Ceph Client librbd Performance Analysis and Learnings - Mahati Chamarthy
0 likes159 views
This document summarizes an analysis of Ceph client (librbd) performance. It discusses RBD features, tracing and profiling tools used in testing, the test environment setup with a single OSD and Mon, and different workload configurations tested, including random read/write with different block sizes. Performance data was collected using eventtracing and analyzed to draw conclusions.
![• Ceph cluster with single OSD, single Mon and a
manger [1]
• Bluestore as backend
• Separate partitions for wal and db
• SSD P3700 NVMe/PCIe 400GB
[1]https://github.com/MahatiC/ceph-single-osd-mon
Test environment setup](https://image.slidesharecdn.com/04-cephclientlibrbdperformanceanalysisandlearnings-mahatichamarthy-180419050457/85/Ceph-Client-librbd-Performance-Analysis-and-Learnings-Mahati-Chamarthy-16-320.jpg)
Ad
More Related Content
What's hot (19)
Similar to Ceph Client librbd Performance Analysis and Learnings - Mahati Chamarthy (20)
Ad
Recently uploaded (20)
Ad
Ceph Client librbd Performance Analysis and Learnings - Mahati Chamarthy
- 1. Ceph Client (librbd) Performance Analysis and Learnings Mahati Chamarthy Software Engineer, Intel
- 2. • RBD • Tracing/profiling tools • Test environment setup • Workload configs • Performance analysis & conclusions Agenda
- 3. RBD • RBD images – block devices striped over objects and stored in RADOS object store • Stripes images across the entire cluster • Common use-case: disks for VMs
- 4. RBD • rbd_data.$rbd_id.$fragment • rbd_directory • rbd_info • rbd_id.$rbd_name • rbd_header.$rbd_id • rbd_object_map.$rbd_id
- 5. • Interacts with OSD via kernel module or librbd • RBD layering • Copy-on-write • Integration with QEMU, libvirt, Linux Kernel • RBD mirroring RBD Features
- 7. • Ceph subsystem logging • Wireshark/t-shark for analyzing network traffic • Zipkin/Blkin • Ceph perf counters along with rbd replay • Perf tool & flame charts • Eventtrace Tracing/Profiling tools
- 13. perf flame graph
- 14. • Lttng based • Ease of use with minimal code changes • Eventtrace support exists in Ceph code • Generates large amount of data – needs to be parsed Eventtrace
- 16. • Ceph cluster with single OSD, single Mon and a manger [1] • Bluestore as backend • Separate partitions for wal and db • SSD P3700 NVMe/PCIe 400GB [1]https://github.com/MahatiC/ceph-single-osd-mon Test environment setup
- 17. • Workloads collected on an rbd image of size 1024MB, in a pool with 100 pgs, and cache disabled • Data collected using eventtrace; the values are averaged out using a python script • Config: o random write o blocksize: 4k o iodepth: 1 o runtime: 6sec Config - I
- 20. • Workloads collected on an rbd image of size 1024MB, in a pool with 100 pgs, and cache disabled • Data collected using eventtrace; the values are averaged out using a python script • Config: o random read o blocksize: 4k o iodepth: 1 o runtime: 6sec Config - II
- 22. • Workloads collected on an rbd image of size 1024MB, in a pool with 100 pgs, and cache disabled • Data collected using eventtrace; the values are averaged out using a python script • Config: o random write o blocksize: 16k o iodepth: 1 o runtime: 6sec Config - III
- 24. • Workloads collected on an rbd image of size 1024MB, in a pool with 100 pgs, and cache disabled • Data collected using eventtrace; the values are averaged out using a python script • Config: o random read o blocksize: 16k o iodepth: 1 o runtime: 6sec Config - IV
- 27. Thank you!