1. Field of the Invention
This invention relates to network computing, and more particularly to grid computing systems.
2. Description of the Related Art
Grid computing enables organizations to use their distributed computing resources more efficiently and flexibly, providing more usable power out of existing systems—and helping organizations gain a competitive business advantage. Conceptually, a grid is quite simple: it is a collection of computing resources connected through a network. Grid middleware aggregates these resources—e.g. servers, storage, databases, and scientific instruments—and provides transparent, remote, and secure access to computing power wherever and whenever it is needed. Grid computing aggregates resources and delivers computing power to every user in the network. A compute grid may include distributed compute resources including one or more of, but not limited to: desktop, server, and High Performance Computing (HPC) systems. Grid computing may provide benefits not available with traditional computing models including one or more of, but not limited to: better utilization of resources, increased user productivity, scalability, and flexibility.
The simplest form of a grid, a Cluster Grid, consists of multiple systems interconnected through a network. Cluster Grids may contain distributed workstations and servers, as well as centralized resources in a data center environment. Typically owned and used by a single project or department, Cluster Grids support both high throughput and high performance jobs. Common examples of the Cluster Grid architecture include compute farms, groups of multi-processor HPC systems, Beowulf clusters, and networks of workstations (NOW).
Cluster Grids typically employ a standard three-tier system architecture, as shown in
Sun's Cluster Grid Architecture
Grid computing systems typically use a traditional model where a Grid Farm has a static view of the network. Sun's Cluster Grid implementation is an exemplary grid computing system that may be used for computation-intensive jobs. The following individual components may be included in Sun's Cluster Grid architecture:
Sun Grid Engine software is a distributed management product that optimizes utilization of software and hardware resources. Sun Grid Engine finds a pool of idle resources and harnesses it productively, so an organization gets as much as five to ten times the usable power out of systems on the network. Sun Grid Engine software aggregates available compute resources and delivers compute power as a network service.
JXTA
Sun's JXTA is an exemplary peer-to-peer platform. Peer-to-peer platforms such as JXTA may provide protocols for building networking applications that thrive in dynamic environments. JXTA technology is a set of open protocols that allow any connected device on the network ranging from cell phones and wireless PDAs to PCs and servers to communicate and collaborate in a peer-to-peer (P2P) manner. JXTA peers create a virtual network where any peer can interact with other peers and resources directly even when some of the peers and resources are behind firewalls and NATs or are on different network transports. In JXTA, every peer is identified by an ID, unique over time and space. Peer groups are user-defined collections of entities (peers) that may share a common interest. Peer groups are also identified by unique IDs. Peers may belong to multiple peer groups, discover other entities and peer resources (e.g. peers, peer groups, services, content, etc.) dynamically, and publish themselves and resources so that other peers can discover them.
Embodiments of a self-healing grid mechanism using peer-to-peer platform protocols are described. In one embodiment, two (or more) nodes on a grid may be configured as master nodes. One of the configured master nodes may serve as the actual master node, and one may be backup or “shadow” master node, not actively serving as a master node. If the active master node goes down, the peer-to-peer protocols may be used to detect that the master node is not active and the backup master node may take over the master node operations for the grid. The backup master node may “shadow” the master to node, keeping updated information about the grid configuration and operations via the peer-to-peer platform protocols, so that it can assume grid management operations seamlessly.
A backup master node may track and maintain information about a grid. In one embodiment, the backup master node may communicate with the master node of the grid using one or more peer-to-peer platform protocols to collect and maintain information about the one or more compute nodes of the grid. In one embodiment, the information about the one or more compute nodes may include, but is not limited to, state information about one or more jobs executing on the compute nodes. In one embodiment, the grid may be configured according to Sun Cluster Grid architecture. Other embodiments may use other grid architectures. In one embodiment, the peer-to-peer platform protocols may be JXTA protocols. Other embodiments may use other peer-to-peer platform protocols.
The master node managing operations of the grid may go down at some point. The backup master node may detect that the master node is down using one or more peer-to-peer platform protocols. In one embodiment, the backup master node may poll the master node using one or more peer-to-peer platform protocols, and may determine that the master node is down if the master node does not respond to the poll. In one embodiment, the master node may send a notification to the backup master node to inform the backup master node that it is going down and that the backup master node needs to assume management of the grid. In other embodiments, other mechanisms may be used to detect if the master node is down.
