Managing the Grid

DeployR Enterprise supports an extensible grid framework, providing load balancing capabilities across a network of node resources.

Each node on the grid contributes its own processor, memory, and disk resources. These resources can be leveraged by the DeployR server to run R code as part of user authenticated, asynchronous, and anonymous operations.

With the DeployR Enterprise, you can dynamically scale the overall workload capacity by enabling and disabling the number of available grid nodes, and by designating these nodes to specific modes of operation. The grid can then automatically distribute the workload across the set of enabled nodes. Or, you can create custom node clusters tuned for specific operation types and resource requirements. Note that DeployR Open supports only a static grid framework of a single, local grid node with a fixed slot limit.

When configuring the grid, you should begin by gaining an understanding of the anticipated workload for the grid. With that workload in mind, you can then scale the grid to handle the load by adding or changing the resources of each grid node.

A well-configured grid maximizes throughput while minimizing resource contention. The server and grid deployment model is described as Server[1] -> Node[1..N] where N is defined as the total number of nodes on the grid. For more on planning and provisioning, see the Scale & Throughput Guide.

Figure: Node List page

Node Operation Types

You must designate an operation type for each node on the grid. The following table describes the distinct types of operations recognized by the grid.

While a mixed mode exists, dedicating a node to a specific operation type provides a simple yet powerful mechanism that guarantees processing isolation between the distinct modes of operation being handled concurrently on the grid.

DeployR Open includes only the local node, DeployR Default Node. Therefore, this node should generally remain designated to the mixed mode.

Once a node is dedicated to a specific operating type, the administrator can further tailor its configuration by provisioning processor, memory, and disk resources appropriate to the nature of the targeted operations.

Each operation type has a distinct set of server policies that impact the time-out policy, IP filters, encryption, and so on.

Grid Workload Distribution

DeployR Enterprise supports an extensible grid framework, providing load balancing capabilities across a network of node resources.

There are two workload distribution models supported by the DeployR grid:

1. Operation-cluster model
2. Named-cluster model

Each node on the grid is assigned an operation type: authenticated, asynchronous, anonymous, or mixed mode. The operation type for a node indicates the types of operations that the node will accept at runtime. For example, a grid node configured for authenticated operations will only accept authenticated (project) operations at runtime and never accept asynchronous or anonymous operations. A grid node configured for 'mixed mode' accepts all types of operations.

The operation type assigned to a node and the cluster named for the request (if one was named) are the biggest determinants in how the workload is distributed. In the absence of a specified custom cluster name, the workload is automatically spread across nodes of the same operation type using the operation-cluster model. However, you can further optimize that distribution by allowing a custom cluster to be named for an operation, which then follows the named-cluster workload distribution model.

Operation-Clusters Workload Distribution

The set of nodes dedicated to the same operation type is automatically members of the same operation-cluster. In this way, the operation-cluster workload distribution model provides a simple and automatic way to partition the workload on the grid. There are four operation-clusters, one for each type of operation (authenticated, asynchronous, or anonymous) and one for mixed mode nodes. The specific manner in which an operation is allocated onto a node using the operation-cluster workload distribution model is a function of:

• The type of operation being requested (authenticated, asynchronous, or anonymous)
• The number of nodes on the grid that match the operation-cluster identity
• The current workload for those nodes

If a request does not name a custom cluster, the grid begins by determining the operation type for that request, then identifying all nodes configured for that type of operation, and finally sending the request to the node with the least burden to minimize resource contention and to maximize throughput.

When all of the nodes of a custom cluster are busy, any additional requests for that cluster will be rejected. Whenever an operation-cluster on the grid or the entire grid is operating a maximum capacity, any additional requests for authenticated or anonymous operations will be rejected and any additional requests for asynchronous operations will be queued.

For example, if the grid determines that a request is made for an asynchronous job, the grid will look for all nodes with the asynchronous type, and then send the request to the least busy node in that operation-cluster. However, if all nodes within the asynchronous operation-cluster are busy, then the request is funneled to the least burdened node of mixed mode type.

