Deadline scheduling: can your mobile device last longer?

Transcription

1 Deadline scheduling: can your mobile device last longer? Juri Lelli, Mario Bambagini, Giuseppe Lipari Linux Plumbers Conference 202 San Diego (CA), USA, August 3 TeCIP Insitute, Scuola Superiore Sant'Anna Area della Ricerca CNR, Via G. Moruzzi 5627 Pisa, Italy

2 Outline Classical Real-Time concepts An implementation Power Aware Deadline Scheduling Conclusions & Discussion 8/30/2 Juri Lelli - PA deadline Plumbers2 2/40

3 Classical Real-Time concepts 8/30/2 Juri Lelli - PA deadline Plumbers2 3/40

4 What's Real-Time? All about deadlines Useful result correct output + before its deadline Go as fast as you can is not always the best answer Go as fast as you need to respect deadlines But not faster! Or you will spend too much energy! 8/30/2 Juri Lelli - PA deadline Plumbers2 4/40

5 Hard vs. Soft RT Hard real-time tasks: a missed deadline may cause catastrophic consequences. Soft real-time tasks: timing constraints are important, but not critical (QoS degradation) 8/30/2 Juri Lelli - PA deadline Plumbers2 5/40

6 Resource Reservation Concurrent real-time tasks compete for resources (CPU time) Resource Reservation mechanism A task is allowed to execute for: Q time units (budget) in every interval of length P (reservation period) Allocated bandwidth is U=Q/P (a.k.a. utilization) 8/30/2 Juri Lelli - PA deadline Plumbers2 6/40

7 Temporal Isolation nothing new? CFS bandwidth control RT throttling Reservations, like containers Each reservation has its own dynamic deadline (Constant Bandwidth Server) Reservations are scheduled by Earliest Deadline First 8/30/2 Juri Lelli - PA deadline Plumbers2 7/40

8 Temporal Isolation EDF gives higher priority to more urgent tasks CBS slows down (or throttles) misbehaving tasks EDF + CBS ensures temporal isolation Admission control provides timing guarantees U i M 8/30/2 Juri Lelli - PA deadline Plumbers2 8/40

9 Choosing parameters Period is easy: P = T or P = NT Budget assignment is painful :'-( Hard real-time WCET calculated with tools from source code estimated via benchmarks Soft real-time measure computation time and adapt budget adaptive reservations 8/30/2 Juri Lelli - PA deadline Plumbers2 9/40

10 Adaptive/Feedback Scheduling 8/30/2 Juri Lelli - PA deadline Plumbers2 0/40

11 Mobile Data Not only Multimedia not strictly periodic no explicit deadline how to cope with other data sources (e.g., Web browsing, s, user interaction, etc.)? pack similar sources on container reservations best effort (we don't care) 8/30/2 Juri Lelli - PA deadline Plumbers2 /40

12 An implementation 8/30/2 Juri Lelli - PA deadline Plumbers2 2/40

13 What we already have SCHED_DEADLINE deadline based real-time task scheduling policy bandwidth isolation each task runs inside its own reservation global/clustered multiprocessor scheduling through dynamic task migration 8/30/2 Juri Lelli - PA deadline Plumbers2 3/40

14 What we already have SCHED_DEADLINE SCHED_RR SCHED_FIFO SCHED_BATCH SCHED_NORMAL SCHED_IDLE dl.c rt.c fair.c kernel/sched/ 8/30/2 Juri Lelli - PA deadline Plumbers2 4/40

15 What we already have "Adaptive Quality of Service Architecture" (for the Linux kernel) adaptive resource reservation layer dynamically adapting CPU reservation QoS aware applications API a bit outdated & UP only :-( 8/30/2 Juri Lelli - PA deadline Plumbers2 5/40

16 Power Aware Deadline Scheduling 8/30/2 Juri Lelli - PA deadline Plumbers2 6/40

17 Power aware RT scheduling Let's bring power efficiency into the picture... Hardware context (simplified) homogeneous / heterogenous single / multiple clock domains discrete operating frequencies transitories to adjust to a new frequency overhead for frequency scaling overhead for switching a core either ON or OFF 8/30/2 Juri Lelli - PA deadline Plumbers2 7/40

