Now think of your graph as the base widget. Now imagine its opaque but defines common interfaces, this is the level of focus for the next layer. Its not concerned with a specific computation flow as much as scaling the graphs out.
The first thing to do for scaling is to realize that you have a template that created the dynamic graph in step one. Once you can create one there is virtually no limit to how many you can have or have small you can slice up the work segments. With the data DSL you make sure that data locality is optimized. With the scheduling DSL, you bring new realities into existence based on seed information.
Note, the actual complexity contained in each self contained graph is abstracted at this layer. Its just a widget that can be replicated.
Life coexists, the rules for your graph should be discoverable in some registry.
No comments:
Post a Comment