node or nodes should initially provide each client service. Additionally, there must be coordination of which node implements which services during a failure scenario. If individual client services were to implement their own mechanisms for determining the best location to execute, and for initiating recovery if the node currently providing the services leaves the cluster, a heavy burden would be placed upon the administrative management of the cluster. Registration and launch service 56 provides cluster awareness to non-cluster-aware applications by choosing which node a client service will execute on, registering the client service with that node, and notifying nodes presently in the cluster that the particular client service is registered with that node. Registration and launch service 56 additionally will provide an optional launching, or execution, capability, which is invoked when the service is registered at a particular node. The launching capability can additionally be used to transfer a service from one node to another in a controlled fashion. The launching capability, referred to as "action functions," enhances cluster-wide availability by providing the ability to initiate, migrate, and terminate client services. Such transfers of the client service may result, for example, from nodes joining and leaving the cluster. Examples of services which may make principle use of the registration and launch service include, by way of example, printing subsystems, floating internet protocol (IP), networking services, and license server, although other services may advantageously make use of the registration and launch service.

1. Registration

Registration is the process where one node, or more than one node, in a cluster can claim, or is (are) assigned, responsibility for a previously defined client service. It is performed on a cluster-wide atomic basis. A registration indicates a claim of responsibility that the registered node is fulfilling the obligations of the specified client service. As a consequence of registration, or de-registration, optional service start-up, notification, and shut-down commands, known as "action functions", will be invoked. In this manner, registration may initiate, or launch, the client service and provide cluster awareness to the user service. Cluster awareness is a result of notification, upon completion of registration, to other nodes in the cluster, as well as nodes subsequently joining the cluster, of the registration.

2. Choosing

The choice of which node is assigned a particular responsibility, as part of a registration operation, is guided by a set of choosing parameters, or database items, 64. These choosing parameters may include a set of database items which specify when, where, and under what conditions a service should be registered. However, additional criteria may be included in the choosing function including recent performance statistics of particular nodes. Administrator-supplied priority factors may be selected as follows:

  • Allowable_Nodes - The nodes from the cluster where registration is allowed. All nodes must be potential members of the cluster although they need not be powered up. A single wildcard character may be utilized to designate all potential nodes of the cluster.
  • Node Preferences - Node preferences result from the fact that not all nodes will support all client services equally well. Node preferences may be specified as an unordered list or as an ordered list. Selection among unordered members will be influenced by recent performance characteristics of the cluster. Ordered lists are processed beginning with the highest rank member.
  • Disallowable_Nodes - The nodes from the cluster where registration is not allowed. Adding a node to a client service's disallowable_node's field does not automatically initiate a transfer of the service.
  • Auto_Register - This is used when the cluster is first powered up, wherein each user client service potentially needs to be registered and started. The auto-register field allows the administrator to define under what conditions a service should be registered.
  • Placement_Policy - This indicates what type of registering philosophy is in place; namely, whether the client service is to be registered on exactly one node or is to be registered and started on every allowable_node.

As illustrated in Fig. 18, a client service may be started from either an auto-start 66 or an external start 68. The registration and launch service of a decision-making node selects the best node at 64 utilizing choosing parameters as previously set forth. The decision-making node can be any node in the cluster. It is determined by possession of a file lock and is processed using a cluster-wide semaphore. If maturity applies to the client service, the registration and launch service transfers to a state 70 awaiting maturity. Maturity refers to the maturity of the cluster. A cluster is mature with respect to a given client service if at least one node from the allowable_node's list is up and (a) the primary node is available, (b) enough nodes are up, or (c) enough time has gone by. Once the cluster is mature (70), or if maturity does not apply, the registration and launch service notifies (72) the selected node to start the service and the other nodes that the client service has been assigned to a particular node. The client service is then started or launched (74). Each registration and launch client service has independent choosing factors, except as described below (grouping).

3. Grouping

Registration and launch service 56 allows client services to be associated with each other in an association known as a grouping 76 (Figs. 16 and 17). The grouping mechanism is a relationship between a parent client service 78 and one or more child client services 80. The purpose of this grouping arrangement is to allow the administrator to specify associations where specific services must be placed together. The child will be placed wherever the parent is placed. Children services do not have any choosing factors; only the parents' choosing factors are used. A grouping 76' may include child services 80 that are children of another service 80', which is, in turn, the child of parent service 78, as illustrated in Fig. 17, all of the children (80 and 80') would be under the placement of parent service 78.

The registration for each child 80 is pended until its parent is successfully registered and its start_command, if any, has successfully completed. After the parent completes, each child is processed. Similarly, a de-registration of a parent implies the de-registration of the children. Children are de-registered first with stop_commands invoked as appropriate. When transferring a grouping 76, grouped children 80 are always stopped first and started last.

4. Action Functions

Registration and launch service 56 supports a variety of actions to take place as a consequence of registration transactions. Actions are the "launch" aspect of the registration and launch service. When combined with groupings 76 and choosing parameters 64, actions provide that many client services can depend entirely on the registration and launch service for cluster-wide availability. Initiation, migration, and termination may all be carried out directly with registration and launch service 56.

Registration and launch service 56 in the illustrated embodiment includes four action functions:

  • Start_Command - The database is checked to determine if such command is associated with the service upon successful registration of the client service. If such command is present, the client service is executed on the registered node. The registration is not complete until the start has completed successfully. If a start operation fails, an attempt is made to start the client service on another node in the allowable_node list.
  • Stop_Command - The database is checked to determine the presence of this command when a client service intends to terminate. The de-registration is not complete until the stop_command terminates.
  • Notify_Command - This command provides a mechanism whereby other nodes are informed that the client service has been assigned to a particular node. When a client service is successfully registered, the database is checked to determine if this command is associated with the service. If it is, it is executed on all nodes in the allowable-node list except the registered node. If a node in the allowable_node's list joins the cluster after a service is registered, and the service has a notify_command, the command is initiated on the new node. This includes nodes which leave and subsequently rejoin the cluster. If there is a start_command, the notification is pended until successful start.
  • Recovery_Command - This is used when a node ungracefully leaves the cluster. For each service registered on the forced out node, the database is checked to determine if there is a recovery_command associated with the service. If there is, it is executed. The node for the recovery operation is determined using the choosing parameters 64. When the recovery completes, the service is de-registered. Typically, the service will then be registered and started on one of the surviving nodes.

A concept closely related to an action function is that of transfer. Transfer of a service is accomplished through a combination of two action functions. First, the service is located and de-registered. Second, it is registered and started. A service transfer may be very helpful under various circumstances. In one circumstance, the administrator may wish to move a service. In another circumstance, the transfer function is used to transfer all of the services for a node that is being gracefully shut down. In another circumstance, the service, by its nature, may be trivial to move. Because there is no impact to moving it, such service may be automatically transferred when a preferred node joins the cluster if the service is placed on a node other than a preferred node.

