Patent Grant
7788544
Distributed computing systems are increasingly being utilized to support business as well as technical applications. Typically, distributed computing systems are constructed from a collection of computing nodes that combine to provide a set of processing services to implement the distributed computing applications. Each of the computing nodes in the distributed computing system is typically a separate, independent computing device interconnected with each of the other computing nodes via a communications medium, e.g., a network.
One challenge with distributed computing systems is the organization, deployment and administration of such a system within an enterprise environment. For example, it is often difficult to manage the allocation and deployment of enterprise computing functions within the distributed computing system. An enterprise, for example, often includes several business groups, and each group may have competing and variable computing requirements.
In general, the techniques of this invention are directed to autonomic management of autonomic management systems. In particular, the embodiments of this invention use a measure, analyze, and respond model to autonomically manage one or more autonomic management systems. By understanding specific state information of these autonomic management systems, embodiments of the invention may achieve optimal performance for the autonomic management systems through operations monitoring, analyzing current system state against optimal state, and modifying the configurations or resources of the autonomic management systems.
For example, an autonomic management system may be a distributed computing system that conforms to a multi-level, hierarchical organizational model. One or more control nodes provide for the efficient and automated allocation and management of computing functions and resources within the distributed computing system in accordance with the organization model.
As described herein, the model includes four distinct levels: fabric, domains, tiers and nodes that provide for the logical abstraction and containment of the physical components as well as system and service application software of the enterprise. A user, such as a system administrator, interacts with the control nodes to logically define the hierarchical organization of the distributed computing system. The control nodes are responsible for all levels of management in accordance with the model, including fabric management, domain creation, tier creation and node allocation and deployment.
In one embodiment, a method comprises receiving monitoring data from a plurality of application nodes interconnected via a communication network. The method also includes receiving a failure notification from an autonomic management system when the autonomic management system perceives a failure of a first application node based on a fault tolerance parameter. In addition, the method includes analyzing the monitoring data from all of the application nodes to determine whether the perceived failure of the first application node is genuine. The method also comprises adjusting the fault tolerance parameter when the perceived failure of the first application is not genuine.
In another embodiment, a distributed computing system comprises a plurality of application nodes interconnected via a communications network. The system also comprises an autonomic management system to provide autonomic control of the application nodes, wherein the autonomic management system monitors the application nodes to perceive a failure of a first node of the application nodes based on a fault tolerance parameter. In addition, the system includes a system status manager to adjust the fault tolerance parameter based on an analysis of a state of all of the application nodes.
In another embodiment, a computer-readable medium comprises instructions. The instructions cause a processor to receive monitoring data from a plurality of application nodes interconnected via a communication network. The instructions also cause the processor to receive a failure notification from an autonomic management system when the autonomic management system perceives a failure of a first application node based on a fault tolerance parameter. In addition, the instructions also the processor to analyze the monitoring data from all of the application nodes to determine whether the perceived failure of the first application node is genuine. The instructions further cause the processor to adjust the fault tolerance parameter when the perceived failure of the first application is not genuine.
The details of one or more embodiments of the invention are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the invention will be apparent from the description and drawings, and from the claims.
In the illustrated example, the collection of computing nodes forming distributed computing system 10 are logically grouped within a discovered pool 11, a free pool 13, an allocated tiers 15 and a maintenance pool 17. In addition, distributed computing system 10 includes at least one control node 12.
Within distributed computing system 10, a computing node refers to the physical computing device. The number of computing nodes needed within distributed computing system 10 is dependent on the processing requirements. For example, distributed computing system 10 may include 8 to 512 computing nodes or more. Each computing node includes one or more programmable processors for executing software instructions stored on one or more computer-readable media.
Discovered pool 11 includes a set of discovered nodes that have been automatically “discovered” within distributed computing system 10 by control node 12. For example, control node 12 may monitor dynamic host communication protocol (DHCP) leases to discover the connection of a node to network 18. Once detected, control node 12 automatically inventories the attributes for the discovered node and reassigns the discovered node to free pool 13. The node attributes identified during the inventory process may include a CPU count, a CPU speed, an amount of memory (e.g., RAM), local disk characteristics or other computing resources. Control node 12 may also receive input identifying node attributes not detectable via the automatic inventory, such as whether the node includes I/O, such as HBA. Further details with respect to the automated discovery and inventory processes are described in U.S. patent application Ser. No. 11/070,851, entitled “AUTOMATED DISCOVERY AND INVENTORY OF NODES WITHIN AN AUTONOMIC DISTRIBUTED COMPUTING SYSTEM,” filed Mar. 2, 2005, the entire content of which is hereby incorporated by reference.
Free pool 13 includes a set of unallocated nodes that are available for use within distributed computing system 10. Control node 12 may dynamically reallocate an unallocated node from free pool 13 to allocated tiers 15 as an application node 14. For example, control node 12 may use unallocated nodes from free pool 13 to replace a failed application node 14 or to add an application node to allocated tiers 15 to increase processing capacity of distributed computing system 10.
In general, allocated tiers 15 include one or more tiers of application nodes 14 that are currently providing a computing environment for execution of user software applications. In addition, although not illustrated separately, application nodes 14 may include one or more input/output (I/O) nodes. Application nodes 14 typically have more substantial I/O capabilities than control node 12, and are typically configured with more computing resources (e.g., processors and memory). Maintenance pool 17 includes a set of nodes that either could not be inventoried or that failed and have been taken out of service from allocated tiers 15.
Control node 12 provides the system support functions for managing distributed computing system 10. More specifically, control node 12 manages the roles of each computing node within distributed computing system 10 and the execution of software applications within the distributed computing system. In general, distributed computing system 10 includes at least one control node 12, but may utilize additional control nodes to assist with the management functions.
Other control nodes 12 (not shown in
Network 18 provides a communications interconnect for control node 12 and application nodes 14, as well as discovered nodes, unallocated nodes and failed nodes. Communications network 18 permits internode communications among the computing nodes as the nodes perform interrelated operations and functions. Communications network 18 may comprise, for example, direct connections between one or more of the computing nodes, one or more customer networks maintained by an enterprise, local area networks (LANs), wide area networks (WANs) or a combination thereof. Communications network 18 may include a number of switches, routers, firewalls, load balancers, and the like.
In one embodiment, each of the computing nodes within distributed computing system 10 executes a common general-purpose operating system. One example of a general-purpose operating system is the Windows™ operating system provided by Microsoft Corporation. In some embodiments, the general-purpose operating system such as the Linux kernel may be used.
In the example of
System administrator 20 may interact with control node 12 and identify the particular types of software images to be associated with application nodes 14. Alternatively, administration software executing on control node 12 may automatically identify the appropriate software images to be deployed to application nodes 14 based on the input received from system administrator 20. For example, control node 12 may determine the type of software image to load onto an application node 14 based on the functions assigned to the node by system administrator 20. Application nodes 14 may be divided into a number of groups based on their assigned functionality. As one example, application nodes 14 may be divided into a first group to provide web server functions, a second group to provide business application functions and a third group to provide database functions. The application nodes 14 of each group may be associated with different software images.
Control node 12 provides for the efficient allocation and management of the various software images within distributed computing system 10. In some embodiments, control node 12 generates a “golden image” for each type of software image that may be deployed on one or more of application nodes 14. As described herein, the term “golden image” refers to a reference copy of a complete software stack for providing an execution environment for applications.