The backup master node may notify the compute nodes of the grid that it is assuming master node operations for the grid after determining that the master node is down. To determine the compute nodes of the grid, in one embodiment, the backup master node may discover compute nodes of the grid using one or more peer-to-peer platform protocols. In another embodiment, information about the one or more compute nodes of the grid may be obtained from the master node and maintained by the backup master node using one or more peer-to-peer platform protocols. In one embodiment, the backup master node may notify the one or more compute nodes of the grid that the backup master node is assuming management of the grid using one or more peer-to-peer platform protocols. The compute nodes may then update their grid configuration information to indicate that the backup master node is assuming management of the grid in response to the notification. In one embodiment, the backup master node may advertise that it is assuming management of the grid using one or more peer-to-peer platform protocols. Potential or existing compute nodes and any other node, such as job submitter nodes, may then discover that the backup master node is now the master node for the grid using one or more peer-to-peer platform protocols. The backup master node may then assume normal master node operations for the grid.
In one embodiment, if one of the compute nodes in a grid becomes unavailable, a master node can determine, using peer-to-peer platform protocols, that the compute node is no longer part of the grid. In one embodiment, the compute node may send notification of departure to the master node using a peer-to-peer platform protocol. Alternatively, the master node may poll the compute nodes using one or more peer-to-peer platform protocols to determine if and when a compute node goes away. If a job the compute node was working on was in progress, the master node may resubmit the job to another compute node for completion using one or more peer-to-peer platform protocols.
While the invention is described herein by way of example for several embodiments and illustrative drawings, those skilled in the art will recognize that the invention is not limited to the embodiments or drawings described. It should be understood, that the drawings and detailed description thereto are not intended to limit the invention to the particular form disclosed, but on the contrary, the intention is to cover all modifications, equivalents and alternatives falling within the spirit and scope of the present invention as defined by the appended claims. The headings used herein are for organizational purposes only and are not meant to be used to limit the scope of the description or the claims. As used throughout this application, the word “may” is used in a permissive sense (i.e., meaning having the potential to), rather than the mandatory sense (i.e., meaning must). Similarly, the words “include”, “including”, and “includes” mean including, but not limited to.
Embodiments of an autonomic grid computing mechanism are described. Integrating peer-to-peer platform protocols with a grid computing system may enable the grid computing system to run in a heterogeneous network and to be self-maintaining. In embodiments, integrating peer-to-peer platform protocols (e.g. JXTA protocols) with a grid computing system such as the Sun Cluster Grid architecture and Grid Engine may enhance the grid computing system by enabling or improving one or more of, but not limited to: self-deployment and self-configuration, self-updating, node removal, monitoring, administration, load balancing, grid brokering, and decentralized grids. Embodiments may also allow multi-grid support with multiple master nodes in a single grid. Embodiments of the autonomic grid computing mechanism may provide one or more of, but not limited to: a reduced number of manual configurations; quicker deployment; reduced cost of deployment; and preferably simplified addition and removal of nodes in a grid farm.
Embodiments may use peer-to-peer platform protocols (e.g., JXTA protocols) as the transport for the grid computing system. In the prior art, grids typically use TCP/IP. Using peer-to-peer platform protocols may allow a grid computing system to extend and operate on a wider network, across firewalls and other such network barriers.
Embodiments of an autonomic grid computing mechanism may provide one or more mechanisms for adding new compute nodes and for removing compute nodes. Embodiments may allow for the rapid deployment of grid farms. Embodiments of an autonomic grid computing mechanism may provide capabilities including one or more of, but not limited to: multi-grid support, grid management, remote access/management, load balancing, and grid cell selection.
In one embodiment, grid master nodes may be configured manually. Peer-to-peer platform protocols and discovery mechanisms may be used to automatically configure compute nodes. In one embodiment, a bootstrapping mechanism or agent based on a peer-to-peer platform (e.g. JXTA) may be installed on (potential) compute nodes. The bootstrapping mechanism discovers a master node using a peer-to-peer platform discovery mechanism, informs the master node that the node is joining the network using one or more peer-to-peer platform protocols, and provides details of the node to the master node using one or more peer-to-peer platform protocols. When a potential compute node comes up, the node (through the bootstrapping mechanism) may search for a master node using the peer-to-peer platform discovery mechanism. When the potential compute node finds a master node, it may provide detail information about the compute node to the master node using one or more peer-to-peer platform protocols. Based on those details, the master node may know that the new compute node has joined the grid network, record the fact, and, depending on what the compute node's details are, configure the new compute node accordingly. In one embodiment, once the compute node discovers a master node, the peer-to-peer platform bootstrapping mechanism may terminate.