When a client service is in the process of being transferred from one node to another, a "transfer intent" flag is set by the transferring node. Effect of the transfer intent flag on the transitions of registration and launch service 56 may be seen by reference to Fig. 19. Registration and launch service 56 includes a starting state 82, a registered state 84, a stopping state 86, a de-registered state 88, and a recovering state 90. Each of the stopping, starting, and recovering states will be skipped if their respective command does not exist. The starting state 82 indicates that the service is in the process of starting. If start is successful, the service goes to a registered state 84, which indicates that the service claims to be operational on some node. The registered service transfers to stopping state 86 if the transfer intent flag is set as part of transferring to another node in a graceful leave of the operational node. The registered service transfers to stopping state 86 as part of an external de-register or a transfer operation. Upon completion of the stopping command, the service transitions from the stopping state 86 to the de-registered state 88, indicating that the service is currently not registered and that no node transitions are currently underway. A service being transferred normally proceeds immediately from de-registered state 88 to starting state 82. Recovering state 90 only occurs if a node ungracefully leaves the cluster while the service was in the registered state 84 or in the stopping state 86. A more detailed state transition diagram is illustrated in Fig. 20, illustrating various intermediate states.

Table 2 illustrates an example of state transition in a three-node cluster having nodes N0, N1, and N2. The example is based upon the allowable_nodes being node N0 and node N1, with a maturity_count equal to 2 and a maturity_time equal to 5. Auto_register is set to auto. The example applies to a single client service. At the beginning of the example, all nodes are down. At time t1, node N2 boots so that the status of the node changes to an up condition. No change occurs for the service because node N2 is not in the allowable_node's list. At time t2, the service is awaiting cluster maturity and node N0 boots. At time t3, the maturity_time lapses and the service start_command is executed on node N0. Start completes at time t4 and the service is registered on node N0. At time t5, node N1 boots. Because the service is already registered, no action occurs. However, if a "notification action" is defined, it would be executed on node N1 at this time. At time t6, node N2 gracefully leaves the cluster, which also has no influence on the service. At time t7, node N0 begins a graceful leave. This results in the service being in the stopping state with the transfer intent set. Stopping is completed at time t8. The service is transferred to node N1 in the starting state at time t9. The service becomes registered on N1 at time t10 when start completes. At time t11, node N0 completes the graceful leave and goes down. This does not represent a change in the database because the up/down status of a node is not a state maintained in the database. At time t12, node N1 begins a graceful leave which places the service in a stopping state. When stop completes at time t13, the service becomes de-registered and the transfer intent flag is cleared. STATE TRANSITION EXAMPLE A Event Time Summary State Summary Flags Node N0 State Node N0 Flags Node N1 State Node N1 Flags Node N2 State Node N2 Flags t0 - down I down I down N2 boots t1 - down I down I up N0 boots t2 - M up I down I up 5 minutes elapse t3 Starting Starting I down I up Start completes t4 Registered Registered I down I up N1 boots t5 Registered Registered I up up N2 gracefully leaves cluster t6 Registered Registered I up I down N0 begins graceful leave t7 Stopping T Stopping I up I down Stop completes t8 - T I up I down t9 Starting T - I Starting I down Start completes t10 Registered ?? I Registered I up N0 completes graceful leave t11 Registered down I Registered I up N1 begins graceful leave t12 Stopping T down I Stopping I up stop completes t13 - down I - I up
T = transfer intent

M = awaiting maturity

I = available node set

In an example illustrated in Table 3, node N0 is defined as a preferred node with the remaining nodes as allowable. The maturity_count is set to 2. Auto_register is set to auto. Placement_policy is set to register_on_one. Transfer_cost is free. At the beginning of the example, the three nodes, N0, N1, and N2, in the cluster are all down. At time t1, node N0 boots and the service enters the starting state on node 0. At time t2, start completes and the service is registered on node N0. At time t3, node N1 boots and is notified that the service is registered on node N0. Between times t4 and t7, node N0 gracefully leaves the cluster. The service transfers to node N1 and is registered there. At time t8, node N2 boots and is notified that the service is registered on node N1. At time t9, node N0 boots and is notified that the service is registered on node 1. At time t10, shortly after it boots, node N0 notices that free transfers are allowed and that it is preferred over node N1. Node N0 will automatically initiate a transfer. The service enters the stopping state 86 and stop completes at time t11. At time t12, the service enters the starting state on node N0. Starting is complete at time t13 and the service is registered on node N0. Nodes N1 and N2 are notified that the service is registered on node N0. STATE TRANSITION EXAMPLE B Event Time Summary State Summary Flags Node N0 State Node N0 Flags Node N1 State Node N1 Flags Node N2 State Node N2 Flags t0 - N down I down I down I 0 boots t1 Starting N Starting I down I down I Start completes t2 Registered N Registered I down I down I 1 boots t3 Registered N Registered I -(notified) I down I 0 leaves cluster t4-t7 Registered N down I Registered I down I 2 boots t8 Registered N down I Registered I -(notified) I 0 boots t9 Registered N -(notified) I Registered I - I Takeover t10 Stopping N,T - I Stopping I - I Stop completes t11 - N,T - I - I - I t12 Starting N,T Starting I - I - I Start completes t13 Registered N,T Registered I -(notified) I -(notified) I
T = transfer intent

M = awaiting maturity

I = available node set

N = has notify

5. Event Manager Daemon

Registration and launch service 56 will automatically request event manager daemon 58 to monitor for a condition relating to a client service. Event manager daemon 58 responds by defining an event_group. The registration and launch service will request the event manager daemon to enable this event_group upon registration of the service and disable the event_group in response to de-registration of the service. If event manager daemon 58 detects a problem, an event action 62 will be invoked to resolve the problem. No direct communication will be returned to the registration and launch service.

Table 4 illustrates the fields associated with an event monitored by event manager daemon 58. In addition to the event name field, there are fields for IN parameters and OUT parameters which define the event that the appropriate event watcher is set to detect and the OUT parameters are filled in when the event occurs. The output is made available to the action function 62 associated with the particular event. Event_groups are used to logically associate otherwise independent events in order to specify when, where, and under what conditions to enable them. EVENT MANAGER DAEMON Field Description Event Name A string that identifies an event instance; it is unique within the cluster. IN Parameters A fixed set of name-value pairs that define an event; they are used by the appropriate event watcher to detect the event. OUT Parameters A fixed set of name-value pairs that describe an occurrence of the event. Action Function This command line describes what happens if the event occurs; it may reference values from the IN and OUT parameters.

