Shisha: Online Scheduling of CNN Pipelines on Heterogeneous Architectures

Publikation: Beitrag in Buch/Konferenzbericht/Sammelband/GutachtenBeitrag in KonferenzbandBeigetragenBegutachtung

Beitragende

Abstract

Many modern multicore processors integrate asymmetric core clusters. With the trend towards Multi-Chip-Modules (MCMs) and interposer-based packaging technologies, platforms will feature heterogeneity at the level of cores, memory subsystem and the interconnect. Due to their potential high memory throughput and energy efficient core modules, these platforms are prominent targets for emerging machine learning applications, such as Convolutional Neural Networks (CNNs). To exploit and adapt to the diversity of modern heterogeneous chips, CNNs need to be quickly optimized in terms of scheduling and workload distribution among computing resources. To address this we propose Shisha, an online approach to generate and schedule parallel CNN pipelines on heterogeneous MCM-based architectures. Shisha targets heterogeneity in compute performance and memory bandwidth and tunes the pipeline schedule through a fast online exploration technique. We compare Shisha with Simulated Annealing, Hill Climbing and Pipe-Search. On average, the convergence time is improved by ∼ 35 × in Shisha compared to other exploration algorithms. Despite the quick exploration, Shisha’s solution is often better than that of other heuristic exploration algorithms.

Details

OriginalspracheEnglisch
TitelParallel Processing and Applied Mathematics - 14th International Conference, PPAM 2022, Revised Selected Papers
Redakteure/-innenRoman Wyrzykowski, Jack Dongarra, Ewa Deelman, Konrad Karczewski
Herausgeber (Verlag)Springer Science and Business Media B.V.
Seiten249-262
Seitenumfang14
ISBN (Print)9783031304415
PublikationsstatusVeröffentlicht - 2023
Peer-Review-StatusJa

Publikationsreihe

ReiheLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Band13826 LNCS
ISSN0302-9743

Konferenz

Titel14th International Conference on Parallel Processing and Applied Mathematics, PPAM 2022
Dauer11 - 14 September 2022
StadtGdansk
LandPolen

Externe IDs

ORCID /0000-0002-5007-445X/work/160049123

Schlagworte

Schlagwörter

  • CNN parallel pipelines, Design space exploration, Online tuning, Processing on chiplets, Processing on heterogeneous computing units