In one embodiment, the master node may send the new compute node one or more scripts, files, and other grid configuration information using one or more peer-to-peer platform protocols so that the node can self-configure to be grid-ready as a compute node. Before the master node provides the information, the compute node may have had no information about the grid, and may contain no software related to the grid other than the bootstrapping mechanism. Through the configuration over peer-to-peer platform protocols, the master node may send all necessary files and instructions (e.g., mount instructions, directories to share, what directories to put files in, permission changes, etc.) to a new compute node. Using this information, the new compute node may self-configure to become part of the grid network. Whenever a job comes in, the master node knows about the new compute node and may send the new job to the new compute node, in one embodiment using one or more peer-to-peer platform protocols.
Note that, in a network, a potential compute node discovering a master node may be a new node joining the network, a compute node that needs upgrading, or a compute node that was shut down and is coming up again. In one embodiment, hen a compute node reboots, it may inform the master node that it is a new node but that it is already partially or fully configured. In this instance, the master node may not need to fully configure the new node. Depending on the status of the compute node provided to the master node, the master node may determine that the compute node is partially configured and may provide additional configuration information, if necessary.
One embodiment may provide a mechanism for patch and/or upgrade management. For example, a compute node may have older software (e.g. version 1.0). The master node may be running newer software (e.g. version 1.1). The master node may determine from the information passed to it by the compute node that the compute node has the wrong version of software and may provide information such as how to upgrade, patches to install, and/or other upgrade information instructions to the compute node.
In one embodiment, peer-to-peer platform protocols may be used by compute nodes to discover master nodes and by master nodes to discover compute nodes. In one embodiment, if a master node knows there are (potential) compute nodes somewhere on the network, the master node may search for the compute nodes using one or more peer-to-peer platform protocols and discovery mechanisms and invite the compute nodes to join the grid.
In one embodiment, deployment (installation and configuration) of the nodes in a grid farm may be automated using peer-to-peer platform (e.g. JXTA) discovery and communication protocols. In one embodiment, the master node 200 may be installed manually. In one embodiment, other nodes (e.g. compute nodes 204A and 204B or potential compute nodes) may be preloaded with a peer-to-peer platform-based (preferably, small footprint) bootstrapping mechanism or agent (not shown). When a new node 204 is deployed in a grid farm network, the bootstrapping mechanism searches for a master node 200 using one or more peer-to-peer platform protocols (e.g. JXTA Discovery protocols). Upon discovery, the new compute node 204 notifies the master node 200 about its presence and capabilities/configuration (e.g. hardware and software capabilities/configuration such as CPU power, disk space, installed software, OS, etc.) using one or more peer-to-peer platform communication protocols (e.g. JXTA Pipe protocols). In one embodiment, the new compute node 204 may notify the master node 200 about its presence and capabilities using advertisements (e.g. JXTA advertisements). The master node 200 then may configure the new compute node 204 and make it part of the grid farm, in one embodiment using one or more peer-to-peer platform communication protocols (e.g. JXTA Pipe protocols). In one embodiment, the master node 200 may send configuration data (e.g. binaries, scripts, NFS mount points, etc.) to the compute node 204. In one embodiment, when a session (job) is complete, the master node 200 may use one or more one or more peer-to-peer platform communication protocols (e.g. JXTA Pipe protocols) to signal the end of session to the compute nodes 204.
In one embodiment, as an alternative to the compute nodes 204 discovering the master nodes 200, a master node 200 may discover compute nodes 204. In this embodiment, the master nodes 200 may be manually configured. The compute nodes 204 may be automatically configured. A master node 200 may use a peer-to-peer platform discovery mechanism (e.g. JXTA's discovery service) to discover one or more compute nodes 204. The master node 200 may then configure the discovered compute nodes 204 using one or more peer-to-peer platform communication services (e.g. JXTA Pipe service). In one embodiment, when a session (job) is complete, the master node 200 may use one or more one or more peer-to-peer platform communication services (e.g. JXTA Pipe service) to signal the end of session to the compute nodes 204.
