Institutional Repository of Vilnius University
Investigating the impact of linux file systems on object storage performance /
Abstract
dc:descriptionS3-compatible object storage is commonly deployed on Linux clusters, but the local storage backend (filesystem or raw-block design) is often chosen by convention rather than validated under the expected concurrency and object size mix. This thesis measures how the local storage backend influences throughput, latency, and saturation characteristics of two open-source S3 implementations under controlled benchmarks. The work analyses two open-source object storage solutions: MinIO and Ceph RADOS Gateway (RGW). MinIO stores objects on a local filesystem, so filesystem choice directly affects it. Ceph RGW is tested with two backends: FileStore (filesystem-backed) and BlueStore (raw-block backend to give analysis perspective). For filesystem-based setups, four Linux filesystems are compared: EXT4, XFS, BTRFS, and ZFS. Experiments are performed on a four-node cluster. Workloads are generated using MinIO Warp and include PUT and GET operations, mixed read/write workloads, object-size scaling (from small to larger objects), concurrency scaling (increasing the number of clients), and namespace growth up to 10,000 objects. Across tests, the environment is kept the same, and only the local storage backend is changed, so the differences can be attributed mainly to the backend choice. Results show that backend choice has the clearest impact under high load, especially with mixed traffic and higher concurrency. At low load, many configurations behave similarly, while at higher concurrency, throughput can stay the same as latency continues to increase. This shows saturation and queuing. Small objects are more sensitive to metadata and per-request overhead, while larger objects place greater stress on the storage configuration. The comparison of Ceph BlueStore and FileStore indicates that a filesystem-backed design can introduce measurable overhead compared to a raw-block approach. However, raw-block designs still show limits in the designed configuration. In this test environment, MinIO shows smaller performance differences across filesystems than Ceph. Based on these observations, backend selection is best checked by running workload-representative benchmarks, particularly mixed workloads and concurrency scaling. The emphasis is on latency stability under load rather than peak throughput alone. Future work should evaluate larger namespaces and separate cold-cache vs warm-cache behaviour more thoroughly.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Narbutas, Renaldas,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD277691089&prefLang=en_US