courseware Power-aware scheduling Jan Madsen Informatics and Mathematical

courseware Power-aware scheduling Jan Madsen Informatics and Mathematical Modelling Technical University of Denmark Richard Petersens Plads, Building 321 DK 2800 Lyngby, Denmark Jan@imm. dtu. dk

Mission critical embedded systems § Based on work by § § J. Liu, P. H. Chou, N. Bagherzadeh, F. Kurdahi § University of California, Irvine § CODES’ 01 & DAC’ 01 So. C-MOBINET courseware Jan Madsen

Mars Rover – Mission § Perform experiments § Autonomous mobile vehicle § Alpha proton X-ray spectrometer § Imaging § Travel between different target locations So. C-MOBINET courseware Jan Madsen

Mars Rover – Conditions § Surface temperature [-40 o. C; -80 o. C] § Communication ~ 11 minute § No real-time control § Supervised autonomous control So. C-MOBINET courseware Jan Madsen

Mars Rover - System composition § CPU § 3 images per day § Motors § 60 cm per min § Hazard detection § Heaters § -80 o. C requires motors to be heathed So. C-MOBINET courseware Jan Madsen

Mars Rover – Power? § Power sources § Battery (non-rechargeable) § Solar panel (free) § Power consumers § Digital: imaging, communication, control § Mechanical: driving, steering § Thermal: heating motors in the low-temperature environment So. C-MOBINET courseware Jan Madsen

System-level power manager § Amdalhs’ law applies to power § Power savings of a component is scaled to its contribution to power usage of the whole system § If a component draws 2% of the power in a system, a 50% power reduction amounts to 1% saving to the system § The power manager must consider all power consumers in the entire system and identify the major power consumers So. C-MOBINET courseware Jan Madsen

System-level power manager § System-level power consumers § (Digital) computation domain § Processors, memory, I/O, ASIC § Non-computation domains § Mechanical: motors § Thermal: heaters § Major power consumers: mechanical and thermal So. C-MOBINET courseware Jan Madsen

Power-aware vs. low-power § Low-power § § Minimize power usage Just enough power to meet performance requirement No distinction between costly power and free power Component-level power managers § Power-aware § Best use of available power § Minimize power usage with low power budget § Deliver high performance with high power budget § Distinguish different models of power sources § Battery, solar, nuclear, etc. § Track variant power availability § System-level power managers So. C-MOBINET courseware Jan Madsen

Low-power scheduling § Shutting down subsystems § Variable-voltage processor scheduling § Limited applicability to power-aware designs § § Timing constraints are not strongly guaranteed Power usage is handled as a by-product No tracking to power availability No distinction to different energy sources So. C-MOBINET courseware Jan Madsen

Low-power scheduling - Example p 1 r 1 p 2 p 3 r 1 idle p 1 r 1 idle r 1 p 2 p 3 r 1 idle So. C-MOBINET courseware Jan Madsen

Power-aware scheduling § Min/max timing constraints on tasks § Min timing constraint § Subsumes precedence as special cases § Max timing constraint § Subsumes deadline as special cases § Min/max power constraints on the system § Max power constraint § Total power budget from the available sources § Hard constraint, must be guaranteed § Min power § Free power (solar), minimize power jitter § soft constraint, best effort So. C-MOBINET courseware Jan Madsen

Constraint graph G(V, E) § Vertices V: tasks § d(v), execution delay § p(v), power consumption § r(v), resource mapping So. C-MOBINET courseware Jan Madsen § Edges E: timing constraints § Forward edge: min constraint § Backward edge: max constraint

Constraint graph G(V, E) § Schedule § Timing-valid schedule § Time assignments to tasks § Finish time So. C-MOBINET courseware Jan Madsen § Timing constraints satisfied § No resource conflict

Power-aware Gantt chart § Time view § Power view § Bins – tasks § Horizontal axis – start time, duration § Vertical axis – power § Tracks – parallel resources So. C-MOBINET courseware Jan Madsen § Power profile § Power constraints § Power properties § Spikes, gaps § Energy cost § Utilization