Node 250 may include, in memory 254, a bootstrapping mechanism 256. In one embodiment, node 250 may be preloaded with bootstrapping mechanism 256. When node 250 is deployed in a grid farm network, the bootstrapping mechanism 256 searches for a master node 260 using one or more peer-to-peer platform protocols. Upon discovery of master node 260, node 250 may notify the master node 260 of its presence and capabilities/configuration using one or more peer-to-peer platform communication protocols. The master node 260 then may send configuration information to node 250 using one or more peer-to-peer platform protocols, and node 250 may then use the configuration information to self-configure as a compute node in the grid farm.
In one embodiment, the installation and setup of the P2P platform (e.g., JXTA) master 302 may be performed manually. In one embodiment, a setup file on P2P platform master 302 may include, but is not limited to, the following information: Master Hostname, Master IP, Mount Point, and Command port.
In one embodiment, P2P platform master 302 may include one or more scripts that are configured for use by a P2P platform client 304 to self-install as an execution host (i.e. compute node).
In one embodiment, as indicated at 310, installation may begin with the P2P platform client 304 sending a multicast discovery with its IP address and hostname embedded in the message. As indicated at 312, P2P platform master 302 may then add the execution host to the grid. In one embodiment, this may include setting up the necessary files and informing a grid master 300 of the addition of this host to the grid, for example by sending a command (in an exemplary embodiment using Sun's Grid Engine and JXTA, a qconf-ah command). In one embodiment, this command is preferably run on an administrative host within the grid. In one embodiment, the host that this command is run on is preferably able to resolve the hostname of the P2P platform client 304. In one embodiment, this may be a simple command, either adding this host to a file (e.g. /etc/hosts file) or checking if the P2P platform master 302 can resolve this host and adding on the negative. The peer-to-peer platform IDs (e.g. JXTA ID) of the systems that have been found may be tracked, for example for use in keeping track of the status of the system.
As indicated at 314, script(s) and/or arguments may be sent from P2P platform master 302 to P2P platform client 304. In one embodiment, P2P platform master 302 sends one or more scripts along with arguments to be run on the script(s) in one or more messages (e.g. peer-to-peer platform protocol messages). In one embodiment, P2P platform client 304 may then read the script(s) from the message(s), create a file, and then run the script(s) with the arguments from the message specified within the script.
As indicated at 316, P2P platform client 304 may then send return code to P2P platform master 302. Once P2P platform client 304 is finished running the script, it may send return code back to P2P platform master 302. This return code may, for example, be used to deal with errors that may occur within the scripts, may be logged, and/or may be used in carrying out further action dependent on the return code.
In one embodiment, 314 and 316 (receiving and running the script(s) and arguments) may be repeated with other scripts until the installation has completed successfully. Script actions that may be performed may include one or more of, but are not limited to:
Note that, in one embodiment, this list may reflect the order that the script actions are preferably run in, but in other embodiments, the actions may be performed in other orders, one or more actions may be left out, and/or one or more additional actions may be performed. In one embodiment, the grid computing system may be the Sun Grid Engine (SGE). Once these actions have been completed, the compute node (on P2P platform client 304) will be part of the grid and may be able to run jobs that have been submitted to its queue by P2P platform master 302.
In one embodiment, a peer-to-peer platform (e.g. JXTA) agent on a master node may track the peer-to-peer platform identifiers of each installed host (i.e. compute node) and/or the peer-to-peer platform identifiers of each host that attempts to install. This may allow a status or state to be maintained for each host (compute node) that is installed (or not installed). In one embodiment, this state may be one of: Installed, Uninstalled, or Error.
In one embodiment, the node may be a new node or a node that was previously configured as a compute node. In the latter case, to self-configure, the node may update its compute node configuration using the grid configuration information. In one embodiment, if the node was previously configured as a compute node, the information the node sends to the master node may include compute node configuration information for the node. The master node may determine that the compute node configuration needs to be updated from the compute node configuration information. The master node may then include update information for the compute node configuration in the grid configuration information sent to the node.
