Scylla Summit 2022: IO Scheduling & NVMe Disk Modelling
0 likes753 views
This document discusses IO scheduling and modeling NVMe disks. It explains that different components compete for limited disk resources with different priorities and may overconsume resources if not scheduled properly. It then describes using an IO scheduler to get maximum concurrency from the disk and apply request priorities while avoiding overconsumption. The document outlines a token bucket algorithm for rate limiting and previews ongoing work to implement a new scheduler in Seastar and Scylla and add related metrics and tuning capabilities.
Related topics:1 of 14
Download to read offline
More Related Content
What's hot (20)
Ad
Similar to Scylla Summit 2022: IO Scheduling & NVMe Disk Modelling (20)
Ad
More from ScyllaDB (20)
Recently uploaded (20)
Scylla Summit 2022: IO Scheduling & NVMe Disk Modelling
- 1. IO Scheduling & NVMe Disk Modelling Pavel Emelyanov Principal Engineer
- 2. Pavel Emelyanov ■ Ex Linux kernel hacker ■ Founder of the CRIU project ■ Co-maintainer of Seastar & ScyllaDB Principal Engineer
- 3. Why scheduling at all ■ Different components compete for limited resources ■ They have different priorities ■ They have no idea how not to over-consume the resource
- 4. How disk should work Little’s law Internal parallelism
- 5. How to schedule IO ■ Get the maximum concurrency value • With the help of iotune tool ■ Put that much requests into the disk ■ … apply priorities somehow ■ https://www.scylladb.com/2021/04/06/scyllas-new-io-scheduler/
- 6. How to schedule mixed IO
- 7. How disk does works
- 8. Disk math
- 9. Token bucket algorithm (and its math)
- 10. Scheduler math
- 12. Current status and future work ■ The algorithm is implemented in Seastar ■ Going to be merged into next Scylla OpenSource ■ Works well with “legacy” io_properties.yaml • Updated iotune yet to come ■ New per-class metrics reporting the accumulated requests costs ■ On-the-fly tuning of rate limiting factor
- 14. Thank you! Stay in touch Pavel Emelyanov @xemulp (not very active) [email protected]