Everything described here harks back to FC7 days and is retained here for historical purposes only.
Hypervisor Development
This page is a repository of information relating to Hypervisor Development in Fedora.
Links & Resources
- Xen Project
- fedora-xen mailing list
- Herbert Xu's hypercall docs (coming soon)
- Steven Rostedt's hv docs (coming soon)
- Red Hat Emerging Technologies includes cool bleeding edge virtualization stuff
- paravirt queue mercurial repo containing paravirt ops patches and Rusty's lhype project
- Credit Scheduler Documentation
- CKRM v/s Xen CPU Scheduler
- Xen benchmarks from Bull (not comprehensive, but interesting)
- Virtualization hardware extensions useful basic info on Pacifica & VT
Papers
- Scheduling Algorithms and Operating Systems Support for Real-Time Systems, 1994, Krithi Ramamritham, John A. Stankovic. (covers EDF)
- K. J. Duda, D. R. Cheriton. Borrowed-Virtual-Time (BVT) Scheduling: Supporting Latency-sensitive Threads in a General-purpose Scheduler.
TODO List
Task Ownership Legend
- jm - JamesMorris
- hx - HerbertXu
- sct - StephenTweedie
- sr - StevenRostedt
Add yourself to the above if you wish to participate. Once you select a task to work on, add your initials (in bold) next to it below. To add a new item, discuss it first on the fedora-xen mailing list then update this page.
Current Items (FC7)
- Security audit [hx]
- Network performance [hx]
- Analyze & document Credit Scheduler, identify issues and potential improvements [sr]
- lhype [jm]
Future Items
- Enforce binary interface versioning for HV & VMs
- SMP scalablility ^1^
- Investigate re-writing timer code using latest Linux kernel ideas (sr ?)
- XSM (jm ?)
- Investigate NAPI-like mitigation within the HV
- May need to add HV heap ballooning
- NUMA functionality (upstream ?)
- ia64
- ppc64
- big pages (upstream work)
- MSI (upstream work)
- PM (upstream work)
- Virtualized block I/O ^2^
- Migration from 32 to 64-bit host (investigate some time, probably very difficult)
- PAE > 16GB ?
- Xen Share ?
^1^ Andrew Theurer from IBM presented a paper at OLS on a number of SMP scalability issues. The big ones were writeable pagetables (with his tests it appears to be slower than emulation even on UP); page table lock contention and global counter when writing to CR3.
^2^ sct: Initially, just the normal virtual disk front/back-end pair. For FC5, and certainly for RHEL-5, ideally we'd like to have something that looks like a real SCSI block device. ie. that supports proper SCSI ID inquiries from domU etc.