In event manager subsystem 54, event manager daemon 58 is the center of control. All watchers 60a-60g connect via a communication library to the event manager daemon. One of the watchers provided in event manager subsystem 54 is membership manager watcher 60e, which receives notifications from membership manager subsystem 46 of node transitions in the manner previously described and provides an interface to transition notification framework 44. Event manager subsystem 54 provides awareness to registration and launch service 56 of such node transitions.

One example of an application for which registration and launch service 56 is especially apropos is to provide a floating license server on cluster system 25. The licensed software could be established as a service and could be allowed to execute on a given number of nodes in the cluster. The registration and launch service will run the start program that brings up the licensed software on one of the nodes. If that node goes down gracefully, or ungracefully, the registration and launch service will transfer the licensed software to a new node, after recovery if the leave was ungraceful.

Thus, it is seen that the present embodiment provides a tightly coordinated cluster membership manager framework which coordinates the joining or leaving among all nodes in a cluster, including taking the multiple layers of involved subsystems through the transitions. One of the subsystems may be in user space and carries out the transfers of client services, including user applications, resulting from nodes joining and leaving the cluster. Other user space applications may register with the membership manager transition notification framework at run time. Thus, a robust system is provided which enhances the high aggregate performance of the multiprocessor cluster technology.

The present embodiment facilitates the use of multiprocessor cluster systems with operating systems having multiple subsystems which are layered by taking all of the involved subsystems through node transitions. It also brings cluster awareness to non-cluster-aware client services, which include a wide variety of computing activities including user applications. This allows users to treat the cluster system as a single unit with the cluster system providing cluster-wide availability to the client service, including initiation of the client service on a particular node, migration of the client service between nodes, and termination of the client service.

Changes and modifications in the specifically described embodiments can be carried out without departing from the principles of the invention, which is intended to be limited only by the scope of the appended claims.


