Uploaded November 2025 | Updated September 2026, 3 weeks ago
Efficient Ceph Performance Troubleshooting in Production Using eBPF - Dongdong Tao, Canonical
Dongdong Tao will introduce some powerful eBPF-based tools he developed for analyzing Ceph performance in production environments.
One of the key tools, radostrace, enables users to trace every I/O request sent from Ceph clients to OSDs. It provides deep visibility into each individual I/O request, including the target OSDs, object names, operation types, and round-trip latency. With these low-level per IO data collected via eBPF, it's possible to accurately identify underperforming OSDs in large-scale production clusters and understand which types of requests are degrading performance.
In this session, Dongdong will showcase real-world production performance issues and demonstrate how these tools enabled him to resolve them more efficiently than ever before.
For instance, bug tracker.ceph.com/issues/69014, radostrace helped pinpoint that the bottleneck in the slow rbd sparsify operation was due to the extent comparison phase. And by optimizing the comparison function from O(M*N) to O(M), we see a 60x performance improvement.
The tools are open-sourced and available at: github.com/taodd/cephtrace
Efficient Ceph Performance Troubleshooting in Production Using eBPF - Dongdong Tao, Canonical
Dongdong Tao will introduce some powerful eBPF-based tools he developed for analyzing Ceph performance in production environments.
One of the key tools, radostrace, enables users to trace every I/O request sent from Ceph clients to OSDs. It provides deep visibility into each individual I/O request, including the target OSDs, object names, operation types, and round-trip latency. With these low-level per IO data collected via eBPF, it's possible to accurately identify underperforming OSDs in large-scale production clusters and understand which types of requests are degrading performance.
In this session, Dongdong will showcase real-world production performance issues and demonstrate how these tools enabled him to resolve them more efficiently than ever before.
For instance, bug tracker.ceph.com/issues/69014, radostrace helped pinpoint that the bottleneck in the slow rbd sparsify operation was due to the extent comparison phase. And by optimizing the comparison function from O(M*N) to O(M), we see a 60x performance improvement.
The tools are open-sourced and available at: github.com/taodd/cephtrace
