OLAF: Programmable Data Plane Acceleration for Asynchronous Distributed Reinforcement Learning

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Krishna, Nehal Baganal, Tahir, Anam, Khamis, Firas, Arashloo, Mina Tahmasbi, Zink, Michael, Rizk, Amr
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866913932281643008
author Krishna, Nehal Baganal
Tahir, Anam
Khamis, Firas
Arashloo, Mina Tahmasbi
Zink, Michael
Rizk, Amr
author_facet Krishna, Nehal Baganal
Tahir, Anam
Khamis, Firas
Arashloo, Mina Tahmasbi
Zink, Michael
Rizk, Amr
contents Asynchronous Distributed Reinforcement Learning (DRL) can suffer from degraded convergence when model updates become stale, often the result of network congestion and packet loss during large-scale training. This work introduces a network data-plane acceleration architecture that mitigates such staleness by enabling inline processing of DRL model updates as they traverse the accelerator engine. To this end, we design and prototype a novel queueing mechanism that opportunistically combines compatible updates sharing a network element, reducing redundant traffic and preserving update utility. Complementing this we provide a lightweight transmission control mechanism at the worker nodes that is guided by feedback from the in-network accelerator. To assess model utility at line rate, we introduce the Age-of-Model (AoM) metric as a proxy for staleness and verify global fairness and responsiveness properties using a formal verification method. Our evaluations demonstrate that this architecture significantly reduces update staleness and congestion, ultimately improving the convergence rate in asynchronous DRL workloads.
format Preprint
id arxiv_https___arxiv_org_abs_2507_05876
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle OLAF: Programmable Data Plane Acceleration for Asynchronous Distributed Reinforcement Learning
Krishna, Nehal Baganal
Tahir, Anam
Khamis, Firas
Arashloo, Mina Tahmasbi
Zink, Michael
Rizk, Amr
Networking and Internet Architecture
Hardware Architecture
Asynchronous Distributed Reinforcement Learning (DRL) can suffer from degraded convergence when model updates become stale, often the result of network congestion and packet loss during large-scale training. This work introduces a network data-plane acceleration architecture that mitigates such staleness by enabling inline processing of DRL model updates as they traverse the accelerator engine. To this end, we design and prototype a novel queueing mechanism that opportunistically combines compatible updates sharing a network element, reducing redundant traffic and preserving update utility. Complementing this we provide a lightweight transmission control mechanism at the worker nodes that is guided by feedback from the in-network accelerator. To assess model utility at line rate, we introduce the Age-of-Model (AoM) metric as a proxy for staleness and verify global fairness and responsiveness properties using a formal verification method. Our evaluations demonstrate that this architecture significantly reduces update staleness and congestion, ultimately improving the convergence rate in asynchronous DRL workloads.
title OLAF: Programmable Data Plane Acceleration for Asynchronous Distributed Reinforcement Learning
topic Networking and Internet Architecture
Hardware Architecture
url https://arxiv.org/abs/2507.05876