Anspruch[de]
  1. Betriebsverfahren für ein Multiprozessorsystem mit mehreren Knoten, einer gemeinsamen, für alle Knoten zugänglichen Ressource und mehreren interdependenten Ebenen von Betriebssystem-Subsystemen auf jedem der genannten Knoten, wobei das Verfahren die folgenden Schritte umfasst:
    • Kombinieren spezieller der genannten Knoten in einem Cluster, der Benutzern des genannten Systems im Wesentlichen als ein vereintes System erscheint, und
    • Benachrichtigen der Ebenen von Subsystemen, die auf gegenwärtig in dem Cluster befindlichen Knoten abgearbeitet werden, über Übergänge von sich dem Cluster anschließenden und es verlassenden Knoten, um eine konsistente Ansicht aktiver Mitgliedschaft in dem Cluster bereitzustellen.
  2. Verfahren nach Anspruch 1, bei dem die Subsysteme in Ebenen interdependent sind, wobei Subsysteme höherer Ebenen von dem Betrieb von Subsystemen niedrigerer Ebenen abhängig sind.
  3. Verfahren nach Anspruch 2, das die Benachrichtigung von einem der genannten Subsysteme auf allen der genannten, gegenwärtig in dem Cluster befindlichen Knoten über einen Übergang und das Verarbeiten dieser Benachrichtigung an dem genannten einen der genannten Subsysteme, bevor ein weiteres der genannten Subsysteme auf allen der genannten, gegenwärtig in dem Cluster befindlichen Knoten über den Übergang benachrichtigt wird, aufweist.
  4. Verfahren nach Anspruch 3, das das Benachrichtigen von Subsystemen, angefangen mit Subsystemen niedrigerer Ebene und der Reihe nach durch höhere Subsystemebenen fortschreitend, über einen Übergang eines sich dem Cluster anschließenden Knotens aufweist.
  5. Verfahren nach Anspruch 3, das das Benachrichtigen von Subsystemen, angefangen mit Subsystemen höherer Ebene und der Reihe nach durch niedrigere Subsystemebenen fortschreitend, über einen Übergang eines das Cluster kontrolliert verlassenden Knotens aufweist.
  6. Verfahren nach Anspruch 3, das das Benachrichtigen von Subsystemen, angefangen mit Subsystemen niedrigerer Ebene und der Reihe nach durch höhere Subsystemebenen fortschreitend, über einen Übergang eines von anderen Prozessoren aus dem Cluster gedrängten Knotens aufweist.
  7. Verfahren nach einem der Ansprüche 2 bis 6, bei dem die genannten Subsysteme ein Subsystem höherer Ebene aufweisen, das mit Benutzerprogrammen interagiert.
  8. Verfahren nach Anspruch 7, bei dem das genannte Subsystem höherer Ebene einen Dienst aufweist, der Benutzerprogramme automatisch und atomisch zu anderen Knoten transferiert, wenn der die Benutzerprogramme ausführende Knoten das Cluster verlässt.
  9. Verfahren nach einem der Ansprüche 2 bis 8, bei dem die genannten Subsysteme ein Verteilter-Sperrenmanager-Subsystem, ein Virtueller-Diskmanager-Subsystem und ein Gemeinschaftsdatei-Subsystem aufweisen.
  10. Verfahren nach einem der Ansprüche 1 bis 9, bei dem das genannte Verfahren für einen Übergang eines sich dem Cluster anschließenden Knotens die folgenden Schritte aufweist:
    • a) Registrieren von Subsystemen des sich anschließenden Knotens zum Empfangen von Übergangsbenachrichtigungen,
    • b) Anschließen des Knotens an das Cluster und
    • c) Benachrichtigen registrierter Subsysteme in dem Cluster darüber, dass der sich anschließende Knoten sich dem Cluster angeschlossen hat.
  11. Verfahren nach einem der Ansprüche 1 bis 10, bei dem das genannte Verfahren für einen Übergang eines von einem anderen Knoten aus dem Cluster gedrängten Knotens die folgenden Schritte aufweist:
    • a) Benachrichtigen registrierter Subsysteme durch den anderen Knoten darüber, dass der eine Knoten aus dem Cluster gedrängt wird, und
    • b) Übertragen registrierter Programme, die auf dem genannten einen Knoten ausgeführt werden, atomisch auf einen anderen Knoten und Wiederherstellen der Programme zum Ausführen auf dem genannten anderen Knoten.
  12. Verfahren nach einem der Ansprüche 1 bis 11 einschließlich Initiieren von Client-Diensten auf speziellen der genannten Knoten in einem Cluster auf eine Weise, die den Client-Diensten im Wesentlichen als ein vereintes System erscheint, Wählen eines Knotens für jeden Client-Dienst, Registrieren des Client-Dienstes bei diesem Knoten und Benachrichtigen von gegenwärtig in dem Cluster befindlichen Knoten darüber, dass der spezielle Client-Dienst bei dem speziellen Knoten registriert ist, wodurch der spezielle Dienst auf einen anderen Knoten übertragen werden kann, wenn der spezielle Knoten das Cluster verlässt.
  13. Verfahren nach Anspruch 12, das ferner das Starten eines Client-Dienstes auf einem Knoten gemäß einem in dem Client-Dienst enthaltenen Aktionsparameter in Reaktion auf das Registrieren dieses Client-Dienstes bei diesem Knoten aufweist.
  14. Verfahren nach Anspruch 13, das ferner das Gruppieren von Client-Diensten als Eltern-Client-Dienst und wenigstens einen Kind-Client-Dienst, das Registrieren gruppierter Client-Dienste bei dem gleichen Knoten und das Starten gruppierter Client-Dienste gemäß einem in dem Eltern-Client-Dienst enthaltenen Aktionsparameter aufweist.
  15. Verfahren nach Anspruch 12, 13 oder 14, bei dem das genannte Wählen eines Knotens das Bereitstellen einer Datenbank von Wählfaktoren für den Client-Dienst und das Anwenden der genannten Wählfaktoren auf die gegenwärtig in dem Cluster befindlichen Knoten aufweist, wobei die genannten Wählfaktoren Regeln festlegen, die Knoten mit dem Client-Dienst in Bezug setzen.
  16. Verfahren nach Anspruch 15, bei dem die genannten Wählfaktoren aus der Gruppe ausgewählt werden, die zulässige Knoten, nicht zulässige Knoten und Knotenpräferenzen aufweist.
  17. Verfahren nach einem der Ansprüche 12 bis 16, das ferner das Benachrichtigen von sich an das Cluster anschließenden Knoten darüber aufweist, dass der spezielle Client-Dienst bei den speziellen Knoten registriert ist.
  18. Verfahren nach einem der Ansprüche 12 bis 17, das die Überwachung eines bei einem Knoten registrierten Client-Dienstes unter Verwendung einer Ereignisüberwachung an diesem Knoten aufweist.
  19. Verfahren nach Anspruch 18, das das Aktivieren der Ereignisüberwachung in Reaktion auf das Registrieren des Client-Dienstes und das Deaktivieren der Ereignisüberwachung in Reaktion auf das Deregistrieren des Client-Dienstes aufweist.
  20. Verfahren nach einem der Ansprüche 12 bis 19, bei dem die genannten mehreren Subsysteme einen Cluster-Mitgliedschaftsmanager haben, der bestimmt, welche der genannten Knoten gegenwärtig in dem Cluster sind, und bei dem der genannte Cluster-Mitgliedschaftsmanager Benachrichtigung darüber bereitstellt, welche Knoten in dem Cluster sind.
  21. Verfahren nach Anspruch 20, bei dem das genannte Initiieren von Diensten auf einem speziellen der genannten Knoten von einem anderen der genannten mehreren Subsysteme durchgeführt wird.
  22. Verfahren nach einem der Ansprüche 1 bis 21, das das Initiieren von Client-Diensten auf speziellen der genannten Knoten in einem Cluster auf eine Weise, die dem Client-Dienst im Wesentlichen als ein vereintes System erscheint, das Registrieren eines Client-Dienstes bei einem der genannten Knoten und das Starten des Client-Dienstes auf diesem Knoten gemäß einem von dem Client-Dienst aufgewiesenen Aktionsparameter in Reaktion auf das Registrieren dieses Benutzer-Dienstes bei diesem Knoten aufweist.
  23. Verfahren nach Anspruch 22, das das Transferieren des Client-Dienstes zu einem anderen Knoten aufweist, wenn dieser Knoten das Cluster verlässt.
  24. Verfahren nach Anspruch 23, bei dem das genannte Transferieren das Neustarten des Client-Dienstes auf dem anderen Knoten gemäß dem genannten Aktionsparameter aufweist.
  25. Multiprozessor-Cluster-System, umfassend:
    • mehrere Knoten (26),
    • eine gemeinsame Ressource (28), die für alle Knoten zugänglich ist,
    • ein Cluster-Kommunikationsmedium (36) zwischen den genannten Knoten,
    • mehrere interdependente Ebenen von Betriebssystem-Subsystemen (46-54) auf jedem der genannten Knoten,
