What is it about?
Modern cloud applications rely on storage systems to save and retrieve data quickly and reliably. Longhorn is a popular open-source tool that manages storage in Kubernetes — the platform most organizations use to run cloud applications. While Longhorn is easy to set up and widely used, it was not designed to take full advantage of today's fast hardware, such as high-speed NVMe solid-state drives and ultra-fast network connections. This research identifies the key bottlenecks slowing Longhorn down and introduces three targeted improvements to its core storage engine. First, the way Longhorn talks to the operating system was replaced with a modern, more efficient method called ublk, which dramatically reduces delays. Second, the internal communication between the storage controller and its data replicas was redesigned to allow many requests to be handled at the same time, instead of one after another. Third, a new lightweight storage layer called Direct Block Store (DBS) was built to manage data and snapshots more efficiently than the previous approach using sparse files. Together, these changes deliver up to ten times more storage operations per second and significantly higher data transfer speeds, all while keeping Longhorn's simple, familiar design intact. The improved system was also compared to Mayastor, a high-performance storage engine known for its speed, and matched or exceeded it in several important tests.
Featured Image
Photo by imgix on Unsplash
Why is it important?
Storage performance is a silent bottleneck in many cloud and on-premises deployments. This work shows that targeted, pragmatic engineering — rather than a full system rewrite — can unlock an order-of-magnitude improvement in a widely deployed open-source storage system. The results directly benefit organizations running data-intensive workloads such as machine learning, video analytics, and high-transaction databases on Kubernetes. The improvements have been submitted as pull requests to the upstream Longhorn project, meaning the broader community of users can benefit. The paper also establishes ublk as a viable, high-performance frontend for cloud-native storage systems — a reference that other projects can build on.
Perspectives
Cloud infrastructure is evolving rapidly, and storage is struggling to keep up. Hardware is getting faster — NVMe drives and high-speed networks are now commonplace even in mid-range servers — but the software managing that hardware often lags behind. This work is a reminder that the gap between hardware capability and software utilization is a real and solvable problem, and that open-source communities benefit enormously from performance-focused contributions that respect the existing design. Looking ahead, the principles demonstrated here — minimizing data copies, eliminating serialization bottlenecks, and using modern OS interfaces like io_uring — are broadly applicable across the cloud-native storage ecosystem. As Kubernetes continues to grow as the dominant platform for running workloads at scale, the demand for storage solutions that are both easy to operate and genuinely high-performance will only increase.
Konstantinos Kampadais
Foundation for Research and Technology - Hellas (FORTH)
Read the Original
This page is a summary of: Optimizing Longhorn for High Performance Cloud-native Storage, April 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3805687.3806255.
You can read the full text:
Resources
Contributors
The following have contributed to this page