System administrator 20 may create a golden image by installing an operating system, middleware and software applications on a computing node and then making a complete copy of the installed software. In this manner, a golden image may be viewed as a “master copy” of the software image for a particular computing function. Control node 12 maintains a software image repository 26 that stores the golden images associated with distributed computing system 10.
Control node 12 may create a copy of a golden image, referred to as an “image instance,” for each possible image instance that may be deployed within distributed computing system 10 for a similar computing function. In other words, control node 12 pre-generates a set of K image instances for a golden image, where K represents the maximum number of image instances for which distributed computing system 10 is configured for the particular type of computing function. For a given computing function, control node 12 may create the complete set of image instance even if not all of the image instances will be initially deployed. Control node 12 creates different sets of image instances for different computing functions, and each set may have a different number of image instances depending on the maximum number of image instances that may be deployed for each set. Control node 12 stores the image instances within software image repository 26. Each image instance represents a collection of bits that may be deployed on an application node.
Further details of software image management are described in co-pending U.S. patent application Ser. No. 11/046,133, entitled “MANAGEMENT OF SOFTWARE IMAGES FOR COMPUTING NODES OF A DISTRIBUTED COMPUTING SYSTEM,” filed Jan. 28, 2005 and co-pending U.S. patent application Ser. No. 11/046,152, entitled “UPDATING SOFTWARE IMAGES ASSOCIATED WITH A DISTRIBUTED COMPUTING SYSTEM,” filed Jan. 28, 2005, each of which is incorporated herein by reference in its entirety.
In general, distributed computing system 10 conforms to a multi-level, hierarchical organizational model that includes four distinct levels: fabric, domains, tiers and nodes. Control node 12 is responsible for all levels of management, including fabric management, domain creation, tier creation and node allocation and deployment.
As used herein, the “fabric” level generally refers to the logical constructs that allow for definition, deployment, partitioning and management of distinct enterprise applications. In other words, fabric refers to the integrated set of hardware, system software and application software that can be “knitted” together to form a complete enterprise system. In general, the fabric level consists of two elements: fabric components or fabric payload. Control node 12 provides fabric management and fabric services as described herein.
In contrast, a “domain” is a logical abstraction for containment and management within the fabric. The domain provides a logical unit of fabric allocation that enables the fabric to be partitioned amongst multiple uses, e.g. different business services.
Domains are comprised of tiers, such as a 4-tier application model (web server, application server, business logic, persistence layer) or a single tier monolithic application. Fabric domains contain the free pool of devices available for assignment to tiers.
A tier is a logically associated group of fabric components within a domain that share a set of attributes: usage, availability model or business service mission. Tiers are used to define structure within a domain e.g. N-tier application, and each tier represents a different computing function. A user, such as administrator 20, typically defines the tier structure within a domain. The hierarchical architecture may provide a high degree of flexibility in mapping customer applications to logical models which run within the fabric environment. The tier is one construct in this modeling process and is the logical container of application resources.
The lowest level, the node level, includes the physical components of the fabric. This includes computing nodes that, as described above, provide operating environments for system applications and enterprise software applications. In addition, the node level may include network devices (e.g., Ethernet switches, load balancers and firewalls) used in creating the infrastructure of network 18. The node level may further include network storage nodes that are network connected to the fabric.
System administrator 20 accesses administration software executing on control node 12 to logically define the hierarchical organization of distributed computing system 10. For example, system administrator 20 may provide organizational data 21 to develop a model for the enterprise and logically define the enterprise fabric. System administrator 20 may, for instance, develop a model for the enterprise that includes a number of domains, tiers, and node slots hierarchically arranged within a single enterprise fabric.
More specifically, system administrator 20 defines one or more domains that each correspond to a single enterprise application or service, such as a customer relation management (CRM) service. System administrator 20 further defines one or more tiers within each domain that represent the functional subcomponents of applications and services provided by the domain. As an example, system administrator 20 may define a storefront domain within the enterprise fabric that includes a web tier, an application tier and a database tier. In this manner, distributed computing system 10 may be configured to automatically provide web server functions, business application functions and database functions.
For each of the tiers, control node 12 creates a number of “node slots” equal to the maximum number of application nodes 14 that may be deployed. In general, each node slot represents a data set that describes specific information for a corresponding node, such as software resources for a physical node that is assigned to the node slot. The node slots may, for instance, identify a particular software image instance associated with an application node 14 as well as a network address associated with that particular image instance.
In this manner, each of the tiers include one or more node slots that reference particular software image instances to boot on the application nodes 14 to which each software image instance is assigned. The application nodes 14 to which control node 12 assigns the image instances temporarily inherit the network address assigned to the image instance for as long as the image instance is deployed on that particular application node. If for some reason the image instance is moved to a different application node 14, control node 12A moves the network address to that new application node.
System administrator 20 may further define specific node requirements for each tier of the fabric. For example, the node requirements specified by system administrator 20 may include a central processing unit (CPU) count, a CPU speed, an amount of memory (e.g., RAM), local disk characteristics and other hardware characteristics that may be detected on the individual computing nodes. System administrator 20 may also specify user-defined hardware attributes of the computing nodes, such as whether I/O (like HBA) is required. The user-defined hardware attributes are typically not capable of detection during an automatic inventory. In this manner, system administrator 20 creates a list of attributes that the tier requires of its candidate computing nodes. In addition, particular node requirements may be defined for software image instances.
In addition to the node requirements described above, system administrator 20 may further define policies that are used when re-provisioning computing nodes within the fabric. System administrator 20 may define policies regarding tier characteristics, such as a minimum number of nodes a tier requires, an indication of whether or not a failed node is dynamically replaced by a node from free pool 13, a priority for each tier relative to other tiers, an indication of whether or not a tier allows nodes to be re-provisioned to other tiers to satisfy processing requirements by other tiers of a higher priority or other policies. Control node 12 uses the policy information input by system administrator 20 to re-provision computing nodes to meet tier processing capacity demands.
After receiving input from system administrator 20 defining the architecture and policy of the enterprise fabric, control node 12 identifies unallocated nodes within free pool 13 that satisfy required node attributes. Control node 12 automatically assigns unallocated nodes from free pool 13 to respective tier node slots of a tier. As will be described in detail herein, in one embodiment, control node 12 may assign computing nodes to the tiers in a “best fit” fashion. Particularly, control node 12 assigns computing nodes to the tier whose node attributes most closely match the node requirements of the tier as defined by administrator 20. The assignment of the computing nodes may occur on a tier-by-tier basis beginning with a tier with the highest priority and ending with a tier with the lowest priority. Alternatively, or in addition, assignment of computing nodes may be based on dependencies defined between tiers.
As will be described in detail below, control node 12 may automatically add unallocated nodes from free pool 13 to a tier when more processing capacity is needed within the tier, remove nodes from a tier to the free pool when the tier has excess capacity, transfer nodes from tier to tier to meet processing demands, or replace failed nodes with nodes from the free pool. Thus, computing resources, i.e., computing nodes, may be automatically shared between tiers and domains within the fabric based on user-defined policies to dynamically address high-processing demands, failures and other events.
In this example, e-commerce fabric 32 includes a storefront domain 34A and a financial planning domain 34B. Storefront domain 34A corresponds to the enterprise storefront domain and allows customers to find and purchase products over a network, such as the Internet. Financial planning domain 34B allows one or more employees to perform financial planning tasks for the enterprise.
