Loading…

Design, Implementation, and Evaluation of Host-based In-band Network Telemetry for TCP

In-band Network Telemetry (INT) is an innovative solution for capturing telemetry of the core network that has only recently been made possible due to the advent of programmable switches. INT data has been used very successfully to provide accurate core network conditions and perform TCP optimizatio...

Full description

Saved in:
Bibliographic Details
Main Authors: Sathyanarayana, Sandesh Dhawaskar, Hollingsworth, Max, Wu, Wenji, Cziva, Richard
Format: Conference Proceeding
Language:English
Subjects:
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In-band Network Telemetry (INT) is an innovative solution for capturing telemetry of the core network that has only recently been made possible due to the advent of programmable switches. INT data has been used very successfully to provide accurate core network conditions and perform TCP optimization for larger data throughput or lower latency. Unfortunately, there are a number of drawbacks. First, the particular network conditions that the switches monitor is statically controlled and this can result in significant bandwidth wastage. Second, telemetry data is processed within the core network and the end-host is subject to synchronization delays to access the telemetry data. These delays reduce the efficiency of real-time tuning by the end-host. And third, the network telemetry only covers the core of the network and this leaves out last mile telemetry which in various deployments, especially mobile networks, is crucial for the success of TCP optimization. We identify this gap and, for the first time, design and develop a kernel-based infrastructure and eBPF-based APIs for INT that provide flexible telemetry gathering, real-time access to data for end-hosts, and includes last-mile telemetry. This is done by addressing significant challenges to handle TCP packet modification at the end-host. Our developed framework has only 1.73% and 1.08% performance overhead in public and controlled networks. Our system gives a new dimension for TCP performance optimization that can now utilize the last mile telemetry and will prevent network operators from doing manual TCP tuning.
ISSN:2150-329X
DOI:10.1109/GIIS56506.2022.9937001