18 Ideal UP example Associate a reservation to each task Track actual CPU utilization (active servers contribute) U Dynamically change CPU speed S to follow U Admission Control don't go below U if you don't want deadline misses Power efficiency slow down as much closer to U as you can to save power 8/30/2 Juri Lelli - PA deadline Plumbers2 8/40

19 Ideal UP example τ τ 2 S 8/30/2 Juri Lelli - PA deadline Plumbers2 9/40

20 Ideal UP example τ reservations τ 2 S 8/30/2 Juri Lelli - PA deadline Plumbers2 20/40

21 Ideal UP example τ τ 2 S background best effort 8/30/2 Juri Lelli - PA deadline Plumbers2 2/40

22 UP with thresholds No existing processor can vary its frequency with continuity set some thresholds on the values of the total system bandwidth CPU with N different frequencies compute N different bandwidth thresholds use them to know when a CPU speed change is needed, for example... 8/30/2 Juri Lelli - PA deadline Plumbers2 22/40

23 UP with thresholds U = U = 0.7 U = 0.5 8/30/2 Juri Lelli - PA deadline Plumbers2 23/40

24 Going SMP SCHED_DEADLINE performs active load-balancing Global-EDF on a M-CPUs system the M earliest deadline ready tasks are always running achieved through push/pull migrations (as in rt) deadline based decisions 8/30/2 Juri Lelli - PA deadline Plumbers2 24/40

25 Power aware push/pull hybrid scheduling EDF on each CPU U driven migrations different policies single clock domain minimize mean U multiple clock domains compact tasks to turn off CPUs as long as possible big.little compact on little core, move only big tasks on big cores? 8/30/2 Juri Lelli - PA deadline Plumbers2 25/40

26 PA push/pull an example single clock domain spread load to go slower /30/2 Juri Lelli - PA deadline Plumbers2 26/40

27 PA push/pull an example /30/2 Juri Lelli - PA deadline Plumbers2 27/40

28 PA push/pull an example /30/2 Juri Lelli - PA deadline Plumbers2 28/40

29 PA push/pull an example new task enters the system /30/2 Juri Lelli - PA deadline Plumbers2 29/40

30 PA push/pull an example /30/2 Juri Lelli - PA deadline Plumbers2 30/40

31 PA push/pull an example lower mean frequency /30/2 Juri Lelli - PA deadline Plumbers2 3/40

32 PA push/pull Tegra like /30/2 Juri Lelli - PA deadline Plumbers2 32/40

33 PA push/pull Tegra like user starts a console class video game 8/30/2 Juri Lelli - PA deadline Plumbers2 33/40

34 PA push/pull Tegra like /30/2 Juri Lelli - PA deadline Plumbers2 34/40

35 PA push/pull Tegra like /30/2 Juri Lelli - PA deadline Plumbers2 35/40

36 PA push/pull Tegra like /30/2 Juri Lelli - PA deadline Plumbers2 36/40

37 Conclusions & Discussion 8/30/2 Juri Lelli - PA deadline Plumbers2 37/40

38 Conclusions prototype implementation should be straightforward account for runqueue actual utilization modify push/pull mechanism integration in a real system (Android?) could be more difficult the plan start simple and then add bits one ofter the other :-) 8/30/2 Juri Lelli - PA deadline Plumbers2 38/40

39 Open Questions When is it worth to wake up a CPU? Interaction with other components (e.g., thermal management, drivers, etc.)? Test platforms? add here your own concern or... 8/30/2 Juri Lelli - PA deadline Plumbers2 39/40

40 Thanks! Questions? The work leading to this talk has been supported by the European Community's Seventh Framework Programme n S(o)OS - Service-oriented Operating Systems ( 8/30/2 Juri Lelli - PA deadline Plumbers2 40/40