Tier level 31C includes one or more tiers within each domain that represent the functional subcomponents of applications and services provided by the domain. For example, storefront domain 34A includes a web server tier (labeled “web tier”) 36A, a business application tier (labeled “app tier”) 36B, and a database tier (labeled “DB tier”) 36C. Web server tier 36A, business application tier 36B and database tier 36C interact with one another to present a customer with an online storefront application and services. For example, the customer may interact with web server tier 36A via a web browser. When the customer searches for a product, web server tier 36A may interacts with business application tier 36B, which may in turn access a database tier 36C. Similarly, financial planning domain 34B includes a financial planning tier 36D that provides subcomponents of applications and services of the financial planning domain 34B. Thus, in this example, a domain may include a single tier.
Tier level 31D includes one or more logical node slots 38A-38H (“node slots 38”) within each of the tiers. Each of node slots 38 include node specific information, such as software resources for an application node 14 that is assigned to a respective one of the node slots 38. Node slots 38 may, for instance, identify particular software image instances within image repository 26 and map the identified software image instances to respective application nodes 14. As an example, node slots 38A and 38B belonging to web server tier 36A may reference particular software image instances used to boot two application nodes 14 to provide web server functions. Similarly, the other node slots 38 may reference software image instances to provide business application functions, database functions, or financial application functions depending upon the tier to which the node slots are logically associated.
Although in the example of
In this example, organizational data 21 associates free node pool 13 with the highest-level of the model, i.e., e-commerce fabric 32. As described above, control node 12 may automatically assign unallocated nodes from free node pool 13 to at least a portion of tier node slots 38 of tiers 36 as needed using the “best fit” algorithm described above or another algorithm. Additionally, control node 12 may also add nodes from free pool 13 to a tier when more processing capacity is needed within the tier, remove nodes from a tier to free pool 13 when a tier has excess capacity, transfer nodes from tier to tier to meet processing demands, and replace failed nodes with nodes from the free tier.
Although not illustrated, the model for the enterprise fabric may include multiple free node pools. For example, the model may associate free node pools with individual domains at the domain level or with individual tier levels. In this manner, administrator 20 may define policies for the model such that unallocated computing nodes of free node pools associated with domains or tiers may only be used within the domain or tier to which they are assigned. In this manner, a portion of the computing nodes may be shared between domains of the entire fabric while other computing nodes may be restricted to particular domains or tiers.
During this process, control node 12 may receive input specifying node requirements of each of the tiers of the hierarchical model (52). As described above, administrator 20 may specify a list of attributes, e.g., a central processing unit (CPU) count, a CPU speed, an amount of memory (e.g., RAM), or local disk characteristics, that the tiers require of their candidate computing nodes. In addition, control node 12 may further receive user-defined custom attributes, such as requiring the node to have I/O, such as HBA connectivity. The node requirements or attributes defined by system administrator 20 may each include a name used to identify the characteristic, a data type (e.g., integer, long, float or string), and a weight to define the importance of the requirement.
Control node 12 identifies the attributes for all candidate computing nodes within free pool 13 or a lower priority tier (54). As described above, control node 12 may have already discovered the computing nodes and inventoried the candidate computing nodes to identify hardware characteristics of all candidate computing nodes. Additionally, control node 12 may receive input from system administrator 20 identifying specialized capabilities of one or more computing nodes that are not detectable by the inventory process.
Control node 12 dynamically assigns computing nodes to the node slots of each tier based on the node requirements specified for the tiers and the identified node attributes (56). Population of the node slots of the tier may be performed on a tier-by-tier basis beginning with the tier with the highest priority, i.e., the tier with the highest weight assigned to it. As will be described in detail, in one embodiment, control node 12 may populate the node slots of the tiers with the computing nodes that have attributes that most closely match the node requirements of the particular tiers. Thus, the computing nodes may be assigned using a “best fit” algorithm.
Next, control node 12 retrieves the node requirements associated with the selected tier (62). Control node 12 may, for example, maintain a database having entries for each node slot, where the entries identify the node requirements for each of the tiers. Control node 12 retrieves the node requirements for the selected tier from the database.
In addition, control node 12 accesses the database and retrieves the computing node attributes of one of the unallocated computing nodes of free pool 13. Control node 12 compares the node requirements of the tier to the node attributes of the selected computing node (64).
Based on the comparison, control node 12 determines whether the node attributes of the computing node meets the minimum node requirements of the tier (66). If the node attributes of the selected computing node do not meet the minimum node requirements of the tier, then the computing node is removed from the list of candidate nodes for this particular tier (68). Control node 12 repeats the process by retrieving the node attributes of another of the computing nodes of the free pool and compares the node requirements of the tier to the node attributes of the computing node.
If the node attributes of the selected computing node meet the minimum node requirements of the tier (YES of 66), control node 12 determines whether the node attributes are an exact match to the node requirements of the tier (70). If the node attributes of the selected computing node and the node requirements of the tier are a perfect match (YES of 70), the computing node is immediately assigned from the free pool to a node slot of the tier and the image instance for the slot is associated with the computing node for deployment (72).
Control node 12 then determines whether the node count for the tier is met (74). Control node 12 may, for example, determine whether the tier is assigned the minimum number of nodes necessary to provide adequate processing capabilities. In another example, control node. 12 may determine whether the tier is assigned the ideal number of nodes defined by system administrator 20. When the node count for the tier is met, control node 12 selects the next tier to enable, e.g., the tier with the next largest priority, and repeats the process until all defined tiers are enabled, i.e., populated with application nodes (60).
If the node attributes of the selected computing node and the node requirements of the tier are not a perfect match control node 12 calculates and records a “processing energy” of the node (76). As used herein, the term “processing energy” refers to a numerical representation of the difference between the node attributes of a selected node and the node requirements of the tier. A positive processing energy indicates the node attributes more than satisfy the node requirements of the tier. The magnitude of the processing energy represents the degree to which the node requirements exceed the tier requirements.
After computing and recording the processing energy of the nodes, control node 12 determines whether there are more candidate nodes in free pool 13 (78). If there are additional candidate nodes, control node 12 repeats the process by retrieving the computing node attributes of another one of the computing nodes of the free pool of computing nodes and comparing the node requirements of the tier to the node attributes of the computing node (64).
When all of the candidate computing nodes in the free pool have been examined, control node 12 selects the candidate computing node having the minimum positive processing energy and assigns the selected computing node to a node slot of the tier (80). Control node 12 determines whether the minimum node count for the tier is met (82). If the minimum node count for the tier has not been met, control node 12 assigns the computing node with the next lowest calculated processing energy to the tier (80). Control node 12 repeats this process until the node count is met. At this point, control node 12 selects the next tier to enable, e.g., the tier with the next largest priority (60).
In the event there are an insufficient number of computing nodes in free pool 13, or an insufficient number of computing nodes that meet the tier requirements, control node 12 notifies system administrator 20. System administrator 20 may add more nodes to free pool 13, add more capable nodes to the free pool, reduce the node requirements of the tier so more of the unallocated nodes meet the requirements, or reduce the configured minimum node counts for the tiers.
Control node 12 then determines whether there are any computing nodes in the free pool of nodes that meet the minimum node requirements of the tier (92). When there are one or more nodes that meet the minimum node requirements of the tier, control node 12 selects the node from the free pool based the node requirements of the tier, as described above, (94) and assigns the node to the tier (95). As described in detail with respect to
If there are no adequate computing nodes in the free pool, i.e., no nodes at all or no nodes that match the minimal node requirements of the tier, control node 12 identifies the tiers with a lower priority than the tier needing more processing capacity (96).
