Hybrid TPU Thread Scheduling via Hardware Feedback

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional operating systems struggle to optimize thread management in data processing systems with heterogeneous processor topologies, leading to inefficiencies in thread scheduling and resource utilization.

Innovation Solution

A single version of the operating system is developed with logic to detect processor topology, analyze hardware feedback, and manage threads based on these factors, enabling optimized thread scheduling across different types of processing cores, including big and small cores with varying power consumption and multithreading capabilities.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If a conventional operating system with homogenous thread scheduling algorithms is used, then the system is simple to implement, but thread scheduling efficiency deteriorates in heterogeneous processor topologies

Engineering Contradiction:
Improveease of implementationVSAvoidthread scheduling efficiency
Core Design Contradiction:
Ease of manufactureVSProductivity

Solution Approach 1:

The thread scheduling algorithm dynamically adapts to the heterogeneous processor topology by detecting core types (performance cores vs. efficiency cores) and adjusting scheduling decisions in real-time. The OS modifies its behavior based on runtime conditions, selecting appropriate scheduling strategies for different thread types and processor configurations, thereby resolving the contradiction between implementation simplicity and scheduling efficiency.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

Different scheduling algorithms are applied to different parts of the system: performance-oriented scheduling for performance cores, power-efficient scheduling for efficiency cores, and specialized handling for SMT threads. This localized approach allows each processor type to be optimized independently while maintaining overall system simplicity through a unified scheduling framework.

Inventive Principle:
Principle #3Local quality

2Device complexity

If threads are scheduled without considering processor topology, then the scheduling algorithm is simple, but power consumption increases and performance decreases

Engineering Contradiction:
Improvescheduling algorithm complexityVSAvoidpower consumption
Core Design Contradiction:
Device complexityVSLoss of energy

Solution Approach 1:

The operating system performs preliminary detection and classification of processor topology during system initialization, identifying performance cores, efficiency cores, and SMT capabilities before thread scheduling begins. This advance knowledge allows the scheduler to make informed decisions without adding complex real-time analysis, reducing both power consumption and algorithmic complexity.

Inventive Principle:
Principle #10Preliminary action

3Ease of operation

If threads are scheduled without considering processor topology, then the system is easier to operate, but system performance and efficiency deteriorate

Engineering Contradiction:
Improvesystem operabilityVSAvoidsystem performance
Core Design Contradiction:
Ease of operationVSProductivity

Solution Approach 1:

The operating system automatically detects and adapts to the heterogeneous processor topology without requiring user intervention or manual configuration. The scheduler self-adjusts based on detected core types and thread characteristics, maintaining ease of operation while achieving optimized performance through automated topology-aware scheduling decisions.

Inventive Principle:
Principle #25Self-service

4Device complexity

If hardware feedback is not utilized, then the system complexity is lower, but thread configuration optimization is insufficient

Engineering Contradiction:
Improvesystem complexityVSAvoidthread configuration optimization
Core Design Contradiction:
Device complexityVSManufacturing precision

Solution Approach 1:

The system incorporates hardware feedback mechanisms that provide runtime performance and power consumption data to the scheduler. This feedback enables continuous optimization of thread configuration by comparing actual performance against targets and adjusting scheduling decisions accordingly, achieving precise optimization without excessive system complexity through efficient feedback processing.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3929743B1Technology for optimizing hybrid processor utilization
Publication Date: 2024.07.03 INTEL CORP
  • EP3929743B1 patent drawingFigure 1
  • EP3929743B1 patent drawingFigure 2
  • EP3929743B1 patent drawingFigure 3A

AI summary

A data processing system comprises a hybrid processor comprising a big TPU and a small TPU. At least one of the TPUs comprises an LP of a processing core that supports SMT. The hybrid processor further comprises hardware feedback circuitry. A machine-readable medium in the data processing system comprises instructions which, when executed, enable an OS in the data processing system to collect (a) processor topology data from the hybrid processor and (b) hardware feedback for at least one of the TPUs from the hardware feedback circuitry. The instructions also enable the OS to respond to a determination that a thread is ready to be scheduled by utilizing (a) an OP setting for the ready thread, (b) the processor topology data, and (c) the hardware feedback to make a scheduling determination for the ready thread. Other embodiments are described and claimed.