Named-Clusters Workload Distribution

While the default operation-cluster model works well in many cases, there are some limitations to that model when operations within a single operation-cluster do not share the exact same runtime characteristics. To address this situation, the named-cluster workload distribution model was introduced in DeployR 7.4 to allow administrators to assign custom cluster names to nodes.

As an example, it is not uncommon that authenticated operations that would otherwise execute on the same authenticated operation-cluster differ in need from one another. Some authenticated operations may require nodes that deliver fast processing power, while other authenticated operations may require nodes that deliver large amounts of memory. Named-cluster workload distribution solves this problem by allowing the administrator to designate a custom cluster name for an individual or groups of nodes on the grid that are intended for special purpose operations, such as "high-cpu" or "high-mem". Once a custom cluster has been designate to at least one node in the grid, individual operations can be targeted for execution within that named custom cluster, which ensures suitable runtime resources are available on the node where an operation eventually executes.

The specific manner in which an operation is allocated onto a node using this model is a function of:

• The custom cluster's identity specified as an API parameter for the operation being requested
• The number of nodes on the grid that match the name of the custom cluster
• The current workload for those nodes

Whenever a custom cluster has been named for the requested operation, the grid begins by identifying the nodes that fall within that cluster, then sends the request to the node within that cluster that has the least burden to minimize resource contention and to maximize throughput.

When all of the nodes of a custom cluster are busy, any additional requests for that cluster are funneled to a node configured for the mixed mode operation type, if one is available. Whenever a custom cluster on the grid or the entire grid is operating a maximum capacity, any additional requests for authenticated or anonymous operations will be rejected and any additional requests for asynchronous operations will be queued.

For example, if the grid determines that a request is made for an operation to use the "high-mem" cluster, the grid will look for all nodes within the "high-mem" cluster and then send the request to the least busy node within that cluster. However, if all nodes within the "high-mem" cluster are busy, then the request is funneled to the least burdened node of mixed mode type.

Any node can belong to a custom cluster if the administrator has assigned the node to that cluster by entering its custom name in the Cluster name field. Custom cluster names are not case-sensitive.

Node Validation and Errors

The Administration Console performs a verification of node configurations whenever you:

• Display the Grid tab
• Create or update a grid node configuration
• Import a grid node configuration

When you visit the Grid tab, any enabled nodes found to be unresponsive are highlighted to help you visually identify that these nodes are not operating properly.

When you attempt to create, update or import a node, the verification process is run. If an issue is detected, the following message appears: “Grid node details provided failed to validate, please check your node configuration and try again.”

Common validation issues include:

• The configuration points to a grid node that is improperly installed or not yet installed.
• The configuration points to a grid node that is currently unresponsive for any number of reasons including the RServe process not running on that node.
• The node is behind a firewall or temporarily unreachable.
• The node’s configuration details are incomplete such as an empty name field when the name is required.

Node Properties

These constraints govern limits for processor and memory usage at runtime.

After a default installation of DeployR, it is highly unlikely that these settings require any editing under most operating conditions. However, if DeployR is in conflict with another application on the server machine or if you have made edits to the file Rserv.conf, you must mirror your changes here as well.

If the external data directories are properly configured across all grid nodes and an R script updates the data in the external directory, then those updates are automatically mirrored across all nodes without any manual or external intervention by the admin.

Slots in Use

You can see which slots are currently in use and how they are being used by a particular node in the Slots In Use section. For each slot, you can review the following details:

You can force the termination of a current operation on slot, by clicking the Shutdown link.

Creating New Nodes

While many default node settings are preconfigured during the installation process, the DeployR Enterprise administrator must still declare and configure each node individually in the Administration Console. For more information about installing, see the installation instructions for your server operating system.

To create a new node:

1. From the main menu, click The Grid.

2. From the Grid Node List page, click the New Grid Node link in the menu bar.

   Figure: New Node page

   

3. In the New Grid Node page, define the applicable properties for the node, such as the Name, Host, Operating Type, and Storage Context.