After the node has self-configured as a compute node in the grid, the master node may record, track, and manage the presence and activity of the new compute node in the grid. In one embodiment, the master node may receive jobs from one or more job submitter nodes using one or more peer-to-peer platform protocols. The master node may distribute one or more jobs to the compute nodes, including the new compute node, for execution using one or more peer-to-peer platform protocols. The master node may receive results of the execution from the compute nodes using one or more peer-to-peer platform protocols, and the master node may then send the results to the job submitter node using one or more peer-to-peer platform protocols.
In one embodiment, if the node later needs to leave the grid, the node may send the master node notification that it is leaving the grid using one or more peer-to-peer platform protocols. In one embodiment, the master node may poll or otherwise track the compute nodes in the grid using one or more peer-to-peer platform protocols. In one embodiment, if the master node has submitted a job to the node for execution that has not completed, after determining or detecting that the node has left the grid, the master node may resubmit the job to another compute node of the grid for execution using one or more peer-to-peer platform protocols.
In one embodiment, peer-peer-platform (e.g. JXTA) discovery protocols may enable a grid farm to be dispersed geographically at different locations. This may enable a grid farm to be spread among two or more different physical locations instead of all the nodes being deployed at the same physical location (e.g. room).
Self-Updating Grids
In one embodiment, a grid farm may be self-updating. In one embodiment, each node in a grid farm may maintain information including one or more of, but not limited to: the version of the grid engine it is using, if it failed during configuration process, etc. In one embodiment, when a compute node reboots, it may inform the master node about its current status, and the master node may appropriately upgrade/configure the compute node, if necessary. In one embodiment, if a compute node is determined to be properly configured and is running a correct version of grid computing system (grid engine) software, it is preferably not reconfigured. In one embodiment, software upgrades may be performed using a centralized solution such as JNLP, CVS, or other software upgrade methods.
In one embodiment, rather than relying on centralized updating for compute nodes, compute nodes may be updated from a “nearby” peer using one or more peer-to-peer platform (e.g. JXTA) protocols. The nearby peer may be a master node, another compute node, or even a peer that is not currently participating in the grid farm. For the purpose of this document, a “nearby peer” may be considered a logically nearby peer. A logically nearby peer is a peer that may respond as fast as or faster than other peers, but that may or may not be physically nearby or physically closer than at least some other peers. A physically nearby peer is a peer in close physical proximity. However, a physically nearby peer may or may not be logically nearby. In summary, a nearby peer may be considered a peer that is logically nearby which may or may not be physically nearby.
In one embodiment, a compute node may communicate with a nearby peer using one or more peer-to-peer platform (e.g. JXTA) protocols to determine if its grid computing software and/or data is up-to-date and, if not, may download updated software and/or data from the nearby peer. In another embodiment, a compute node may communicate with the master node using one or more peer-to-peer platform (e.g. JXTA) protocols to determine if its grid computing software and/or data is up-to-date and, if not, may communicate with a nearby peer (which may be another execution peer) using one or more peer-to-peer platform (e.g. JXTA) protocols to update its software/data.
Node Removal
Embodiments of an autonomic grid computing mechanism may provide a mechanism for node removal. When a compute node needs to leave a grid farm (temporarily or permanently), it may notify the master node using one or more peer-peer-platform (e.g. JXTA) protocols, and the master node may then gracefully remove the compute node from the grid farm. Embodiments of the installation, self-configuration and removal mechanisms described herein, using peer-peer-platform (e.g. JXTA) protocols, preferably enable any node to dynamically join a grid farm, self-configure, and to remove itself if necessary or desired.
Multi-Grid Support
Embodiments of an autonomic grid computing mechanism may provide a mechanism for multi-grid support. In this embodiment, there may be multiple master nodes, and thus multiple smaller grids (or grid cells), in a single (larger) grid. Through multi-grid support, job submitters may be configured to automatically select one of the grid cells for performing a particular job. Multi-grid support may also provide load balancing.
In the prior art, grids are monolithic, with one master node and multiple compute nodes. In this prior art architecture, the master node may be a bottleneck for performance and a single point of failure. There is just one master node, and under some conditions, the master node may not be able to dispatch jobs quickly and efficiently. In an embodiment of an autonomic grid computing mechanism that provides multi-grid support, instead of just having one big grid network, the grid network may be split into multiple smaller grids. The job submitter may use peer-to-peer platform protocols to determine which one of the smaller grids may be able to do the job more quickly, and/or to discover other information about the smaller grids that may be used to determine which of the smaller grids is best suited to do the job.
