What question did this study set out to answer?

This research aims to compare the performance of different TCP congestion control algorithms across various application workloads in cloud environments.

July 12, 2026Open Access

Performance evaluation of TCP congestion control variants across application workloads in cloud based networks

ADAravinda Swamy D.AHAnanya HegdeUniversidad del Noreste AMAnantha Chary C. M.

Key Points

This research aims to compare the performance of different TCP congestion control algorithms across various application workloads in cloud environments.
Evaluated three TCP variants: Cubic, Reno, and BBR.
Conducted experiments using four workload types: synthetic testing, real-time stream processing, distributed I/O, and compute-intensive sorting.
Used a dumbbell network topology with AWS EC2 instances to create controlled conditions.
Cubic showed superior throughput in synthetic testing versus BBR and Reno, with higher metrics.
Reno exhibited lower end-to-end latency in real-time stream processing workloads, yielding quicker response times.
BBR resulted in fewer retransmissions in distributed I/O, outperforming others in that regard.

Abstract

Choosing the right Transmission Control Protocol (TCP) congestion control algorithm matters more in shared cloud environments than is often appreciated, yet head-to-head comparisons across heterogeneous, concurrently running workloads remain rare in the literature. We evaluated three widely used variants Cubic, Reno, and Bottleneck Bandwidth and Round-trip propagation time (BBR) under four workload types: synthetic throughput testing with iperf3, real-time stream processing using Apache Kafka and Apache Flink, distributed I/O, and compute-intensive sorting via Hadoop TeraSort. All experiments ran inside a dumbbell network topology built on four Amazon Web Services (AWS) m7i-flex.large Elastic Compute Cloud (EC2) instances, with a dedicated forwarding node and static routing forcing every flow through that single contention point. Across every metric we measured throughput, end-to-end latency, retransmission count, and job completion time the three variants behaved quite differently depending on the workload. The results give concrete, workload-specific guidance for operators choosing a congestion control policy in multitenant cloud deployments.

Ask AI

Helpful

Bookmark

View Full Paper