(→QoS (Quality of Service))
|Line 35:||Line 35:|
Revision as of 11:58, 27 March 2008
Grand Unified Scheduler Project + QoS
Glossary of terms
EDF = earliest deadline first
HRT = hard real time
SRT = soft real time
BE = best effort
Can be used (1) as a hard real time policy, but also be (2) used as a way of getting minimal lag to service a burt of events like interrupts.
Cyclic/periodic in nature.
- How does a static priority system like Linux express itself in terms of an EDF scheduler ? This potentially complicated because of non-standard real time uses of real time policies. How do things like hardness in a SCHED_FIFO task map to in EDF ? How does SCHED_RR map onto EDF ? For a special performance driven BE burst case of use of SCHED_FIFO, a lagging or differing is not considered a failure, how is that mapped onto EDF ? This isn't considered a failure for this use the policy.
- Is M-CASH, EDF-HSB etc... flexible enough as an abstract container or common factor mathematically or algorithmically so that all three stock scheduler policies, SCHED_FIFO, SCHED_RR and SCHED_OTHER (CFS) can be constructed in terms of that algorithm ?
- How would a practical EDF system look like with overload code ? Have EDF be largely run queue localized with manual assignments and with crude not-so-rigorous aperiodic overload handling (below) ?
- If these algorithms overlook the problem of aperiodic overloads, then what kind of crude not so rigorous overload handling can we do ? apply what we have already with the current rt-overload logic where we scan run queues across the system (or possibly with a specific CPU set) to migrate a task to another CPU that isn't running a real time task (SCHED_FIFO/RR) ? What about the use of a cheaply precomputed slack span that can be quickly read during a cross processor run queue scan for finding a suitable BE slot span to handle an overload migration with near future soft deadline ? Let's call this EDF overload (rebalancing).
- How cheap is it to compute or pre-compute a span of slack slots for BE threads like irq-threads ?
- Localize slack computation so that we can bypass the use of a global EDF computation if possible for certain aperiodic tasks.
QoS (Quality of Service)
- Do we need EDF for QoS ? Or should we just used special manually tuned IO handling daemon servers to issue requests in terms of bandwidth ? It would construct its scheduling policy statically and/or dynamically in a custom manner with SCHED_FIFO.
- How much CPU time is need to handle a single interrupt ? can this be used as a center piece for an EDF algorithm to use to help pre-compute a BE span ?
mplayer + SCHED_RR + /dev/rtc emulation using timerfd
- [done] lock analysis of the timerfd paths, Posix locking code, etc... [It doesn't really hit locks for file descriptors that are backed with an anonymous inode]
-  think about getting SCHED_RR to be driven by a VBL interrupt and phase lock loop it so that the period is precise and correct.
-  write a ld.so preload library to emulate /dev/rtc so that there can be multiple openings of /dev/rtc from the perspective of the system. /dev/rtc is normally a single open device.
-  think about migration of SCHED_RR tasks across CPUs and synchronization with a common the VBL across those processors.