In one embodiment, a job submitter may use the peer-to-peer platform to search for and discover grids, and then determine which grid may perform a job faster and/or cheaper. The job submitter may use peer-to-peer platform (e.g. JXTA) protocols to acquire information about the various grids, and then the job submitter may determine which of the grids it wants to send a job to based on the acquired information. Based on one or more criteria, the job submitter may determine which grid it wants to perform the job. For example, one of the grids may be overloaded, but another grid may be idle. A job submitter may also, for example, look for a grid that matches particular requirements. A particular job may have particular processing requirements, for example graphics processing requirements, and the job submitter may look for a grid that includes at least one compute node with a graphics engine.
In one embodiment, the plurality of grids may be configured according to Sun Cluster Grid architecture. Other embodiments may use other grid architectures. In one embodiment, the peer-to-peer platform protocols may be JXTA protocols. Other embodiments may use other peer-to-peer platform protocols. As indicated at 422, the job submitter node may select one of the grids to perform a particular job. In one embodiment, the job submitter node may determine from the obtained information that the selected grid is best suited to perform the particular job based on one or more criteria including, but not limited to: speed of job execution (how quickly it is expected the grid can execute the job and return results) and one or more special processing requirements (e.g. graphics capabilities) for the compute nodes of the grid. As indicated at 424, after selecting the grid, the job submitter node may submit the particular job to the selected grid (and, potentially, additional jobs) using one or more peer-to-peer platform protocols.
Load Balancing
Embodiments of an autonomic grid computing mechanism may provide a mechanism for load balancing among grid farms using one or more peer-to-peer platform (e.g. JXTA) protocols. In one embodiment, before submitting a job to a grid farm, a job submitter may use one or more peer-to-peer platform protocols to search for a suitable grid farm in a multi-grid environment. Determining a suitable grid farm may be based on criteria such as how many jobs are pending in a grid's queue, if a grid farm has particular resources required to complete a job, etc. In one embodiment, a job submitter may determine which one of the grids is likely to complete the job faster considering the current load.
Self-Healing Grid Mechanism
Embodiments of a grid computing mechanism may provide a mechanism for self-healing in grid farms. In one embodiment, two or more nodes may be (manually) configured as master nodes. One or more of the configured master nodes may serve as actual or active master nodes, and one or more of the configured master nodes may be backup or “shadow” master nodes, not actively serving as master nodes. If an active master node goes down or becomes unavailable, the peer-to-peer (e.g. JXTA) protocols may be used to detect that the master node is not active and one of the other nodes configured as a master node may take over the master node operations for the grid of the (no longer available) master node.
Conventionally, grids have one master node, which presents the problem of a single point of failure. As illustrated in
In one embodiment, a backup master node 502 may discover that the master node 500 has failed using the peer-to-peer platform protocols. In one embodiment, a backup master node 502 may periodically poll or attempt to discover the master node 500 using one or more peer-to-peer platform protocols, and upon determining that it can no longer discover the master node 500 may take over the role of master node for the grid. In one embodiment, the master node 500 may notify the backup master node 502 that it is leaving the grid using one or more peer-to-peer platform protocols.
Typically, in a grid, the compute nodes 504 do not know about each other. A master node 500 may communicate with, control, and maintain information about the compute nodes 504. As illustrated in
As indicated at 520, the master node managing operations of the grid may go down. As indicated at 522, the backup master node may detect that the master node is down using one or more peer-to-peer platform protocols. In one embodiment, the backup master node may poll the master node using one or more peer-to-peer platform protocols, and may determine that the master node is down if the master node does not respond to the poll. In one embodiment, the master node may send a notification to the backup master node to inform the backup master node that it is going down and that the backup master node needs to assume management of the grid. In other embodiments, other mechanisms may be used to detect if the master node is down.
