1. Supporting Time-Sensitive Applications on a Commodity OS
Ashvin Goel, Luca Abeni, Charles Krasic, Jim Snow, Jonathan Walpole
Presenter: Madhura S Rama

2. Agenda Goal Why TSL? Time Sensitive Requirements TSL Implementation
Firm Timers
Fine-Grain Kernel Preemptibility
CPU Scheduling
Experimental Results
Conclusion

3. Goal
Integrate efficient support for time-sensitive applications in a commodity OS without significantly degrading the performance of traditional applications.

4. Why TSL? Soft real-time applications are time sensitive
Resources must be allocated at appropriate times
Commodity OS provides coarse-grained resource allocation to maximize throughput
Conflicts the needs of RTOS
Time-sensitive Linux (TSL) - provides good performance to both time-sensitive and throughput–oriented applications.

5. Time Sensitive Requirements
Allocate resources at “appropriate times”
Kernel latency – latency between the event and its activation
Should be low in a time-sensitive application
Three requirements to reduce this latency
Preemption
Latency
Timer Latency
Scheduling Latency
Time
Interrupt
Handler
Another app
Scheduled
Wall-clock time
event
Timer
Interrupt
Scheduler
Application Scheduled
(Activation)

6. Different Timers Periodic Timers One-Shot (Hard) Timers Soft Timers
Implemented with Periodic timer interrupts. Max 10ms latency
Decrease in latency increases interrupt overhead
One-Shot (Hard) Timers
Interrupts only when needed
Timer reprogramming, fielding interrupts
Soft Timers
Avoids interrupts
Cost of polling, timer latency
Firm Timers
Combines all the advantages of the above timers
Incurs very low overhead

7. Three key techniques Accurate Timing Mechanism Responsive Kernel
Firm Timers
Responsive Kernel
Lock-breaking preemptible kernel
Appropriate CPU Scheduling Algorithm
Proportion-Period Scheduler
Priority-based Scheduler

8. Firm Timers Provides accurate timing mechanism with low overhead
Combines one-shot timers with soft timers by exposing a timer overshoot parameter
One-shot timer is programmed to fire after an overshoot amount of time after the next timer expiry
If a system call occurs after a timer has expired but before the one-shot timer expired, soft timers become effective.
Reprogram the one-shot timer for the next timer expiry – does not incur any overhead

9. Overshoot Accuracy vs Overhead tradeoff
Reprogram One-shot timer
and fire the event
Poll for timer
expiry
User Space
System call
Kernel Space
Overshoot Parameter
Time
Time-Sensitive
Application ready
For execution
One-shot timer
expiry
Timer Latency

10. Firm Timer Implementation
Maintains a timer queue for each processor
One-shot APIC timer is programmed to generate an interrupt at the next timer expiry event + global overshoot value (can be made dynamic)
Soft timers enabled using non-zero timer overshoot value
Periodic timer used when a timer needs longer timeout
TSL use the standard POSIX interface calls - modified the implementation to use firm timers

11. Fine-Grained Kernel Premptibility
Kernel latency increases with increase in the length of the non-preemptible critical section
Approaches to reduce kernel latency
Explicit insertion of preemption points
Allow preemption anytime the kernel is not accessing shared data structures
Robert Love’s lock-breaking preemptible kernel

12. CPU Scheduling Uses a combination of Proportion-Period CPU Scheduling
Priority CPU Scheduling

13. Proportion-Period CPU Scheduling
Provides “temporal protection” by allocating each task a fixed proportion of the CPU at each task period
Proportion can be assigned either statically or dynamically using a feedback mechanism
Improves accuracy of scheduling analysis
2/3 T1
Proportion
Time
Period
1/3 T2
Time
Proportion

14. Priority CPU Scheduling
Real-time priorities assigned based on application needs
TSL schedules fixed priority tasks in the background
Exception: Shared server tasks -> Priority Inversion
Use Highest locking priority protocol (HLP) to cope with priority inversion
Minimizes scheduling Latency

15. Latency in Real Applications

16. System Overhead

17. Conclusion
TSL can support applications requiring fine-grained resource allocation and low latency
TSL is truly a general-purpose system – can be used effectively for both time-sensitive and throughput-oriented applications

18. Backup Experimental Results
Firm timers are effective in reducing the overhead of one-shot timers when the ratio of number of the soft timers that fire to the number of soft timer checks is sufficiently large (2.1%)
Overhead of TSL on throughput-oriented applications is low
Provides allocation to time-sensitive applications with a variation of less than 400 usec even under heavy load.