Control node 12 determines which of the nodes of the lower priority tiers meet the minimum requirements of the tier in need of processing capacity (98). Control node 12 may, for example, compare the attributes of each of the nodes assigned to node slots of the lower priority tiers to the node requirements of the tier in need of processing capacity. Lower priority tiers that have the minimum number of computing nodes may be removed from possible tiers from which to harvest an application node. If, however, all the lower priority tiers have the minimum number of computing nodes defined for the respective tier, the lowest priority tier is selected from which to harvest the one or more nodes.
Control node 12 calculates the processing energy of each of the nodes of the lower priority tiers that meet the minimum requirements (100). The energies of the nodes are calculated using the differences between the node attributes and the node requirements of the tier needing additional capacity. Control node 12 selects the computing node with the lowest processing energy that meets the minimum requirements, and assigns the selected computing node to the tier in need of processing capacity (102, 95).
When harvesting a node, control node 12 calculates the processing energy of all the nodes in the tier as described above with reference to
Returning the node to the free pool may involve remotely powering off the computing node and updating the database to associate the harvested node with free pool 13. In addition, control node 12 updates the database to disassociate the returned node with the node slot to which it was assigned. At this point, the node no longer uses the network address associated with the image instance mapped to the node slot. Control node 12 may, therefore, assign a temporary network address to the node while the node is assigned to free pool 13.
When configuring a tier, system administrator 20 may elect to activate email alerts. For example, system administrator 20 may activate the email alerts feature in order to receive email alerts providing system administrator 20 with critical and/or non-critical tier information, such as a notification that a tier has been upgraded, a node of the tier has failed or the like. Furthermore, system administrator 20 may input various policies, such node failure rules. For example, system administrator 20 may identify whether control node 12 should reboot a node in case of failure or whether the failed node should automatically be moved to maintenance pool 17. Similarly, system administrator 20 may identify whether nodes assigned to the tier may be harvested by other tiers.
User interface 130 may also allow system administrator 20 to input node requirements of a tier. In order to input node requirements of a tier, system administrator 20 may click on the “Requirements” tab 132, causing user interface 130 to present an input area to particular node requirements of the tier.
User interface 140 also displays properties that control node 12 has identified during the computing node inventory process. In this example, user interface 140 presents system administrator 20 with the a CPU node count, a CPU speed, the amount of RAM, the disk size and other characteristics that are identifiable during the automated node inventory. User interface 140 additionally presents interface information to system administrator 20. Specifically, user interface 140 provides system administrator 20 with a list of components and their associated IP and MAC addresses.
User interface 140 also allows system administrator 20 to define other custom requirements. For example, system administrator 20 may define one or more attributes and add those attributes to the list of node attributes presented to system administrator 20.
System administrator 20 or another user may select an image by clicking on the box in front of the image identifier/name and perform one or more actions on the image. Actions that system administrator 20 may perform on an image include deleting the image, updating the image, and the like. System administrator 20 may select one of the image actions via dropdown menu 152. In some embodiments, user interface 150 may further display other details about the images such as the node to which the images are assigned (if the node status is “active”), the network address associated with the images and the like.
For each node, user interface 160 presents a node ID, a status of the node, the tier to which the node belongs, a hostname associated with the node, a NIC IP address, a rack location, a slot location, the number of CPU's of the node, the amount of RAM on the node, the number of disks on the node, whether the node has I/O HBA, and the number of NICs of the node.
System administrator 20 or other user may select a node by clicking on the box in front of the node identifier/name and perform one or more actions on the node. Actions that system administrator 20 may perform on the node include deleting the node, updating the node attributes or other properties of the node, and the like. System administrator 20 may select one of the node actions via dropdown menu 162.
System administrator 20 or other user may select an alert by clicking on the box in front of the logged alert and perform one or more actions on the logged alert. Actions that system administrator 20 may perform include deleting the alert, changing the status of the alert, or the like. System administrator 20 may specify the log actions via dropdown menu 192.
Monitoring subsystem 202 provides real-time monitoring of the distributed computing system 10. In particular, monitoring subsystem 202 dynamically collects status data 203 from the hardware and software operating within distributed computing system 10, and feeds the status data in the form of monitor inputs 208 to SLAI 204. Monitoring inputs 208 may be viewed as representing the actual state of the fabric defined for the organizational model implemented by distributed computing system 10. Monitoring subsystem 202 may utilize well defined interfaces, e.g., the Simple Network Management Protocol (SNMP) and the Java Management Extensions (JMX), to collect and export real-time monitoring information to SLAI 204.
SLAI 204 may be viewed as an automation subsystem that provides support for autonomic computing and acts as a central nervous system for the controlled fabric. In general, SLAI 204 receives monitoring inputs 208 from monitoring subsystem 202, analyzes the inputs and outputs appropriate action requests 212 to BLT 206. In one embodiment, SLAI 204 is a cybernetic system that controls the defined fabric via feedback loops. More specifically, administrator 20 may interact with BLT 206 to define an expected state 210 for the fabric. BLT 206 communicates expected state 210 to SLAI 204. SLAI 204 receives the monitoring inputs from monitoring subsystem 202 and applies rules to determine the most effective way of reducing the differences between the expected and actual states for the fabric.
For example, SLAI 204 may apply a rule to determine that a node within a high priority tier has failed and that the node should be replaced by harvesting a node from a lower priority tier. In this example, SLAI 204 outputs an action request 212 to invoke BLT 206 to move a node from one tier to the other.
In general, BLT 206 implements high-level business operations on fabrics, domains and tiers. SLAI 204 invokes BLT 206 to bring the actual state of the fabric into accordance with the expected state. In particular, BLT 206 outputs fabric actions 207 to perform the physical fabric changes. In addition, BLT 206 outputs an initial expected state 210 to SLAI 204 and initial monitoring information 214 to SLAI 204 and monitoring subsystem 202, respectively. In addition, BLT 206 outputs notifications 211 to SLAI 204 and monitoring subsystem 202 to indicate the state and monitoring changes to distributed computing system 10. As one example, BLT 206 may provide control operations that can be used to replace failed nodes. For example, BLT 206 may output an action request indicating that a node having address 10.10.10.10 has been removed from tier ABC and a node having address 10.10.10.11 has been added to tier XYZ. In response, monitoring subsystem 202 stops attempting to collect status data 203 from node 10.10.10.10 and starts monitoring for status data from node 10.10.10.11. In addition, SLAI 204 updates an internal model to automatically associate monitoring inputs from node 10.10.10.11 with tier XYZ.
As described further below, monitoring subsystem 202 employs a modular architecture that allows new detection and monitoring collectors 224 to be “plugged-in” for existing and new protocols and for existing and new hardware and software. As illustrated in
In one embodiment, collectors 224 are loaded at startup of control node 12 and are configured with information retrieved from BLT 206. Monitoring engine 222 receives collection requests from SLAI 204, sorts and prioritizes the requests, and invokes the appropriate one of collectors 224 based on the protocol specified in the collection requests. The invoked collector is responsible for collecting the required status data and returning the status data to monitoring engine 222. If the collector is unable to collect the requested status data, the collector returns an error code.
In one embodiment, collectors 224 are Java code compiled into ajar file and loaded with a class loader at run time. Each of collectors 224 has an associated configuration file written in a data description language, such as the extensible markup language (XML). In addition, a user may interact with BLT 206 to add run-time configuration to dynamically configure collectors 224 for specific computing environments. Each of collectors 224 expose an application programming interface (API) to monitoring engine 222 for communication and data exchange.
