Here's a tentative development plan on resource management from our side.
Would request others to comment/add their plans as part of this discussion.
This plan, alongwith appropriate usage scenarios, is being requested for
kernel-summit discussion. Most likely, the resource management plans
will be presented as part of container discussion.
1. Infrastructure
Userland interface for task-grouping, specifying group usage limits etc
Paul Menage's container patches will be the basis for this
infrastructure.
2. cpu controller
- Atleast 1-level deep hierarchical proportional fair-share scheduling
- Soft and Hard limit cpu usage of group
- Fine grained cpu control (ex: parts per 1000?)
- SCHED_FIFO policy for groups
Groups having this policy will recv all cpu cycles as long
as they have any task runnable. This is at the cost of
starving other non-SCHED_FIFO groups.
- Sleeper fairness - Will help bursty workloads
- Accounting - group cpu usage
3. memory controller
- Basic rss controller
- user space oom handler
- Improve shared page accounting
- Handling of soft limits
- Provide support for accounting kernel/user memory, slabs, page tables,
dcache entries, vma's, etc
- Improved statistics (folding)
- Add support for better tuning of containers
i. Low and high water marks per container
ii. Per container swappiness support
- Per container swap file(s)
4. Disk io controller
[I don't have sufficient details at the moment on disk io controller
requirements. Will add after I talk to relevant groups here]
--
Regards,
vatsa
-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2005.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
_______________________________________________
ckrm-tech mailing list
https://lists.sourceforge.net/lists/listinfo/ckrm-tech