    • wobei die genannten Subsysteme ein Cluster-Mitgliedschaftsmanager-Subsystem (46) zum Benachrichten von Subsystemen, die auf gegenwärtig in dem Cluster befindlichen Knoten abgearbeitet werden, über sich dem Cluster anschließende und es verlassende Knoten aufweist, um eine konsistente Ansicht aktiver Mitgliedschaft in dem Cluster bereitzustellen, und
    • ein Registrierungs- und Start-Subsystem (56) zum Initiieren von Client-Diensten auf speziellen der genannten Knoten in einem Cluster auf eine Weise, die den Client-Diensten im Wesentlichen als ein vereinter Knoten erscheint, wobei das genannte Registrierungs- und Start-Subsystem einen Knoten für jeden Client-Dienst wählt, den Client-Dienst bei diesem Knoten registriert und gegenwärtig in dem Cluster befindliche Knoten darüber benachrichtigt, dass der spezielle Dienst bei dem speziellen Knoten registriert ist.
  26. Multiprozessor-Cluster-System nach Anspruch 25, bei dem die Subsysteme in Ebenen interdependent sind, wobei Subsysteme höherer Ebenen von dem Betrieb von Subsystemen niedrigerer Ebenen abhängig sind.
  27. Multiprozessor-Cluster-System nach Anspruch 26, bei dem das genannte Mitgliedschaftsmanager-Subsystem (46) eines der genannten Subsysteme auf allen der genannten, gegenwärtig in dem Cluster befindlichen Knoten über einen Übergang benachrichtigt und dieses genannte eine der genannten Subsysteme auf allen Knoten diese Benachrichtigung verarbeitet, bevor das genannte Mitgliedschaftsmanager-Subsystem ein weiteres der genannten Subsysteme auf allen der genannten, gegenwärtig in dem Cluster befindlichen Knoten über den Übergang benachrichtigt.
  28. Multiprozessor-Cluster-System nach Anspruch 27, bei dem das genannte Mitgliedschaftsmanager-Subsystem (46) Subsysteme angefangen mit Subsystemen niedrigerer Ebene und der Reihe nach durch höhere Subsystemebenen fortschreitend über einen Übergang eines sich dem Cluster anschließenden Knotens benachrichtigt.
  29. Multiprozessor-Cluster-System nach Anspruch 27 oder Anspruch 28, bei dem das genannte Mitgliedschaftsmanager-Subsystem (46) Subsysteme angefangen mit Subsystemen höherer Ebene und der Reihe nach durch niedrigere Subsystemebenen fortschreitend über einen Übergang eines das Cluster kontrolliert verlassenden Knotens benachrichtigt.
  30. Multiprozessor-Cluster-System nach Anspruch 27, 28 oder 29, bei dem das genannte Mitgliedschattsmanager-Subsystem (46) Subsysteme angefangen mit Subsystemen niedrigerer Ebene und der Reihe nach durch höhere Subsystemebenen fortschreitend über einen Übergang eines von anderen Prozessoren aus dem Cluster gedrängten Knotens benachrichtigt.
  31. Multiprozessor-Cluster-System nach einem der Ansprüche 25 bis 30, bei dem das genannte Registrierungs- und Start-Subsystem (56) einen Client-Dienst auf einem Knoten gemäß einem in den Client-Dienst eingeschlossenen Aktionsparameter in Reaktion auf das Registrieren dieses Client-Dienstes bei diesem Knoten startet.
  32. Multiprozessor-Cluster-System nach Anspruch 31, bei dem das genannte Registrierungs- und Start-Subsystem (56) zum Gruppieren von Client-Diensten als einen Eltern- Client-Dienst und wenigstens einen Kind-Client-Dienst ausgeführt ist und bei dem das genannte Registrierungs- und Start-Subsystem ferner gruppierte Client-Dienste bei dem gleichen Knoten registriert und gruppierte Client- Dienste gemäß einem in den Eltern-Client-Dienst eingeschlossenen Aktionsparameter startet.
  33. Multiprozessor-Cluster-System nach einem der Ansprüche 25 bis 32, bei dem das genannte Registrierungs- und Start-Subsystem (56) eine Datenbank mit Wählfaktoren für den Client-Dienst aufweist und die genannten Wählfaktoren auf die gegenwärtig in dem Cluster befindlichen Knoten anwendet, um den Knoten zum Registrieren eines Client-Dienstes zu wählen, wobei die genannten Wählfaktoren Regeln festlegen, die Knoten mit dem Client-Dienst in Bezug setzen.
  34. Multiprozessor-Cluster-System nach einem der Ansprüche 25 bis 33, bei dem das genannte Registrierungs- und Start-Subsystem (56) ferner sich an das Cluster anschließende Knoten darüber benachrichtigt, dass der spezielle Client-Dienst bei dem speziellen Knoten registriert ist.
  35. Multiprozessor-Cluster-System nach einem der Ansprüche 25 bis 34, das ein Ereignismanager-Subsystem (54) zum Feststellen von Clusterfehlern und zum Reagieren auf sie aufweist und bei dem das Registrierungs- und Start-Subsystem (56) auf das Ereignismanager-Subsystem reagiert.
  36. Multiprozessor-Cluster-System nach Anspruch 35, bei dem das genannte Ereignismanager-Subsystem (54) eine Ereignisüberwachung (60) zum Überwachen eines bei einem Knoten registrierten Client-Dienstes an diesem Knoten aufweist.
  37. Computer-geeignetes Medium, in das Programmcode eingegliedert ist, wobei der genannte Programmcode ein Betriebssystem für ein Multiprozessor-Cluster-System definiert, das mehrere Knoten (26) hat, eine gemeinsame, für alle Prozessoren zugängliche Ressource (28) und mehrere interdependente Ebenen von Subsystemen (46 - 54) aufweist, wobei eines der genannten Subsysteme ein Cluster-Mitgliedschaftsmanager-Subsystem (46) zum Kombinieren spezieller der genannten Knoten in einem Cluster ist, das Benutzern des genannten Systems im Wesentlichen als ein vereintes System erscheint, wobei der genannte Cluster-Mitgliedschaftsmanager zum Benachrichtigen der Ebenen von Subsystemen, die auf gegenwärtig in einem Cluster befindlichen Knoten abgearbeitet werden, über Übergänge von sich dem Cluster anschließenden und es verlassenden Knoten ausgeführt ist, um eine konsistente Ansicht aktiver Mitgliedschaft in dem Cluster bereitzustellen.
  38. Computer-geeignetes Medium nach Anspruch 37, wobei das Betriebssystem ferner einen Registrierungs- und Start-Dienst (56) hat zum Initiieren von Client-Diensten an speziellen von Knoten in einem Cluster auf eine Weise, die den Client-Diensten im Wesentlichen als ein vereintes System erscheint, wobei der genannte Registrierungs- und Start-Dienst einen Knoten für jeden Client-Dienst wählt, den Client-Dienst bei diesem Knoten registriert und gegenwärtig in dem Cluster befindliche Knoten darüber benachrichtigt, dass der spezielle Dienst bei dem speziellen Knoten registriert ist.
Anspruch[en]
  1. A method of operating a multiprocessor system having multiple nodes, a shared resource accessible to all nodes, and multiple interdependent levels of operating system subsystems on each of said nodes; the method comprising the steps of:
    • combining particular ones of said nodes in a cluster that appears substantially as a unified system to users of said system; and
    • notifying the levels of subsystems running on nodes presently in the cluster of transitions of nodes joining and leaving the cluster in order to provide a consistent view of active membership in the cluster.
  2. The method of claim 1 wherein the subsystems are interdependent in levels, with higher level subsystems dependent on the operation of lower level subsystems.
  3. The method of claim 2 including notifying one of said subsystems on all of said nodes presently in the cluster of a transition, and processing that notification at said one of said subsystems prior to notifying another of said subsystems on all of said nodes presently in the cluster of the transition.
  4. The method of claim 3 including notifying subsystems beginning with lower level subsystems and proceeding in sequence through higher levels of subsystems of a transition of a node joining the cluster.
  5. The method of claim 3 including notifying subsystems beginning with higher lever subsystems and proceeding in sequence through lower levels of subsystems of a transition of a node gracefully leaving the cluster.
  6. The method of claim 3 including notifying subsystems beginning with lower level subsystems and proceeding in sequence through higher levels of subsystems of a transition of a node forced from the cluster by other processors.
  7. The method of any of claims 2 to 6 wherein said subsystems include a higher level subsystem which interacts with user programs.
  8. The method of claim 7 wherein said higher level subsystem includes a service which automatically and atomically transfers user programs to other nodes when the node executing the user programs leaves the cluster.