A user, such as a system administrator, specifies the protocol or protocols to be used for monitoring a software image when the image is created. In addition, the users may specify the protocols to be used for monitoring the nodes and each service executing on the nodes. Example protocols supported by the collectors 224 include Secure Shell (SSH), Simple Network Management Protocol (SNMP), Internet Control Message Protocol (ICMP) ping, Java Management Extensions (JMX) and the Hypertext Transfer Protocol (HTTP).
Some protocols require special privileges, e.g., root privileges, to perform the required data collection. In this case, the corresponding collectors 224 communicate with a separate process that executes as the root. Moreover, some protocols may require deployment and/or configuration of data providers within the fabric. Software agents may, for example, be installed and configured on nodes and configured on other hardware. If needed, custom in-fabric components may be deployed.
In this example, the modular architecture of monitoring subsystem 202 also supports one or more plug-in interfaces 220 for data collection from a wide range of third-party monitoring systems 228. Third-party monitoring systems 228 monitor portions of the fabric and may be vendor-specific.
In general, sensor subsystem 240 receives actual state data from monitoring subsystem 202 in the form of monitoring inputs 208 and supplies ongoing, dynamic input data to analysis subsystem 244. For example, sensor subsystem 240 is notified of physical changes to distributed computing system 10 by monitoring subsystem 202. Sensor subsystem 240 uses the state data received from monitoring subsystem 202 to maintain ongoing, calculated values that can be sent to analysis subsystem 244 in accordance with scheduler 242.
In one embodiment, sensor subsystem 240 performs time-based hierarchical data aggregation of the actual state data in accordance with the defined organization model. Sensor subsystem 240 maintains organizational data in a tree-like structure that reflects the current configuration of the hierarchical organization model. Sensor subsystem 240 uses the organizational data to perform the real-time data aggregation and map tiers and domains to specific nodes. Sensor subsystem 240 maintains the organizational data based on notifications 211 received from BLT 206.
Sensor subsystem 240 sends inputs to analysis subsystem 244 to communicate the aggregated data on a periodic or event-driven basis. Analysis subsystem 244 may register an interest in a particular aggregated data value with sensor subsystem 240 and request updates at a specified frequency. In response, sensor subsystem 240 interacts with monitoring subsystem 202 and scheduler 242 to generate the aggregated data required by analysis subsystem 244.
Sensor subsystem 240 performs arbitrary data aggregations via instances of plug-in classes (referred to as “triggers”) that define the aggregations. Each trigger is registered under a compound name based on the entity being monitored and the type of data being gathered. For example, a trigger may be defined to aggregate and compute an average computing load for a tier every five minutes. Analysis subsystem 244 requests the aggregated data based on the registered names. In some embodiments, analysis subsystem 244 may define calculations directly and pass them to sensor subsystem 240 dynamically.
Analysis subsystem 244 is composed of a plurality of forward chaining rule engines 246A-246N. In general, rule engines 246 match patterns in a combination of configuration data and monitoring data, which is presented by extraction agent 251 in the form of events. Events contain the aggregated data values that are sent to rule engines 246 in accordance with scheduler 242.
Sensor subsystem 240 may interact with analysis subsystem 244 via trigger listeners 247 that receives updates from a trigger within sensor subsystem 240 when specified events occur. An event may be based on system state (e.g., a node transitioning to an up or down state) or may be time based.
Analysis subsystem 244 allows rule sets to be loaded in source form and compiled at load time into discrimination networks. Each rule set specifies trigger-delivered attributes. Upon loading the rule sets, analysis subsystem 244 establishes trigger listeners 247 to receive sensor notifications and update respective working memories of rule engines 246. As illustrated in
Rule engines 246 process the events and invoke action requests via calls to effector subsystem 248. In addition, rule engines 246 provide a call-back interface so that effector subsystem 248 can inform a rule engine when an action has completed. Rule engines 246 prevent a particular rule from re-firing as long as any action invoked by the rule has not finished. In general, rules contain notification calls and service invocations though either may be disabled by configuration of effector subsystem 248. BLT 206 supplies initial system configuration descriptions to seed each of rule engines 246.
In general, rule engines 246 analyze the events and discover discrepancies between an expected state of the fabric and an actual state. Each of rule engines 246 may be viewed as software that performs logical reasoning using knowledge encoded in high-level condition-action rules. Each of rule engines 246 applies automated reasoning that works forward from preconditions to goals defined by system administrator 20. For example, rule engines 246 may apply modus ponens inferences rules.
Rule engines 246 output requests to effector subsystem 248 which produce actions requests 212 for BLT 206 to resolve the discrepancies. Effector subsystem 248 performs all operations on behalf of analysis subsystem 244. For example, event generator 250, task invocation module 252 and logger 254 of effector subsystem 248 perform event generation, BLT action invocation and rule logging, respectively. More specifically, task invocation module 252 invokes asynchronous operations within BLT 206. In response, BLT 206 creates a new thread of control for each task which is tracked by a unique task identifier (task id). Rules engine 246 uses the task id to determine when a task completes and, if needed, to re-fire any rules that were pended until completion of the task. These tasks may take arbitrary amounts of time, and rules engine 246 tracks the progress of individual task via change notifications 211 produced by BLT 206.
Event generator 250 creates persistent event records of the state of processing of SLAI 204 and stores the event records within a database. Clients uses these event records to track progress and determine the current state of the SLAI 204.
Logger 254 generates detailed trace information about system activities for use in rule development and debugging. The logging level can be raised or lowered as needed without changing operation of SLAI 204.
In addition, working memory 270 includes a second data region 274 that is modifiable (i.e., read/write) and may be updated by monitoring subsystem 202 or used internally by rule engines 246. In general, data region 274 stores aggregated data representing the actual state of the fabric and can be updated by sensor subsystem 240 or by rule engines 246. The actual state may consist of a set of property annotations that can be attached to objects received from BLT 206 or to objects locally defined within a rule engine, such as local object 276.
As described, BLT 206 provides the facilities necessary to create and administer the organizational model (e.g., fabric, domains, tiers and nodes) implemented by distributed computing system 10. In general, BLT 206 abstracts access to the persisted configuration state of the fabric, and controls the interactions with interfaces to fabric hardware services. As such, BLT 206 provides fabric management capabilities, such as the ability to create a tier and replace a failed node. WSDL interfaces 300 provide web service interfaces to the functionality of BLT 206 that may be invoked by web service clients 313. Many of WSDL interfaces 300 offered by BLT 206 allow administrator 20 to define goals, such as specifying a goal of the expected state of the fabric. As further described below, rule engines 246 within SLAI 204, in turn, invoke task manger 312 to initiate one or more BLT tasks to achieve the specified goal. In general, web service clients 313 may be presentation layer applications, command line applications, or other clients.
BLT 206 abstracts all interaction with physical hardware for web service clients 313. BLT 206 is an enabling component for autonomic management behavior, but does not respond to real-time events that either prevent a goal from being achieved or produce a set of deviations between the expected state and the actual state of the system. In contrast, BLT 206 originates goals for autonomic reactions to changing configuration and state. SLAI 204 analyzes and acts upon these goals along with real-time state changes. BLT 206 sets the goals to which SLAI 204 strives to achieve, and provides functionality used by the SLAI in order to achieve the goals.
