Adaptive Parallelism for Accelerator Workload Distribution

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

In clustered computing, existing methods fail to efficiently distribute workloads between computers with varying computational accelerators, such as CPUs and GPUs, leading to suboptimal processing times due to static and dynamic characteristic changes.

Innovation Solution

A computing unit that receives information about the configuration of computational accelerators, including static and dynamic characteristics, divides workloads into subportions and assigns them to appropriate computers for execution, adapting parallelism based on these characteristics to optimize processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If static workload distribution methods are used, then system simplicity is maintained, but processing efficiency deteriorates due to inability to adapt to dynamic conditions

Engineering Contradiction:

Improveprocessing efficiencyVSAvoidworkload management complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The patent implements dynamic workload distribution by continuously monitoring accelerator utilization metrics and adjusting workload allocation in real-time based on current system conditions, transforming the static scheduling approach into an adaptive system that responds to changing computational demands

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system incorporates feedback mechanisms that collect utilization data from accelerators and use this information to inform subsequent workload distribution decisions, creating a closed-loop control system that optimizes processing efficiency through continuous adaptation

Inventive Principle:

Principle #23Feedback

2Productivity

If workload is divided into more subportions, then accelerator utilization improves, but communication overhead increases

Engineering Contradiction:

Improveaccelerator utilizationVSAvoidcommunication overhead time

Core Design Contradiction:

ProductivityVSLoss of time

Solution Approach 1:

The patent dynamically adjusts the granularity of workload subdivision based on accelerator type and current utilization conditions, modifying the parameter of subportion size to optimize the balance between utilization efficiency and communication overhead for different computational scenarios

Inventive Principle:

Principle #35Parameter changes

3Adaptability or versatility

If adaptive parallelism is implemented, then resource utilization optimizes, but system complexity increases

Engineering Contradiction:

Improveresource adaptation capabilityVSAvoidscheduling system complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system adapts parallelism parameters dynamically by adjusting the degree of workload parallelization based on accelerator characteristics and current system state, enabling flexible resource utilization without requiring complete system redesign for different scenarios

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS10203988B2Adaptive parallelism of task execution on machines with accelerators

Publication Date: 2019.02.12 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

A computer system, method, and computer readable product are provided for adaptive parallelism of workload execution on computers with accelerators. In various embodiments, information about both static and dynamic characteristics of computational accelerators for a plurality of computers is received. Based on this information, waves of a workload is divided among this plurality of computers for processing. As the dynamic characteristics of those computational accelerators change over time, future waves may be divided among this plurality of computers differently.