ML Workflow Graph Partitioning for Distributed Scheduling

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional machine learning applications face inefficiencies in distributed execution due to the minimum loop connectedness problem, which affects the scheduling of irreducible partitioned graphs in telecommunications systems, leading to suboptimal run-time performance and convergence rates.

Innovation Solution

The solution involves generating partitioned graphs that can be scheduled in a distributed fashion by determining whether the graph is reducible or irreducible, using Depth First Search Trees to compute an optimal execution order, and selecting algorithms based on topological properties to minimize loop connectedness, thereby enabling efficient distribution and scheduling of machine learning applications.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If conventional schedulers are used for distributed machine learning applications, then the system structure is simple, but the convergence rate and run-time performance are suboptimal due to the minimum loop connectedness problem

Engineering Contradiction:

Improveconvergence rateVSAvoidscheduling complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The patent segments the scheduling problem into two distinct cases: reducible graphs and irreducible graphs. For reducible graphs, a simple distributed scheduling approach is used, while for irreducible graphs, a more complex optimal scheduling algorithm is applied. This segmentation allows the system to achieve high convergence rates without unnecessarily applying complex scheduling to all cases.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent changes the scheduling parameters based on the graph type. For reducible graphs, default scheduling parameters are used, while for irreducible graphs, optimized parameters are applied to minimize loop connectedness. This parameter adaptation enables the system to achieve optimal convergence rates without universally increasing complexity.

Inventive Principle:

Principle #35Parameter changes

2Productivity

If optimal scheduling is computed for irreducible partitioned graphs, then run-time performance improves by up to 50%, but the computational overhead and scheduling complexity increases

Engineering Contradiction:

Improverun-time performanceVSAvoidscheduling computation time

Core Design Contradiction:

ProductivityVSLoss of time

Solution Approach 1:

The patent performs preliminary classification of the partitioned graph to determine whether it is reducible or irreducible before applying the scheduling algorithm. This preliminary action allows the system to avoid unnecessary complex scheduling computations for reducible graphs, thereby reducing computational overhead while maintaining optimal performance for irreducible graphs.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent applies the complex optimal scheduling algorithm only partially - specifically, only to the irreducible case. For reducible graphs, a simpler scheduling approach is used. This partial application of the complex algorithm reduces the overall computational overhead while still achieving the 50% performance improvement where it is most needed.

Inventive Principle:

Principle #16Partial or excessive action

3Adaptability or versatility

If distributed scheduling is implemented without central coordination, then system scalability improves, but handling irreducible graphs becomes more difficult

Engineering Contradiction:

Improvedistributed execution capabilityVSAvoidscheduling ease

Core Design Contradiction:

Adaptability or versatilityVSEase of operation

Solution Approach 1:

The patent introduces an intermediary classification mechanism that determines whether a graph is reducible or irreducible. This intermediary step enables distributed scheduling to handle irreducible graphs effectively by identifying them and applying appropriate scheduling strategies, thereby maintaining both scalability and ease of operation.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent applies different scheduling qualities to different parts of the problem space. For reducible graphs, simple distributed scheduling is used, while for irreducible graphs, optimal scheduling is applied. This local differentiation allows the system to maintain ease of operation for the common case while handling the difficult case effectively.

Inventive Principle:

Principle #3Local quality

Data Source

PatentUS12081636B2Distribution of machine learning workflows on webscale infrastructures

Publication Date: 2024.09.03 NOKIA SOLUTIONS & NETWORKS OY

AI summary

A network device includes processing circuitry configured to cause the network device to: partition a flow graph for an application to generate a partitioned graph of nodes and edges, each of the nodes including computations mapped to an execution unit to execute at least a portion of the application, and each of the edges denoting communications between execution units; determine whether the partitioned graph is an irreducible graph; and schedule the computations and the communications for execution in response to determining that the partitioned graph is an irreducible graph.