Mars Rover - Exercise So. C-MOBINET courseware Jan Madsen

Mars Rover - Exercise Power sources & tasks Duration (sec. ) Power @ -40 o. C Power @ -60 o. C Power @ -80 o. C Solar panel 17 14 11 Battery pack 8 max Constant 2 3 4 Heating two motors 5 8 10 12 Driving 10 8 11 14 Steering 5 4 6 8 Hazard detection 10 3 4 5 CPU So. C-MOBINET courseware Jan Madsen

Mars Rover - Solution Worst case at – 80 o. C Hd St Dr HW 12 HW 34 HW 56 HS 12 HS 34 CPU 9 9 16 16 16 12 18 18 9 9 12 18 18 Power So. C-MOBINET courseware Jan Madsen

Power properties § Power profile P (t) § System-level power consumption curve § Power constraints § Max power constraint Pmax § Power Spike: max power constraint violation § Min power constraint Pmin § Power Gap: min power constraint violation § Power-validity § A timing-valid schedule with no power spikes § Enforce max power budget So. C-MOBINET courseware Jan Madsen § Min power utilization (Pmin) § Energy utilization from free sources § Energy cost Ec (Pmin) § Energy drawn from expensive (non-free) sources § Power-aware trade-off § Performance vs. Energy cost Ec (Pmin)

Mars Rover – Power profile Pmax 20 P (t) Pmin 10 (Pmin) = Ec (Pmin) = So. C-MOBINET courseware (11 x 75) – (2 x 10) (11 x 75) = 95. 2 % 5 x 25+5 x 1+10 x 7 = 3. 4 75 Jan Madsen

Mars Rover – the real thing! § Timing constraints § Three cases w/ different power constraints § Max power: § solar + 10 W § Min power § § solar, free Best: 14. 9 W Typical: 12 W Worst: 9 W So. C-MOBINET courseware Jan Madsen

Scheduling results § Best case § Fast, low cost § Typical case § Slower, increased cost So. C-MOBINET courseware Jan Madsen § Worst case § Slower, high cost § Same as the existing serial schedule

Comparisons to schedules § Existing low-power schedule § Low performance § Low energy cost § Under-utilized free solar power § Does not track power sources § Full serialization by handcrafting So. C-MOBINET courseware Jan Madsen § Power-aware schedules § High performance § High energy cost § Improved utilization of solar power § Tracks available power from different sources § Fully constraint-driven by an automated design tool

Comparisons in a scenario § Scenario § Mission: travel to a target 48 steps away § Existing low-power schedule § Fixed slow speed § Low energy cost in each phase, but high energy cost in worst case § Low performance, high energy cost So. C-MOBINET courseware Jan Madsen § 3 phases: best, typical, worst, 10 min each § Power-aware schedules § Accelerated speed by tracking available power § Finish earlier before working in the worst case § High performance, low energy cost

Conclusion § Power-aware design § Different from low-power § Deliver high performance by tracking power sources § Power-aware schedulers § Incremental scheduling by constraint classification § Potentials on performance speedup and energy saving § System-level design tools § Power manager for the entire system § Aggressive design space exploration So. C-MOBINET courseware Jan Madsen

Incremental scheduling (1) § (1) Timing scheduling § § Topological traversal of the constraint graph Selective serialize tasks that share the same resource Prohibit positive cycles Proven to find a timing-valid schedule So. C-MOBINET courseware Jan Madsen

Incremental scheduling (2) § (2) Max power scheduling § § § Begin with a timing-valid schedule from (1) Enforce max power constraint Reorder tasks to eliminate power spikes Redo (1) for timing violation Heuristics applied So. C-MOBINET courseware Jan Madsen

Incremental scheduling (3) § (3) Min power scheduling § § § Begin with a power-valid schedule from (2) Reorder tasks to reduce power gaps in best-effort Deliver same performance with less energy cost Heuristics applied Results applicable to different constraints So. C-MOBINET courseware Jan Madsen