In general, BLT 206 does not dictate the steps taken in pursuit of a goal since these are likely to change based on the current state of distributed computing system 10 and changes to configurable policy. SLAI 204 makes these decisions based on the configured rule sets for the fabric and by evaluating monitoring data received from monitoring subsystem 202.
Fabric administration service 304 implements a set of methods for managing all aspects of the fabric. Example methods include methods for adding, viewing, updating and removing domains, tiers, nodes, notifications, assets, applications, software images, connectors, and monitors. Other example methods include controlling power at a node, and cloning, capturing, importing, exporting or upgrading software images. Rule engines 246 of SLAI 204 may, for example, invoke these methods by issuing action requests 212.
Task manager 312 receives action requests 212 via task interface 311. In general, task interface 311 provides an interface for receiving action requests 212 from SLAI 204 or other internal subsystem. In response, task manager 312 manages asynchronous and long running actions that are invoked by SLAI 204 to satisfy a goal or perform an action requested by a client.
Task manager 312 generates task data 310 that represents identification and status for each task. Task manager 312 returns a task identifier to the calling web service clients 313 or the internal subsystem, e.g., SLAI 204, that initiated the task. Rule engines 246 and web service clients 313 use the task identifiers to track progress and retrieve output, results, and errors associated with achieving the goal.
In one embodiment, there are no WSDL interfaces 300 for initiating specific tasks. Rather, administrator 20 interacts with BLT 206 though goal interfaces presented by WSDL interfaces 300 to define the goals for the fabric. In contrast, the term task is used to refer to internal system constructs that require no user interaction. Tasks are distinct, low-level units of work that affect the state of the fabric. SLAI 204 may combine tasks to achieve or maintain a goal state.
For example, administrator 20 can request configuration changes by either adding new goals to an object or by modifying the attributes on existing goals. Scheduled goals apply a configuration at a designated time. For example, the goals for a particular tier may specify the minimum, maximum, and target node counts for that tier. As a result, the tier can increase or decrease current node capacity by scheduling goals with different configuration values.
This may be useful, for example, in scheduling a software image upgrade. As another example, entire domains may transition online and offline per a defined grid schedule. Administrator 20 may mix and match goals on a component to achieve configurations specific to the application and environment. For example, a tier that does not support autonomic node replacement would not be configured with a harvesting goal.
In some embodiments, goals are either “in force” or “out of force.” SLAI 204 only works to achieve and maintain those goals that are currently in force. SLAI 204 may applies a concept of “gravity” as the goals transition from in force to out of force. For example, SLAI 204 may transition a tier offline when an online goal is marked out of force. Some goal types may have prerequisite goals. For example, an image upgrade goal may require as a prerequisite that a tier be transitioned to offline before the image upgrade can be performed. In other embodiments, goals are always in force until modified.
SLAI 204 may automatically formulate dependencies between goals or may allow a user to specify the dependencies. For example, a user may request that a newly created tier come online. As a result of this goal, SLAI 204 may automatically direct task manager 312 to generate a task of harvesting a target number of nodes to enable the tier. Generally, all goals remain in-force by SLAI 204 until modified by BLT 206. In one embodiment, each goal remains in-force in one of three states: Satisfied, Warning, or Critical depending on how successful SLAI 204 was in achieving the goal at the time the event record was generated and stored.
In this manner, SLAI 204 controls the life cycle of a goal (i.e., the creation, scheduling, update, deletion of the goal), and provides a common implementation of these and other services such as timeout, event writing, goal conflicts, management of intra-goal dependencies, and tracking tasks to achieving the goals.
Progress toward a goal is tracked though event subsystem 315. In particular, event subsystem 315 tracks the progress of each in force goal based on the goal identifiers. Tasks executed to achieve a particular goal produce events to communicate result or errors. The events provide a convenient time-based view of all actions and behaviors.
Examples of goal types that may be defined by administrator 20 include software image management goals, node allocation goals, harvest goals, tier capacity goals, asset requirement goals, tier online/offline goals, and data gathering goals.
In one embodiment, BLT 206 presents a task interface to SLAI 204 for the creation and management of specific tasks in order to achieve the currently in force goals. In particular, rule engines 246 invoke the task interface based on evaluation of the defined rule sets in view of the expected state and actual state for the fabric. Example task interfaces include interfaces to: reserve node resources; query resources for a node slot; associate or disassociate an image with a node in a tier node slot; allocate, de-allocate, startup or shutdown a node; move a node to a tier; apply, remove or cycle power of a node; create a golden image; create or delete an image instance; and delete an activity, node or tier.
Report generator 302 provides an extensible mechanism for generating reports 314. Typical reports include image utilization reports that contain information with respect to the number of nodes running each software image, inventory reports detailing both the logical and physical aspects of the fabric, and system event reports showing all events that have occurred within the fabric. Report generator 302 gathers, localizes, formats and displays data into report form for presentation to the user. Report generator 302 may include one or more data gathering modules (not shown) that gather events in accordance with a schedule and update an events table to record the events. The data gathering modules may write the events in XML format.
When translated, this example rule marks a tier as overloaded if an application is implemented by the tier and the maximum specified load for the application has been exceeded. Another example rule for outputting a notification that a tier is overloaded and automatically invoking a task within BLT 206 to add a node is:
Rule compiler 344 compiles each of rules 344 and translates match conditions of the rules into a discrimination network that avoids redundant tests during rule execution. Execution engine 346 handles rule administration, object insertion and retrieval, rule invocation and execution of rule actions. In general, execution engine 346 first matches a current set of rules 342 against a current state of working memory 348 and local objects 350. Execution engine 346 then collects all rules that match as well as the matched objects and selects a particular rule instantiation to fire. Next, execution engine 346 fires (executes) the instantiated rule and propagates any changes to working memory 348. Execution engine 346 repeats the process until no more matching rule instantiations can be found.
Firing of a rule typically produces a very small number of changes to working memory 348. This allows sophisticated rule engines to scale by retaining match state between cycles. Only the rules and rule instantiations affected by changes are updated, thereby avoiding the bulk of the matching process. One exemplary algorithm that may be used by execution engine 346 to handle the matching process includes the RETE algorithm that creates a decision tree that combines the patterns in all the rules and is intended to improve the speed of forward-chained rule system by limiting the effort required to re-compute a conflict set after a rule is fired. One example of a RETE algorithm is described in Forgy, C. L.: 1982, ‘RETE: a fast algorithm for the many pattern/many object pattern match problem’, Artificial Intelligence 19, 1737, hereby incorporated by reference. Other alternatives include the TREAT algorithms, and LEAPS algorithm, as described by Miranker, D. P.: ‘TREAT: A New and Efficient Match Algorithm for AI Production Systems’. ISBN 0934613710 Daniel P. Miranker, David A. Brant, Bernie Lofaso, David Gadbois: On the Performance of Lazy Matching in Production Systems. AAAI 1990: 685692, each of which is hereby incorporated by reference.