  9. The method of any of claims 2 to 8 wherein said subsystems include a distributed lock manager subsystem, a virtual disk manager subsystem and a shared file subsystem.
  10. The method of any of claims 1 to 9 wherein, for a transition of a node joining the cluster, said method includes the steps of:
    • a) registering subsystems of the joining node to receive transition notifications;
    • b) joining the node to the cluster; and
    • c) notifying registered subsystems in the cluster that the joining node has joined the cluster.
  11. The method of any of claims 1 to 10 wherein, for a transition of one node being forced out of the cluster by another node, said method includes the steps of:
    • a) the another node notifying registered subsystems that the one node is being forced out of the cluster; and
    • b) transferring registered programs executing on said one node atomically to a different node and recovering the programs to execute on said different node.
  12. The method of any of claims 1 to 11 including initiating client services on particular ones of said nodes in a cluster in a manner that appears substantially as a unified system to the client services, choosing a node for each client service, registering the client service with that node, and notifying nodes presently in the cluster that the particular client service is registered with the particular node, whereby the particular service can be transferred to another node if the particular node leaves the cluster.
  13. The method of claim 12 further including launching a client service on a node according to an action parameter included with the client service in response to registering that client service with that node.
  14. The method of claim 13 further including grouping client services as a parent client service and at least one child client service, registering grouped client services with the same node and launching grouped client services according to an action parameter included with the parent client service.
  15. The method of claim 12, 13 or 14 wherein said choosing a node includes providing a database of choosing factors for the client service and applying said choosing factors to the nodes presently in the cluster, said choosing factors establishing rules relating nodes to the client service.
  16. The method of claim 15, wherein said choosing factors are selected from the group including allowable nodes, disallowable nodes and node preferences.
  17. The method of any of claims 12 to 16 further including notifying nodes joining the cluster that the particular client service is registered with the particular nodes.
  18. The method of any of claims 12 to 17 including monitoring a client service registered with a node at that node using an event watcher.
  19. The method of claim 18 including enabling the event watcher in response to registering the client service and disabling the event watcher in response to de-registering the client service.
  20. The method of any of claims 12 to 19 wherein said multiple subsystems includes a cluster membership manager which controls which of said nodes are presently in the cluster and wherein said cluster membership manager provides notification of which nodes are in the cluster.
  21. The method of claim 20 wherein said initiating services on a particular one of said nodes is performed by another one of said multiple subsystems.
  22. The method of any of claims 1 to 21 including initiating client services on particular ones of said nodes in a cluster in a manner that appears substantially as a unified system to the client service, registering a client service with one of said nodes, and launching the client service on that node according to an action parameter included with the client service in response to registering that user service with that node.
  23. The method of claim 22 including transferring the client service to another node if that node leaves the cluster.
  24. The method of claim 23 wherein said transferring includes relaunching the client service on the another node according to said action parameter.
  25. A multiprocessor cluster system comprising:
    • multiple nodes (26);
    • a shared resource (28) accessible to all nodes;
    • a cluster communication medium (36) between said nodes;
    • multiple interdependent levels of operating system subsystems (46-54) on each of said nodes;
    • said subsystems including a cluster membership manager subsystem (46) adapted to notify subsystems running on nodes presently in the cluster of transitions of nodes joining and leaving the cluster, in order to provide a consistent view of active membership in the cluster; and
    • a registration and launch subsystem (56) adapted to initiate client services on particular ones of said nodes in a cluster in a manner that appears substantially as a unified node to the client services, wherein said registration and launch subsystem chooses a node for each client service, registers the client service with that node, and notifies nodes presently in the cluster that the particular service is registered with the particular node.
  26. The multiprocessor cluster system of claim 25 wherein the subsystems are interdependent in levels, with higher level subsystems dependent on the operation of lower level subsystems.
  27. The multiprocessor cluster system of claim 26 wherein said membership manager subsystem (46) notifies one of said subsystems on all of said nodes presently in the cluster of a transition, and that one of said subsystems on all nodes processes that notification prior to said membership manager subsystem notifying another of said subsystems on all of said nodes presently in the cluster of the transition.
  28. The multiprocessor cluster system of claim 27 wherein said membership manager subsystem (46) notifies subsystems beginning with lower level subsystems and proceeding in sequence through higher levels of subsystems of a transition of a node joining the cluster.
  29. The multiprocessor cluster system of claim 27 or 28 wherein said membership manager subsystem (46) notifies subsystems beginning with higher lever subsystems and proceeding in sequence through lower levels of subsystems of a transition of a node gracefully leaving the cluster.
  30. The multiprocessor cluster system of claim 27, 28 or 29 wherein said membership manager subsystem (46) notifies subsystems beginning with lower level subsystems and proceeding in sequence through higher levels of subsystems of a transition of a node forced from the cluster by other processors.
  31. The multiprocessor cluster system of any of claims 25 to 30 wherein said registration and launch subsystem (56) launches a client service on a node according to an action parameter included with the client service in response to registering that client service with that node.
  32. The multiprocessor cluster system of claim 31 wherein said registration and launch subsystem (56) is adapted to group client services as a parent client service and at least one child client service, and wherein said registration and launch subsystem further registers grouped client services with the same node and launches grouped client services according to an action parameter included with the parent client service.
  33. The multiprocessor cluster system of any of claims 25 to 32 wherein said registration and launch subsystem (56) includes a database of choosing factors for the client service and applies said choosing factors to the nodes presently in the cluster to choose the node for registering a client service, said choosing factors establishing rules relating nodes to the client service.
  34. The multiprocessor cluster system of any of claims 25 to 33 wherein said registration and launch subsystem (56) further notifies nodes joining the cluster that the particular client service is registered with the particular node.