As indicated at 524, the backup master node may assume management of the grid after determining that the master node is down. In one embodiment, assuming management of the grid may include advertising that the backup master node has assumed management of the grid using one or more peer-to-peer platform protocols. This advertisement may be discovered by other nodes, for example job submitter nodes, to determine that the backup master node is now acting as master node for the grid. In one embodiment, assuming management of the grid may include the backup master node discovering the one or more compute nodes of the grid using one or more peer-to-peer platform protocols. In one embodiment, assuming management of the grid may include the backup master node notifying the one or more compute nodes of the grid that the backup master node has assumed management of the grid using one or more peer-to-peer platform protocols. The compute nodes may update their grid configuration information to indicate that the backup master node has assumed management of the grid in response to the notification.
As indicated at 532, the master node managing operations of the grid may go down. As indicated at 534, the backup master node may detect that the master node is down using one or more peer-to-peer platform protocols. In one embodiment, the backup master node may poll the master node using one or more peer-to-peer platform protocols, and may determine that the master node is down if the master node does not respond to the poll. In one embodiment, the master node may send a notification to the backup master node to inform the backup master node that it is going down and that the backup master node needs to assume management of the grid. In other embodiments, other mechanisms may be used to detect if the master node is down.
As indicated at 536, the backup master node may notify the compute nodes of the grid that it is assuming master node operations for the grid after determining that the master node is down. To determine the compute nodes of the grid, in one embodiment, the backup master node may discover compute nodes of the grid using one or more peer-to-peer platform protocols. In another embodiment, information about the one or more compute nodes of the grid may be obtained from the master node and maintained by the backup master node using one or more peer-to-peer platform protocols. In one embodiment, the backup master node may notify the one or more compute nodes of the grid that the backup master node is assuming management of the grid using one or more peer-to-peer platform protocols. The compute nodes may then update their grid configuration information to indicate that the backup master node is assuming management of the grid in response to the notification. In one embodiment, the backup master node may advertise that it is assuming management of the grid using one or more peer-to-peer platform protocols. Potential or existing compute nodes and any other node, for example job submitter nodes, may then discover that the backup master node is now the master node for the grid using one or more peer-to-peer platform protocols. As indicated at 538, the backup master node may then assume normal master node operations for the grid.
Grid Brokering
Embodiments of an autonomic grid computing mechanism may provide a mechanism for grid brokering. In one embodiment, if a grid has too many pending jobs, or if it has certain jobs pending for which it cannot find a suitable execution host (e.g., due to resource requirements which cannot be fulfilled), one or more peer-to-peer platform (e.g. JXTA) protocols may be used by the grid master node to discover one or more other grids and to determine if these other grids can take some of the job load, or can perform a job with special requirements. In one embodiment, out of all possible grids, the grid brokering mechanism may be used to determine which grid is best suited to accept which jobs.
In one embodiment, if the master node determines it may not be able to handle a job submitted by a job submitter (e.g. due to overload or missing resource requirements), the master node may try to find another grid (and another master node) that can handle the job (or part of a job) using one or more peer-to-peer platform protocols. In other words, the master node may subcontract a job, e.g. determine if there is another grid that can perform the job (or part of a job) and forward the job to the other grid for execution using one or more peer-to-peer platform protocols.
Decentralized Grids
Embodiments of an autonomic grid computing mechanism may provide a mechanism for decentralized grids where job submitters can directly select compute nodes and directly submit jobs to the selected compute nodes without going through master nodes.
In a typical prior art grid configuration, the Master node acts as a centralized server, and thus may become a bottleneck. In one embodiment, as illustrated in
Monitoring Nodes in a Grid Farm
Embodiments of an autonomic grid computing mechanism may provide a mechanism for monitoring nodes in a grid farm. In one embodiment, a master node may maintain a database of other nodes (e.g. compute nodes) in the grid farm. In one embodiment, one or more peer-to-peer platform protocols (e.g., JXTA's Peer Information Protocol) may be used by the master node to query the other nodes about the nodes' status. The status may include one or more of, but is not limited to: current queue length, current pending jobs, estimated time to complete a current job, etc.
In one embodiment, a master node may discover information about the compute nodes using one or more peer-to-peer platform protocols. The master node may monitor the compute nodes and perform administration of the grid. In one embodiment, compute nodes may discover information about one or more master nodes using one or more peer-to-peer platform protocols. The compute nodes may monitor the master node; for example, to determine the status of the master node.
