Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SGPRS: Seamless GPU Partitioning Real-Time Scheduler for Periodic Deep Learning Workloads

Amir Fakhim Babaei, Thidapat Chantem

TL;DR

This work conducts GPU speedup gain analysis and proposes SGPRS, the first real-time GPU scheduler considering zero configuration partition switch, which not only meets more deadlines for parallel tasks but also sustains overall performance beyond the pivot point.

Abstract

Deep Neural Networks (DNNs) are useful in many applications, including transportation, healthcare, and speech recognition. Despite various efforts to improve accuracy, few works have studied DNN in the context of real-time requirements. Coarse resource allocation and sequential execution in existing frameworks result in underutilization. In this work, we conduct GPU speedup gain analysis and propose SGPRS, the first real-time GPU scheduler considering zero configuration partition switch. The proposed scheduler not only meets more deadlines for parallel tasks but also sustains overall performance beyond the pivot point.

SGPRS: Seamless GPU Partitioning Real-Time Scheduler for Periodic Deep Learning Workloads

TL;DR

This work conducts GPU speedup gain analysis and proposes SGPRS, the first real-time GPU scheduler considering zero configuration partition switch, which not only meets more deadlines for parallel tasks but also sustains overall performance beyond the pivot point.

Abstract

Deep Neural Networks (DNNs) are useful in many applications, including transportation, healthcare, and speech recognition. Despite various efforts to improve accuracy, few works have studied DNN in the context of real-time requirements. Coarse resource allocation and sequential execution in existing frameworks result in underutilization. In this work, we conduct GPU speedup gain analysis and propose SGPRS, the first real-time GPU scheduler considering zero configuration partition switch. The proposed scheduler not only meets more deadlines for parallel tasks but also sustains overall performance beyond the pivot point.
Paper Structure (12 sections, 4 figures)

This paper contains 12 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Task Model
  3. GPU Speedup Analysis
  4. Real-Time Scheduler Design
  5. Offline Phase
  6. Two Level Priority Assignment
  7. WCET and Virtual Deadline Assignment
  8. Online Phase
  9. Absolute Deadline Assignment
  10. Context Assignment
  11. Stage Queuing
  12. Simulation Results

Figures (4)

  • Figure 1: Speedup gain for different operations when running in isolation
  • Figure 2: SGPRS Overview
  • Figure 3: Result when using 2 contexts (Scenario 1)
  • Figure 4: Result when using 3 contexts (Scenario 2)