Specifying the Deployment Location of the DataFlow Node

Function Description

The configuration file can be used to specify the deployment location of the DataFlow graph node. This function can implement multi-device and multi-instance deployment. The computing workload is prorated to improve the performance.

Constraints

You must specify the deployment locations of all nodes.
You must specify the deployment configuration file based on ge.experiment.data_flow_deploy_info_path in options of AddGraph. The file must exist and be correctly formatted. For details about the format requirements, see Format of the Configuration File.

For details about options, see Options.
The deployment location must be consistent with the location where the node can be deployed. For example, if the UDF supports only the compilation of host deliverables, the node cannot be deployed on a device. Similarly, if the node can be deployed only on a device, it cannot be deployed on a host.
A node cannot be configured repeatedly.
A node cannot be deployed on both the host and device.

How to Use

The following is an example:

     
          flow_node0 = df.FlowNode(input_num=1, output_num=2)
flow_node0_out = flow_node0(data0)

# Construct a FlowGraph. Here focuses on the options usage. For details about how to construct a DataFlow graph, see other sections.
options = {
    "ge.experiment.data_flow_deploy_info_path":"./data_flow_deploy_info.json"
}
dag = df.FlowGraph([out for out in flow_node0_out], options)

Format of the Configuration File

The following is an example of the format requirements of the typical configuration file data_flow_deploy_info.json:

       
        
          
          
            {
  
   "keep_logic_device_order": false,
   "batch_deploy_info": [
        {
            "flow_node_list": ["flowNode1", "flowNode2"],
            "logic_device_list": "0:0:-1:0"
        },
        {
            "flow_node_list": ["flowNode3"],
            "logic_device_list": "0:0:1:1"
 
        },
        {
            "flow_node_list": ["flowNode4"],
            "logic_device_list": "0:0:0:0,0:0:1:0",
            "invoke_list":[
            {
               "invoke_name":"invoked_flow_graph_name",
               "deploy_info_file":"./data_flow_invoke_flow_graph_deploy_info.json"
             },
            {
               "invoke_name":"invoked_flow_graph_name1",
               "logic_device_list": "0:0:0:0"
             }]
        },
        {
            "flow_node_list": ["flowNode5", "flowNode6"],
            "logic_device_list": "0:0:2~3:0~1"
        },
        {
            "flow_node_list": ["flowNode7", "flowNode8"],
            "logic_device_list": "0:0:0~1:0,0:0:2:1"
        }
    ]
}

           

         

       
      

**Table 1** Configuration options
Configuration Option	Description
keep_logic_device_order	Whether to deploy instances in the sequence specified by device_list. Options: true: In multi-instance scenarios, devices are sorted and deployed based on the user-configured device_list. false: In multi-instance scenarios, devices are sorted and deployed based on the internal implementation logic of the framework. Default value: false
batch_deploy_info
flow_node_list	List of FlowNode node names. One or more node names are supported. Use commas (,) to separate multiple node names.
logic_device_list	Deployment location of the DataFlow graph node. The format is clusterid:serverid:deviceid:numaid(pgid/dieid). The fields are described as follows: clusterid: cluster ID. Currently, the value is fixed to 0. serverid: server node ID. Set this parameter to the value of node_id in the numa_config.json file. The following is an example. { "cluster":[ { "cluster_nodes" : [ { "node_id" : 0, .......... deviceid: logical ID of a device. The value corresponds to the index of the item in item_list under node in the numa_config file specified by the RESOURCE_CONFIG_PATH environment variable. The value starts from 0. numaid (pgid/dieid): ID of multiple computing units on a single device. Typical scenarios are as follows: If heavy_load of FunctionPp is set to true, the graph node will be deployed on the host node corresponding to the specified node. When the graph node is deployed on multiple devices, perform configuration based on the following rules: Configure the node based on the devices. For example, 0:0:0:0,0:0:0:1 indicates that the node is deployed on devices 0:0:0:0 and 0:0:0:1 as multiple instances. Configure the node based on the device range. For example, 0:0:0~1:0~1, in which the tilde (~) indicates a range, the number before ~ must be less than or equal to that after ~. 0:0:0~1:0~1 indicates that the node is deployed on two computing units of two devices as multiple instances (0:0:0:0, 0:0:0:1, 0:0:1:0 and 0:0:1:1). Note: If FlowNode is the parent node of a sub-FlowGraph, multiple instances cannot be configured.
invoke_list	(Optional) If this option is not set, the subgraph deployment node is the same as the parent node deployment node. The following parameters belong to invoke_list.
invoke_name	Name of the InvokedClosure nested in the FlowNode, which corresponds to the name parameter of the AddInvokedClosure API.
deploy_info_file	Deployment location of the DataFlow subgraph node. If the InvokedClosure corresponding to invoke_name is a DataFlow subgraph, the deployment location file of the DataFlow subgraph can be configured, for example, ./data_flow_invoke_flow_graph_deploy_info.json. For details about the fields and format requirements in the file, see data_flow_deploy_info.json.
logic_device_list	Deployment location of a DataFlow or an AscendGraph subgraph node. The configuration method is the same as that of logic_device_list in batch_deploy_info. For the same subgraph node, logic_device_list and deploy_info_file cannot be configured at the same time. Otherwise, an error is reported. For an AscendGraph subgraph node, the number of instances in logic_device_list must be the same as that in logic_device_list of the parent node. Otherwise, an error is reported.

Parent topic: Special Topics