In one embodiment, there may be a separate node (e.g. a peer in a peer-to-peer network), referred to as a monitor node, running a monitor application. The monitor node may keep status information for one or more master nodes and their associated grids (compute nodes). In one embodiment, the monitor node may use one or more peer-to-peer platform protocols to poll or otherwise obtain grid information from the master node(s). In one embodiment, the master node(s) may maintain status information about the compute nodes on the grid(s). If a compute node leaves or joins the grid, the peer-to-peer platform protocols may be used to update the master node with the information, and a grid configuration maintained by the monitor application on the separate node may be updated accordingly using the peer-to-peer platform protocols.
The server side of the client/server model is represented by administration host 704 in
In one embodiment, the P2P platform agent 706 may connect to a relay/rendezvous node on the peer-to-peer platform network 700 and create a group. This group may then be joined by other peers within the peer-to-peer network, allowing them to communicate with each other using one or more peer-to-peer platform protocols. The P2P platform agent 706 may then broadcast this information to the rest of the group.
The client side of the client/server model is represented by P2P platform monitor 702 in
Grid Administration Mechanism
In one embodiment, peer-to-peer platform (e.g. JXTA) protocols may be used to remotely administer one or more grids, decentralized grids, and/or grids in a multi-grid environment, for example using communication (e.g. JXTA pipe) protocols. A grid administration node may include one or more grid administration applications for managing the grids using the peer-to-peer platform. In one embodiment, the administration node may also include a monitor mechanism to monitor the grid(s). An exemplary grid administration tasks that may be performed using embodiments is adding or removing one or more machines depending on load. Another example is gathering statistics, for example, price/performance statistics, for use in decision making. The gathered statistics may be used to answer questions such as: Is this grid viable? Are more machines needed? When would be the right time to buy new machines? Gathered statistics may also be used in making load predictions. Other administrative tasks than these exemplary tasks may be performed using embodiments.
Node 750 may include, in memory 754, a grid administration mechanism 756. Master nodes 760 may use one or more peer-to-peer platform protocols to gather information from compute nodes 762. Grid administration mechanism 756 may use one or more peer-to-peer platform protocols to gather grid information from master nodes 760. This grid information may include information collected from compute nodes 762. Grid administration mechanism 756 may be used to remotely perform various administrative and monitoring tasks for the grids using one or more peer-to-peer platform protocols. At least part of the grid information may be displayed for the user on node 750 by grid administration mechanism 756.
Various embodiments may further include receiving, sending or storing instructions and/or data implemented in accordance with the foregoing description upon a carrier medium. Generally speaking, a carrier medium may include storage media or memory media such as magnetic or optical media, e.g., disk or CD-ROM, volatile or non-volatile media such as RAM (e.g. SDRAM, DDR SDRAM, RDRAM, SRAM, etc.), ROM, etc. as well as transmission media or signals such as electrical, electromagnetic, or digital signals, conveyed via a communication medium such as network and/or a wireless link.
The various methods as illustrated in the Figures and described herein represent exemplary embodiments of methods. The methods may be implemented in software, hardware, or a combination thereof. The order of method may be changed, and various elements may be added, reordered, combined, omitted, modified, etc.
Various modifications and changes may be made as would be obvious to a person skilled in the art having the benefit of this disclosure. It is intended that the invention embrace all such modifications and changes and, accordingly, the above description to be regarded in an illustrative rather than a restrictive sense.
This application is a continuation-in-part of U.S. patent application Ser. No. 10/659,820, titled “Autonomic Grid Computing Mechanism” filed Sep. 11, 2003.
Number | Name | Date | Kind |
---|---|---|---|
6009455 | Doyle | Dec 1999 | A |
6192401 | Modiri et al. | Feb 2001 | B1 |
7376693 | Neiman et al. | May 2008 | B2 |
20030074393 | Peart | Apr 2003 | A1 |
20030074402 | Stringer-Calvert et al. | Apr 2003 | A1 |
20030200347 | Weitzman | Oct 2003 | A1 |
20040024891 | Agrusa et al. | Feb 2004 | A1 |
20040034807 | Rostowfske | Feb 2004 | A1 |
20040064548 | Adams et al. | Apr 2004 | A1 |
20040098447 | Verbeke et al. | May 2004 | A1 |
20040210673 | Cruciani et al. | Oct 2004 | A1 |
Number | Date | Country | |
---|---|---|---|
Parent | 10659820 | Sep 2003 | US |
Child | 10828955 | US |