  35. The multiprocessor cluster system of any of claims 25 to 34 including an event manager subsystem (54) adapted to detect and react to cluster errors, and wherein the registration and launch subsystem (56) is responsive to the event manager subsystem.
  36. The multiprocessor cluster system of claim 35 wherein said event manager subsystem (54) includes an event watcher (60) for monitoring a client service registered with a node at that node.
  37. A computer usable medium in which program code is embodied, said program code defining an operating system for a multiprocessor cluster system having multiple nodes (26), a shared resource (28) accessible to all processors, and including multiple interdependent levels of subsystems (46-54), one of said subsystems being a cluster membership manager subsystem (46) for combining particular ones of said nodes in a cluster that appears substantially as a unified system to users of said system, said cluster membership manager adapted to notify the levels of subsystems running on nodes presently in a cluster of transitions of nodes joining and leaving the cluster in order to provide a consistent view of active membership in the cluster.
  38. The computer usable medium of claim 37 wherein the operating system further includes a registration and launch service (56) adapted to initiate client services on particular ones of nodes in a cluster in a manner that appears substantially as a unified system to the client services, wherein said registration and launch service chooses a node for each client service, registers the client service with that node, and notifies nodes presently in the cluster that the particular service is registered with the particular node.
Anspruch[fr]
  1. Procédé d'exploitation d'un système multiprocesseur ayant des noeuds multiples, une ressource partagée accessible à tous les noeuds, et de multiples niveaux interdépendants de sous-systèmes de système d'exploitation sur chacun desdits noeuds ; le procédé comprenant les étapes de :
    • combinaison de noeuds particuliers desdits noeuds en une grappe qui présente substantiellement l'aspect d'un système unifié aux utilisateurs dudit système ; et
    • notification aux niveaux de sous-systèmes exécutés sur les noeuds se trouvant dans la grappe de transitions de noeuds adhérant à et quittant la grappe afin de fournir une vue cohérente d'appartenance active à la grappe.
  2. Procédé selon la revendication 1, dans lequel les sous-systèmes sont interdépendants en terme de niveaux, des sous-systèmes de niveaux supérieurs dépendant du fonctionnement de sous-systèmes de niveaux inférieurs.
  3. Procédé selon la revendication 2, comportant la notification à l'un desdits sous-systèmes sur tous lesdits noeuds se trouvant dans la grappe d'une transition, et le traitement de cette notification au niveau dudit un desdits sous-systèmes avant la notification de la transition à un autre desdits sous-systèmes sur tous lesdits noeuds se trouvant dans la grappe.
  4. Procédé selon la revendication 3, comportant la notification à des sous-systèmes en commençant par des sous-systèmes de niveau inférieur et en continuant en séquence par des niveaux supérieurs de sous-systèmes d'une transition d'un noeud adhérant à la grappe.
  5. Procédé selon la revendication 3, comportant la notification à des sous-systèmes en commençant par des sous-systèmes de niveau supérieur et en continuant en séquence par des niveaux inférieurs de sous-systèmes d'une transition d'un noeud quittant harmonieusement la grappe.
  6. Procédé selon la revendication 3, comportant la notification à des sous-systèmes en commençant par des sous-systèmes de niveau inférieur et en continuant en séquence par les niveaux supérieurs de sous-systèmes d'une transition d'un noeud expulsé de la grappe par d'autres processeurs.
  7. Procédé selon l'une quelconque des revendications 2 à 6, dans lequel lesdits sous-systèmes comportent un sous-système de niveau supérieur qui interagit avec des programmes d'utilisateur.
  8. Procédé selon la revendication 7, dans lequel ledit sous-système de niveau supérieur comporte un service qui transfère automatiquement et atomiquement des programmes d'utilisateur à d'autres noeuds quand le noeud exécutant les programmes d'utilisateur quitte la grappe.
  9. Procédé selon l'une quelconque des revendications 2 à 8, dans lequel lesdits sous-systèmes comportent un sous-système de gestionnaire de verrouillage distribué, un sous-système de gestionnaire de disque virtuel et un sous-système de fichier partagé.
  10. Procédé selon l'une quelconque des revendications 1 à 9, dans lequel, pour une transition d'un noeud adhérant à la grappe, ledit procédé comporte les étapes de :
    • a) enregistrement de sous-systèmes du noeud adhérant afin qu'ils reçoivent des notifications de transitions ;
    • b) adhésion du noeud à la grappe ; et
    • c) notification aux sous-systèmes enregistrés dans la grappe que le noeud adhérant a adhéré à la grappe.
  11. Procédé selon l'une quelconque des revendications 1 à 10, dans lequel, pour une transition d'un noeud étant expulsé de la grappe par un autre noeud, ledit procédé comporte les étapes de :
    • a) notification par l'autre noeud aux sous-systèmes enregistrés que le noeud est expulsé de la grappe ; et
    • b) transfert des programmes enregistrés exécutés sur ledit noeud atomiquement sur un noeud différent et récupération des programmes à exécuter sur ledit noeud différent.
  12. Procédé selon l'une quelconque des revendications 1 à 11, comportant la préparation de services clients sur des noeuds particuliers desdits noeuds dans une grappe d'une manière qui présente substantiellement l'aspect d'un système unifié aux services clients, le choix d'un noeud pour chaque service client, l'enregistrement du service client avec ce noeud, et la notification aux noeuds se trouvant dans la grappe que le service client particulier est enregistré avec le noeud particulier, par lequel le service particulier peut être transféré sur un autre noeud si le noeud particulier quitte la grappe.
  13. Procédé selon la revendication 12, comportant en outre le lancement d'un service client sur un noeud conformément à un paramètre d'action inclus avec le service client en réponse à l'enregistrement de ce service client avec ce noeud.
  14. Procédé selon la revendication 13, comportant en outre le groupement de services client en tant qu'un service client parent et au moins un service client enfant, l'enregistrement de services clients groupés avec le même noeud et le lancement de services clients groupés conformément à un paramètre d'action inclus avec le service client parent.
  15. Procédé selon la revendication 12, 13 ou 14, dans lequel ledit choix d'un noeud comporte la fourniture d'une base de données de facteurs de choix du service client et l'application desdits facteurs de choix aux noeuds se trouvant dans la grappe, lesdits facteurs de choix établissant des règles liant les noeuds au service client.
  16. Procédé selon la revendication 15, dans lequel lesdits facteurs de choix sont sélectionnés dans la grappe comportant des noeuds admissibles, des noeuds non admissibles et des préférences de noeuds.
  17. Procédé selon l'une quelconque des revendications 12 à 16, comportant en outre la notification aux noeuds adhérant à la grappe que le service client particulier est enregistré avec les noeuds particuliers.
  18. Procédé selon l'une quelconque des revendications 12 à 17, comportant la surveillance d'un service client enregistré avec un noeud au niveau de ce noeud en utilisant un observateur d'événements.
  19. Procédé selon la revendication 18, comportant l'activation de l'observateur d'événements en réponse à l'enregistrement du service client et la désactivation de l'observateur d'événements en réponse au désenregistrement du service client.
  20. Procédé selon l'une quelconque des revendications 12 à 19, dans lequel lesdits sous-systèmes multiples comportent un gestionnaire d'appartenance à la grappe qui commande lesdits noeuds qui se trouvent dans la grappe et dans lequel ledit gestionnaire d'appartenance à la grappe fournit une notification des noeuds qui se trouvent dans la grappe.
