Common Flow Development Patterns

Reducing the number of tasks to be used

Assume that we want to run the following flow in a machine that has 4 CPU cores:

from videoflow.producers import IntProducer
from videoflow.processors import IdentityProcessor
from videoflow.consumers.import CommandlineConsumer

A = IntProducer(0, 40, fps = 30)
B = IdentityProcessor(fps = 300)(A)
C = IdentityProcessor(fps = 300)(B)
D = IdentityProcessor(fps = 300)(C)
E = IdentityProcessor(fps = 300)(D)
F = IdentityProcessor(fps = 300)(E)
G = CommandlineConsumer()(F)

The execution engine will take this graph of 7 nodes, and will create 7 corresponding tasks, each one running on its independent process.

Usually, the consumers and producers are I/O bound tasks, while the processors are CPU bound tasks. If you have more CPU bound tasks than CPU physical cores in your system, this can introduced some unwanted or unnecessary synchronization overhead. One way to solve this problem is to group multiple nodes under the same task. In order to do that, we can use the TaskModuleNode.

The TaskModuleNode wraps a subgraph of processor nodes, and has the effect that the execution engine allocates only one task to execute this subgraph. Using the TaskModuleNode the previous example would look like:

from videoflow.producers import IntProducer
from videoflow.processors import IdentityProcessor
from videoflow.consumers.import CommandlineConsumer
from videoflow.core.node import TaskModuleNode

A = IntProducer(0, 40, fps = 30)
B = IdentityProcessor(fps = 300)(A)
C = IdentityProcessor(fps = 300)(B)
D = IdentityProcessor(fps = 300)(C)
E = IdentityProcessor(fps = 300)(D)
F = IdentityProcessor(fps = 300)(E)
task_module = TaskModuleNode(entry_node = B, exit_node = F)
G = CommandlineConsumer()(task_module)

Note

There are certain restrictions in the usage of the TaskModuleNode:

• All nodes on its subgraph must be processor nodes.

• No node on its subgraph can have device_type to be GPU

• There cannot be a TaskModuleNode nested inside a TaskModuleNode.

Notice that now nodes A, B, C, D, E and F are not executed in parallel, but sequentially, and that effectively reduces the fps of the flow in that section of the graph.

Warning

You need to be careful when using TaskModuleNode. You can effectively introduce bottlenecks in the flow if you are not careful.

The picture below contains an example of such a case:

Grouping subgraphs in modules