ML Workload Composition Using Aggregator Node Hierarchies
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Running a machine learning workload is complex due to its tight coupling with infrastructure resource management and embedding within machine learning libraries, leading to time-consuming and error-prone source code updates for different types of workloads.
Innovation Solution
A device receives definition data for a machine learning workload represented as roles and channels, identifying groups of training nodes, selecting intermediate aggregator nodes, and provisioning the workload by configuring nodes and channels to simplify the composition process, decoupling AI/ML algorithms from infrastructure building blocks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning workload is tightly coupled with infrastructure resource management and embedded into machine learning library, then the workload can be executed on existing infrastructure, but the complexity of expressing and managing the workload increases significantly
Solution Approach 1:
The patent segments the machine learning workload into distinct components: data planes (training nodes, aggregator nodes), control planes (workload composition language, provisioning system), and communication channels. This segmentation allows each component to be independently defined and managed, reducing overall complexity while maintaining execution capability.
Solution Approach 2:
The patent introduces an intermediary workload composition language that sits between the user's high-level workload definition and the underlying infrastructure. This intermediary layer translates abstract workload descriptions into concrete infrastructure configurations, shielding users from complexity while ensuring proper resource management.
2Adaptability or versatility
If source code and configuration updates are performed in an ad-hoc fashion for different machine learning workloads, then flexibility to accommodate different workloads is achieved, but time consumption and error rates increase
Solution Approach 1:
The patent creates a universal workload composition language that can express multiple types of machine learning workloads through a consistent set of constructs (data planes, channels, control logic). This universal approach eliminates the need for ad-hoc source code modifications for different workload types, as the same language framework handles all workload variations.
Solution Approach 2:
The patent performs preliminary action by providing a structured framework and language for workload composition before actual workload execution. Users define workloads using predefined patterns and constructs, which are then automatically provisioned and configured, eliminating the need for reactive ad-hoc updates during workload deployment.
3Adaptability or versatility
If machine learning workload expression is embedded into the machine learning library, then integration with existing libraries is achieved, but the ease of operation for users decreases due to required source code updates
Solution Approach 1:
The patent extracts the workload expression functionality from the machine learning library embedding and places it in a separate, dedicated workload composition language. This extraction allows users to define workloads independently of library source code, eliminating the need to modify library code while maintaining full integration capability through the language's provisioning mechanisms.
Data Source
AI summary
In one embodiment, a device receives, via a user interface, definition data for a machine learning workload. The device identifies groups of training nodes in the network that store training datasets, to perform training roles for the machine learning workload by training machine learning models on their respective training datasets. The device selects a set of intermediate aggregator nodes for the groups of training nodes to aggregate their models. The device provisions the machine learning workload by configuring the groups of training nodes, the set of intermediate aggregator nodes, and a global aggregator node for the set of intermediate aggregator nodes and by configuring channels between training nodes in a group, between the groups of training nodes and the set of intermediate aggregator nodes, and between the set of intermediate aggregator nodes and the global aggregator node.