Reinforcement Learning for Runtime Optimization for High Performance and Energy Efficient NoC

Md Farhadur Reza

doi:10.1109/ISOCC53507.2021.9613994

Reinforcement Learning for Runtime Optimization for High Performance and Energy Efficient NoC

Md Farhadur Reza

Eastern Illinois University

Research output: Contribution to journal › Article › peer-review

Abstract

 Hundreds to thousands of cores are now being integrated on a single chip because of the transistor technology
 advancement. Network-on-Chip (NoC) has been the superior interconnect fabric for multi/many-core systems. As applications with unknown demands can arrive in the system and/or applications’ demands can change at runtime, it becomes essential to configure NoC dynamically for energy efficiency and high performance. However, large and complex NoC design space introduces computational and timing complexities to explore optimal
 trade-offs of energy-performance. Therefore, reinforcement learning (RL) techniques are proposed to predict and configure NoC resources quickly based on the learning of the system utilization and applications workloads. Multiple RL solutions with different reward signals, function approximators, and policies are discussed for runtime optimization of NoC. Results have been demonstrated to show the effectiveness of a few RL techniques.

Original language	American English
Journal	18th International SoC Design Conference (ISOCC)
DOIs	https://doi.org/10.1109/ISOCC53507.2021.9613994
State	Published - Oct 1 2021

Keywords

Reinforcement Learning
Networks-on-Chip
Multicore Architectures

Disciplines

Computer Engineering
Computer and Systems Architecture
Digital Communications and Networking
Computer Sciences

Access to Document

10.1109/ISOCC53507.2021.9613994License: Unspecified

ISOCC_Paper_2021_IEEE_Final.pdfFinal published version, 665 KBLicense: Unspecified

Cite this

@article{f189bca94ea84771be761bddc57722a1,

title = "Reinforcement Learning for Runtime Optimization for High Performance and Energy Efficient NoC",

abstract = " Hundreds to thousands of cores are now being integrated on a single chip because of the transistor technology advancement. Network-on-Chip (NoC) has been the superior interconnect fabric for multi/many-core systems. As applications with unknown demands can arrive in the system and/or applications{\textquoteright} demands can change at runtime, it becomes essential to configure NoC dynamically for energy efficiency and high performance. However, large and complex NoC design space introduces computational and timing complexities to explore optimal trade-offs of energy-performance. Therefore, reinforcement learning (RL) techniques are proposed to predict and configure NoC resources quickly based on the learning of the system utilization and applications workloads. Multiple RL solutions with different reward signals, function approximators, and policies are discussed for runtime optimization of NoC. Results have been demonstrated to show the effectiveness of a few RL techniques.",

keywords = "Reinforcement Learning, Networks-on-Chip, Multicore Architectures",

author = "Reza, {Md Farhadur}",

year = "2021",

month = oct,

day = "1",

doi = "10.1109/ISOCC53507.2021.9613994",

language = "American English",

journal = "18th International SoC Design Conference (ISOCC)",

}

TY - JOUR

T1 - Reinforcement Learning for Runtime Optimization for High Performance and Energy Efficient NoC

AU - Reza, Md Farhadur

PY - 2021/10/1

Y1 - 2021/10/1

N2 - Hundreds to thousands of cores are now being integrated on a single chip because of the transistor technology advancement. Network-on-Chip (NoC) has been the superior interconnect fabric for multi/many-core systems. As applications with unknown demands can arrive in the system and/or applications’ demands can change at runtime, it becomes essential to configure NoC dynamically for energy efficiency and high performance. However, large and complex NoC design space introduces computational and timing complexities to explore optimal trade-offs of energy-performance. Therefore, reinforcement learning (RL) techniques are proposed to predict and configure NoC resources quickly based on the learning of the system utilization and applications workloads. Multiple RL solutions with different reward signals, function approximators, and policies are discussed for runtime optimization of NoC. Results have been demonstrated to show the effectiveness of a few RL techniques.

AB - Hundreds to thousands of cores are now being integrated on a single chip because of the transistor technology advancement. Network-on-Chip (NoC) has been the superior interconnect fabric for multi/many-core systems. As applications with unknown demands can arrive in the system and/or applications’ demands can change at runtime, it becomes essential to configure NoC dynamically for energy efficiency and high performance. However, large and complex NoC design space introduces computational and timing complexities to explore optimal trade-offs of energy-performance. Therefore, reinforcement learning (RL) techniques are proposed to predict and configure NoC resources quickly based on the learning of the system utilization and applications workloads. Multiple RL solutions with different reward signals, function approximators, and policies are discussed for runtime optimization of NoC. Results have been demonstrated to show the effectiveness of a few RL techniques.

KW - Reinforcement Learning

KW - Networks-on-Chip

KW - Multicore Architectures

U2 - 10.1109/ISOCC53507.2021.9613994

DO - 10.1109/ISOCC53507.2021.9613994

M3 - Article

JO - 18th International SoC Design Conference (ISOCC)

JF - 18th International SoC Design Conference (ISOCC)

ER -