4. Enter a name in the Name field.

5. Enter an IP address in the Host field. Hostname or IP on which the node is configured. The DeployR Default Node is configured on localhost.

6. Specify the Operating Type. The operating mode you define for a grid node has an impact on the operations you can perform on that node.

7. Under External directory configuration, set the Storage Context to reflect the full path to the external data directory on that node’s machine, such as $DEPLOYR_HOME/deployr/external/data where $DEPLOYR_HOME is the installation directory.

8. Click Create to save the new node. The node configuration is validated. If the configuration works, then it is saved. If not, an error message will appear.

Viewing and Editing Nodes

To view and edit a node:

1. From the main menu, click The Grid.

2. In the Grid Node List, click the node you want to view or edit.

3. The Grid Node Details page appears.

4. Click Edit.

5. Make your changes.

6. Click Update to save the changes. The node configuration is validated. If the configuration works, then it is saved. If not, an error message will appear.

Enabling or Disabling Nodes

You can choose whether or not a configured node will be active in the grid or inactive using the Enable checkbox in the Grid Node Details page. When disabled, the node is no longer available to process any requests from DeployR and it will appear highlighted in the main grid node list.

The Default Grid Node cannot be disabled.

To enable or disable a node (DeployR Enterprise Only):

1. From the main menu, click The Grid. The Grid Node List page appears.

2. Click the node you want to edit. The Grid Node Details page appears.

3. Click Edit.

4. Select the Enable checkbox to activate the node, or unselect it to disable the node.

5. Click Update to save the changes.

Viewing or Stopping Slot Activity

In the Grid Node Details page, you can review the slot activity.

Additionally, you can shut down slots in order to terminate activity. This is a last resort for unwanted activity on the server.

This list is not updated in real time. Refresh your page to see the latest activity.

To review the slot activity for a node and shutdown slots:

1. From the main menu, click The Grid. The Grid Node List page appears.

2. In the Grid Node List, click the node whose slot activity you want to monitor. The Grid Node Details page appears.

3. Look at the Slots in Use section at the bottom. Click a slot link to see details about a particular slot.

4. To terminate all of the activity on the selected node, click Shutdown in the toolbar to terminate all R sessions.

5. To terminate the activity on a specific slot, do the following:

   • In the Grid Node Details page, click the slot whose activity you want to see and terminate.
   • In the Slot Details page, click Shutdown to terminate the activity and release the slot.

Deleting Nodes

The Default Grid Node cannot be deleted.

To delete a node:

1. From the main menu, click The Grid. The Grid Node List page appears.

2. Click the name of the node you want to delete. The Grid Node Details page appears.

3. Click Delete. The node is removed.

Exporting Node Configurations

You can export one or more node configurations into one CSV file. Exporting can be used to copy the basic node configurations to another machine or to preserve them as a backup.

Figure: Export Grid Configuration page

To export node configurations:

1. From the main menu, click The Grid. The Grid Node List page appears.

2. Click Export Grid Configuration.

3. Select individual node configurations or all nodes to be exported.

4. Click Export to File and save the file.

Importing Node Configurations

You can import node configurations into the server from a CSV file you previously exported. You can import the contents of this file to same server from which it was exported or across multiple server deployments.

Grid node configurations are validated upon import. If the configuration works, then the node configuration is imported. If not, an error message will appear.

You can only import these configurations into a server instance matching the same DeployR version from which they were exported.

To import:

1. From the main menu, click The Grid. The Grid Node List page appears.

2. Click Import Grid Configurations.

3. Click Browse and select the CSV file to be imported.

4. Click Load.

   Figure: Choose Grid Nodes to Import page

   

5. To disable grid nodes upon import so that you can individually enable them at a later time, select the Disable grid nodes option. You can always enable/disable grid nodes one at a time later. Enabled nodes are activated immediately in the grid and available for load balancing.

6. Choose the nodes to import.

7. Click Import. If a node by the same name already exists, the incoming node configuration is skipped. The node configuration settings are validated. If the configuration works, then it is saved. If not, an error message will appear.