Application matrix 350 contains all the information needed by control unit 12 to interact with one or more applications or application servers and provide autonomic control over a set of applications. Specifically, application matrix 350 provides a logical definition for deploying and controlling the set of applications to one or more tiers within distributed computing system 10. In one embodiment, application matrix 350 is an electronic document that conforms to a data description language, e.g., the extensible markup language (XML). Application SLAI 358 includes an application rules engine 355 dedicated to processing application-level rules, i.e., forward-chaining rules to provide autonomic control over the applications defined within application matrix 350. Like rules engine 246, application rules engine 355 contains a rule compiler, an execution engine, and a working memory. In order to give effect to changes in application matrix 350, application SLAI 358 automatically updates application rules engine 355 and monitoring subsystem 202. In particular, application matrix 350 sends an alert whenever application matrix 350 changes. In response to this alert, application SLAI 358 captures application-specific attributes from application matrix 350. Specifically, application SLAI 358 captures configuration attributes and rule attributes contained in application matrix 350. Application SLAI 358 transfers any new rule attributes to the working memory of application rules engine 355 to provide autonomic control over the deployment, monitoring and the execution of the applications defined within application matrix 350. In addition, application SLAI 358 updates monitoring subsystem 202 to collect information required to control the applications In this manner, administrator 20 may continue to add new application definitions and configurations to application matrix 350 after distributed control node 12 has started.
As described in further detail below, configuration processor 354 is a software module that generates an application matrix entry based on an application definition and application configuration properties of a “staged” application. A staged application is an application that has been deployed in a staging environment and customized for subsequent deployment within distributed computing system 10. After creating the application matrix entry, administrator 20 may insert the application matrix entry into application matrix 350.
Configuration processor 354 is “pluggable.” That is, administrator 20 can “plug in” different implementations of configuration processor 354 as needed. For example, administrator 20 may need to “plug in” a different implementation of configuration processor 354 to handle applications that do not use an application server.
Application governor 352 is a software engine that performs application-level actions based on requests received from application rules engine 355. In this manner, BLT 206 effects fabric-level actions (e.g., deployment and monitoring of nodes and images) based on request from fabric-level rules engines 246 (
Application governor 352 uses application matrix 350 as a source of parameters when carrying out the application-level operations requested by application rules engine 355. For example, application rules engine 355 may detect that a node to which the application is not deployed is ready for use by the application. As a result, application rules engine 355 directs application governor 352 to handle the details of deploying the application to the node. In turn, application governor 352 accesses application matrix 350 to retrieve application-specific parameters necessary to deploy the application. Storing application-specific parameters in application matrix 350 allows the application-specific parameters to change without having to recompile the rules within working memory of application rules engine 355.
Application governor 352 performs a similar procedure to undeploy an application. That is, application rules engine 355 may detect that a second application needs to use a node more than a first application that is currently deployed to the node. In this situation, application rules engine 355 sends an instruction to application governor 352 to undeploy the first application and deploy the second application. To carry out this instruction, application governor 352 accesses application matrix 350 to discover configuration parameters of both applications. Application governor 352 then uses the discovered configuration parameters to communicate with the applications.
Like configuration processor 354, application governor 352 is also “pluggable.” That is, administrator 20 can easily install or remove implementations of application governor 352 depending on the circumstances. Because application governor 352, and configuration processor 354 represent interchangeable, plug-in modules, the other parts of control unit 12 and system 10, including application rules engine 355, can remain generic and application neutral while providing autonomic control over distributed computing system 10.
In this example, application matrix 350 contains seven columns and two rows. Each row represents a different application entry for deployment within distributed computing system 10. In
Each of the columns represents a different category of elements that make up an application's logical definition. In this example, the columns include:
(1) an application column 360 that includes elements generally related to the deployment of the application,
(2) an application nodes column 362 that contains elements related to the tier node slots to which the application may be assigned,
(3) a services column 364 that contains elements related to the executable services launched when the application is deployed,
(4) a node monitor values column 366 that contains elements related to attributes of the nodes that are to be monitored after the application is deployed to that node,
(5) a service monitored attributes column 368 that contains elements related to attributes of the services that are to be monitored after the application is deployed,
(6) a service levels column 370 that contains elements related to attributes for use when constructing rules to monitor execution of the services, and
(7) a deployment constraints column 372 that contains elements related to attributes for use when constructing rules to control deployment of the application. Different types of applications may have different elements in each column, and different numbers of columns.
In the example of
In addition to these attributes, application 374 contains a series of elements (columns 362-372). In general, application nodes 362 contain a list of all of the available tier-node slots to which control node 12 may deploy application 374. In this example, two tier-node slots are specified. In XML, managed servers 362 appears as:
Each services 364 element identifies a service that control node 12 launches when deploying application 374 on a given application node. A service element comprises a service name and a path to a file containing the executable service. Governor 352 uses the path to locate and launch the service. For example, the following XML code indicates that governor 352 must access the file at “/lib/worklistApp/worklistApp.ear”.
Node Monitored Values 366 elements represent characteristics for use in constructing rules for monitoring nodes to which the applications are deployed. Similarly, Service Monitored Values 368 elements represent characteristics for use in constructing rules for monitoring services that are launched once the application is deployed. In this example, a “nodeMonitoredValues” element defines characteristics of a particular node that are to be monitored. For instance, the amount of free memory in a node is one example of a characteristic listed as a “nodeMonitoredValues” element. On the other hand, a “serviceMonitoredValues” element is specific attribute of a service that is to be monitored. For example, the number of pending operating system-level requests for the service, the number of idle threads, etc., could be service monitored values. In a XML rendition of application matrix 350, node monitored values and service monitored values could appear as follows:
Deployment constraint elements 372 specify characteristics of a node under which application rules engine 355 should (or should not) deploy a service to the node. In this example, a deployment constraint element has five attributes: “attribute”, “expression”, “frequency”, “maxThreshold”, “minThreshold”, and “period.” The “attribute” attribute names the deployment constraint. The “expression” attribute specifies an arithmetic expression manipulating a monitored value. For example, the expression could be “PercentMemoryFree*100”, meaning monitor the value of the “PercentMemoryFree” node monitored value multiplied by 100. The “expression” attribute may specify one or more node monitored values. The “frequency” attribute informs application rules engine 355 how frequently to check the monitored value. The “maxThreshold” attribute tells application rules engine 355 to invoke a rule when the value of the expression exceeds the value specified by the “maxThreshold” attribute. Similarly, the “minThreshold” attribute tells application rules engine 355 to invoke the rule when the value of the expression drops below the value specified by the “minThreshold” attribute. Finally, the “period” attribute informs application rules engine 355 of the period over which to collect monitored value. For example, a deployment constraint element may specify that application rules engine 355 should monitor the PercentMemoryFree attribute of a node every 15 seconds for 60 seconds. If the value of PercentMemoryFree*100 should drop below 1.0 (i.e. 1% of memory free) for 60 seconds, then application rules engine 355 should not deploy the application to that node. In XML, this rule would be represented as:
Service level elements 370 have the same five attributes as deployment constraints elements: “attribute”, “expression”, “frequency”, “maxThreshold”, “minThreshold”, and “period”. However, the “expression” attribute deals with service monitored attributes rather than node monitored attributes. For example, a service level element may specify that application rules engine 355 check the “pendingRequest” service-monitored attribute every 15 seconds for 30 seconds. Then, if there are more than 20 pending requests for more than 30 seconds, application rules engine 355 should take the action of starting a new application. On the other hand, if there are fewer than 5 pending requests for 30 seconds, application rules engine 355 enables the action to remove the application to free up space on a node. Such a service level element could be represented as:
Put together, an XML representation of an application matrix logically defining a single application for deployment within autonomically controlled distributed computing system 10 may appear as follows:
During this process, administrator 20 directs the application to generate an application definition file that represents an application-specific configuration based on the environment specified by the administrator. The application definition file contains information specific to an installation instance of the application. For example, the application definition file typically specifies the locations (paths) for all software components of the application as staged within the staging environment, a list of files to execute, connectivity information for the application, and other information that may be recorded by the configured application.