  21. Procédé selon la revendication 20, dans lequel lesdits services de préparation sur un noeud particulier desdits noeuds est exécuté par un autre desdits sous-systèmes multiples.
  22. Procédé selon l'une quelconque des revendications 1 à 21, comportant la préparation de services clients sur des noeuds particuliers desdits noeuds dans une grappe d'une manière qui présente substantiellement l'aspect d'un système unifié au service client, l'enregistrement d'un service client avec l'un desdits noeuds, et le lancement du service client sur ce noeud conformément à un paramètre d'action inclus avec le service client en réponse à l'enregistrement de ce service d'utilisateur avec ce noeud.
  23. Procédé selon la revendication 22, comportant le transfert du service client à un autre noeud si ce noeud quitte la grappe.
  24. Procédé selon la revendication 23, dans lequel ledit transfert comporte le lancement à nouveau du service client sur l'autre noeud conformément audit paramètre d'action.
  25. Système multiprocesseur en grappe comprenant :
    • des noeuds multiples (26) ;
    • une ressource partagée (28) accessible à tous les noeuds ;
    • un support de communication de grappe (36) entre lesdits noeuds ;
    • de multiples niveaux interdépendants de sous-systèmes de système d'exploitation (46 à 54) sur chacun desdits noeuds ;
    • lesdits sous-systèmes comportant un sous-système de gestionnaire d'appartenance à la grappe (46) adapté pour notifier aux sous-systèmes exécutés sur des noeuds se trouvant dans la grappe des transitions de noeuds adhérant à et quittant la grappe, afin de fournir une vue cohérente d'appartenance active à la grappe ; et
    • un sous-système d'enregistrement et de lancement (56) adapté pour initialiser des services clients sur des noeuds particuliers desdits noeuds dans une grappe d'une manière qui présente substantiellement l'aspect d'un noeud unifié aux services clients, dans lequel le sous-système d'enregistrement et de lancement choisit un noeud pour chaque service client, enregistre le service client avec ce noeud, et notifie aux noeuds se trouvant dans la grappe que le service particulier est enregistré avec le noeud particulier.
  26. Système multiprocesseur en grappe selon la revendication 25, dans lequel les sous-systèmes sont interdépendants en terme de niveaux, des sous-systèmes de niveaux supérieurs dépendant du fonctionnement de sous-systèmes de niveaux inférieurs.
  27. Système multiprocesseur en grappe selon la revendication 26, dans lequel ledit sous-système de gestionnaire d'appartenance (46) notifie à l'un desdits sous-systèmes sur tous lesdits noeuds se trouvant dans la grappe d'une transition, et ce dit sous-système desdits sous-systèmes sur tous les noeuds traite cette notification avant que le sous-système de gestionnaire d'appartenance ne notifie la transition à un autre desdits sous-systèmes sur tous lesdits noeuds se trouvant dans la grappe.
  28. Système multiprocesseur en grappe selon la revendication 27, dans lequel ledit sous-système de gestionnaire d'appartenance (46) notifie à des sous-systèmes en commençant par des sous-systèmes de niveau inférieur et en continuant en séquence par les niveaux supérieurs de sous-systèmes une transition d'un noeud adhérant à la grappe.
  29. Système multiprocesseur en grappe selon la revendication 27 ou 28, dans lequel ledit sous-système de gestionnaire d'appartenance notifie à des sous-systèmes en commençant par des sous-systèmes de niveau supérieur et en continuant en séquence par des niveaux inférieurs de sous-systèmes une transition d'un noeud quittant harmonieusement la grappe.
  30. Système multiprocesseur en grappe selon la revendication 27, 28 ou 29, dans lequel ledit sous-système de gestionnaire d'appartenance (46) notifie à des sous-systèmes en commençant par des sous-systèmes de niveau inférieur et en continuant en séquence par des niveaux supérieurs de sous-systèmes une transition d'un noeud expulsé de la grappe par d'autres processeurs.
  31. Système multiprocesseur en grappe selon l'une quelconque des revendications 25 à 30, dans lequel ledit sous-système d'enregistrement et de lancement (56) lance un service client sur un noeud conformément à un paramètre d'action inclus avec le service client en réponse à l'enregistrement de ce service client avec ce noeud.
  32. Système multiprocesseur en grappe selon la revendication 31, dans lequel ledit sous-système d'enregistrement et de lancement (56) est adapté pour grouper des services clients en tant qu'un service client parent et au moins un service client enfant, et dans lequel ledit sous-système d'enregistrement et de lancement enregistre en outre des services clients groupés avec le même noeud et lance des services clients groupés conformément à un paramètre d'action inclus avec le service client parent.
  33. Système multiprocesseur en grappe selon l'une quelconque des revendications 25 à 32, dans lequel ledit sous-système d'enregistrement et de lancement (56) comporte une base de données de facteurs de choix du service client et applique lesdits facteurs de choix aux noeuds se trouvant dans la grappe afin de choisir le noeud pour l'enregistrement d'un service client, lesdits facteurs de choix établissant des règles liant des noeuds au service client.
  34. Système multiprocesseur en grappe selon l'une quelconque des revendications 25 à 33, dans lequel ledit sous-système d'enregistrement et de lancement (56) notifie en outre aux noeuds adhérant à la grappe que le service client particulier est enregistré avec le noeud particulier.
  35. Système multiprocesseur en grappe selon l'une quelconque des revendications 25 à 34, comportant un sous-système de gestionnaire d'événements (54) adapté pour détecter et réagir à des erreurs de grappe, et dans lequel le sous-système d'enregistrement et de lancement (56) est sensible au sous-système de gestionnaire d'événements.
  36. Système multiprocesseur en grappe selon la revendication 35, dans lequel ledit sous-système de gestionnaire d'événements (54) comporte un observateur d'événements (60) pour surveiller un service client enregistré avec un noeud au niveau de ce noeud.
  37. Support utilisable par un ordinateur dans lequel un code programme est intégré, ledit code programme définissant un système d'exploitation pour un système multiprocesseur en grappe ayant des noeuds multiples (26), une ressource partagée (28) accessible à tous les processeurs, et comportant de multiples niveaux interdépendants de sous-systèmes (46 à 54), l'un desdits sous-systèmes étant un sous-système de gestionnaire d'appartenance à la grappe (46) pour combiner des noeuds particuliers desdits noeuds dans une grappe qui présente substantiellement l'aspect d'un système unifié aux utilisateurs dudit système, ledit gestionnaire d'appartenance à la grappe étant adapté pour notifier aux niveaux de sous-systèmes exécutés sur les noeuds se trouvant dans une grappe des transitions de noeuds adhérant à et quittant la grappe afin de fournir une vue cohérente d'appartenance active à la grappe.
  38. Support utilisable par un ordinateur selon la revendication 37, dans lequel le système d'exploitation comporte en outre un service d'enregistrement et de lancement (56) adapté pour préparer des services clients sur des noeuds particuliers de noeuds dans une grappe d'une manière qui présente substantiellement l'aspect d'un système unifié aux services clients, dans lequel ledit service d'enregistrement et de lancement choisit un noeud pour chaque service client, enregistre le service client avec ce noeud, et notifie aux noeuds se trouvant dans la grappe que le service particulier est enregistré avec le noeud particulier.






IPC
A Täglicher Lebensbedarf
B Arbeitsverfahren; Transportieren
C Chemie; Hüttenwesen
D Textilien; Papier
E Bauwesen; Erdbohren; Bergbau
F Maschinenbau; Beleuchtung; Heizung; Waffen; Sprengen
G Physik
H Elektrotechnik

Anmelder
Datum

Patentrecherche

Patent Zeichnungen (PDF)

Copyright © 2008 Patent-De Alle Rechte vorbehalten. eMail: info@patent-de.com