After staging the application, administrator 20 defines a set of application configuration properties that specify desired behavior of the application within distributed computing system 10 (382). Examples of such application configuration properties include a minimum number of nodes, minimum resource requirements for a node, deployment timing information, network addresses of tier node slots that are able to execute the application and other properties.
Next, administrator 20 directs configuration processor 354 to generate an application entry for the application using the application definition and the application configuration properties (384). The application entry contains configuration attributes used by application governor 352 to interact with the application. In addition, the application entry contains rule attributes used by application rules engine 355 that define how control node 12 monitors the deployment and execution of the application within distributed computing environment 10.
After configuration processor 354 creates the application entry, administrator 20 may modify the application entry (386). In particular, administrator 20 may update start scripts or shell scripts for the application due to path changes between the staging environment and distributed computing system 10.
Once administrator 20 has finished modifying the application entry, administrator 20 inserts the application entry into application matrix 350 (388). Because application matrix 350 detects that it has changed, application matrix 350 sends an alert to application SLAI 358 (390).
In response to the alert, application SLAI 358 may update application rules engine 355 and monitoring subsystem 202 (392). In particular, application SLAI 358 automatically scans application matrix 350. If application SLAI 358 detects new rule attributes, application SLAI 358 creates local objects reflecting the new rule attributes in the working memory of application rules engine 355. In addition, if application SLAI 358 detects new monitored values, application SLAI 358 updates monitoring subsystem 202 to add new monitoring collectors 224 (
After application SLAI 358 updates application rules engine 355 and monitoring subsystem 202, control node 12 has autonomic control over the deployment of applications to tiers based on the configuration of application matrix 350 (394). For example, control node 12 deploys images, deploys applications, monitors the state of nodes and the execution of the applications, and applies tier-level rules as well as application-specific rules. Application SLAI 358 continues to listen for alerts from application matrix 350.
In some instances, autonomic management systems may have difficulty distinguishing between genuine failures and perceived failures of application nodes 408. For instance, an autonomic management system may falsely perceive a failure of an application node when the application node is experiencing uncommon load or decreased system response time. Because of the difficulty in distinguishing genuine from falsely perceived failures, autonomic management systems may respond aggressively to any perceived failure. For example, a slow network response time could cause an autonomic management system to think that a node has failed. The difficulty in distinguishing between actual and perceived failures may arise from the fact that many autonomic management systems evaluate status of a single entity, such as a node or a switch, and determine normal or failure state based on information from just that entity. Evaluating status based on information from a single application node may lead to erroneous decisions because factors reported by other entities may affect the status information of that entity.
In autonomic monitoring system 400, control node 406 generates a failure notification when the control node perceives a failure or condition based on one or more fault tolerance parameters. When a system status manager 402 receives the failure notification, system status manager 402 considers a state of all of application nodes 408 when determining actual versus false positive failures. System status manager 402 assesses monitoring data from other services related to the failing component before deciding on the component's health and status. For example, system status manager 402 may emulate the actions of a human system administrator by continuously evaluating monitors on all components of autonomically monitored system 400 and making inferences based on the state of autonomically monitored system 400 as a whole.
System status manager 402 uses the initial status of the single underperforming node as a starting point to begin collecting additional health information from other components to determine the root cause of a degraded system. Multiple system status managers may work in parallel to evaluate different potential fault paths within the environment to diagnose the problem. The different potential fault paths may be organized in a backward chaining mode.
System status manager 402 makes health and status decisions for individual components based on historical information as well as pre-programmed responses to known failure scenarios. In the case of periodically polled monitoring data, when system status manager 402 recognizes a failure of a single application node, system status manager 402 evaluates the state of other application nodes in autonomically monitor system 400 that could affect the perceived state of the failing application node. If system status manager 402 determines the root cause of the failure is not a genuine failure, but a service issue, system status manager 402 may increase a fault tolerance parameter for the application node based on the analysis of all the components in autonomically monitored system 400.
For example, system status manager 402 may increase the time between attempts to connect to the failing component to collect data if the network traffic monitor indicates that utilization of the component is rising. In such situations monitoring may decrease the application node's use of degraded system resources. This change in retry interval would result in a reduction of network load attributable to monitoring.
Depending on the root cause of the perceived component failures, system status manager 402 may cause control node 406 to essentially ignore or downgrade monitoring failures during periods of high system service utilization. Instead of adjusting the monitoring retry interval for each managed component, control node 406 ignores or downgrade failures for all components—or a subset of identified components—until the system resources return to normal utilization.
This rule-based, autonomic response to failure response may result in a system that is less susceptible to the impact of system resource capacity issues, and that provides a more accurate view of the health of control node 406 and the components control node 406 manages. Instead of enforcing hard, inflexible rules on response time and retry intervals for a monitored component, autonomically monitored system 400 uses those rules as guidelines for monitoring system resources and components and responds to real world situations without spurious failures caused by rigid rules.
In practice, system status manager 402 may comprises one or more software module having instructions stored within a computer-readable storage medium and executing within an operating environment provided by control node 406 or other processing node.
MRE 412 contains a backward chaining rule set designed to perform root cause analysis when control node 406 detects a component failure. When a rule is fired, MRE 412 collects information from monitoring component 410 (if information is available), or sends an action request to a task manager 414 to have the information collected by monitoring component 410. Task manager 414 provides a pluggable infrastructure for adding or augmenting actions required to complete a task.
As MRE 412 receives required information, MRE 412 attempts to determine if the original component failure is a genuine failure or a side effect of a degraded or failed system administration service. If MRE 412 determines that the detected failure is a genuine component failure, MRE 412 does nothing. However, if the failure is a side effect of a system administrative service, MRE 412 determines the scope of the impact of the failure and sends an action request to task manager 414. The action request contains a set of operations to be performed by control node 406. The set of operations may include modifications to monitoring settings for the affected components, or instructions to control node 406 to ignore monitoring settings for the set of components. Task manager 414 is responsible for instituting the action request passed to task manager 414 by MRE 412.
Various embodiments have been described. For example, an embodiment has been described in which a system status manager is a separate component from a control node. Nevertheless, the following claims encompass many alternative embodiments. For example, functionality of a system status manager may be integrated into a control node. These and other embodiments are within the scope of the following claims.
This application claims the benefit of U.S. provisional Application Ser. No. 60/797,214, filed May 3, 2006, the entire content of which is incorporated herein by reference.
| Number | Name | Date | Kind |
|---|---|---|---|
| 6990518 | Secer | Jan 2006 | B1 |
| 7277826 | Castelli et al. | Oct 2007 | B2 |
| 20030079160 | McGee et al. | Apr 2003 | A1 |
| 20050040223 | Yigit et al. | Feb 2005 | A1 |
| 20060107086 | Walker et al. | May 2006 | A1 |
| 20060233312 | Adams et al. | Oct 2006 | A1 |
| 20060276995 | Breitgand et al. | Dec 2006 | A1 |
| 20070260911 | Marilly et al. | Nov 2007 | A1 |
| 20070271276 | Allen et al. | Nov 2007 | A1 |
| 20070288791 | Allen et al. | Dec 2007 | A1 |
| 20080010533 | Hirano et al. | Jan 2008 | A1 |
| 20090138590 | Lee et al. | May 2009 | A1 |
| 20090271511 | Peracha | Oct 2009 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20070288791 A1 | Dec 2007 | US |
| Number | Date | Country | |
|---|---|---|---|
| 60797214 | May 2006 | US |
