SLIDE 1
The imagination driving Australia’s ICT fu
The imagination driving Australia’s ICT fu
Q UESTIONS ? David.Snowdon@nicta.com.au http://ertos.nicta.com.au - - PowerPoint PPT Presentation
Q UESTIONS ? David.Snowdon@nicta.com.au http://ertos.nicta.com.au - - PowerPoint PPT Presentation
R UN T IME P REDICTION O F P ERFORMANCE AND E NERGY WHEN F REQUENCY S CALING David Snowdon, Stefan Petters and Gernot Heiser David.Snowdon@nicta.com.au The imagination driving Australias ICT fu R UN T IME P REDICTION O F P ERFORMANCE AND E
SLIDE 2
SLIDE 3
MOTIVATION
The imagination driving Australia’s ICT fu
➜ Embedded systems are often restricted by battery life. ➜ Total system energy consumption.
Our work looks at effective DVFS in real systems
! " # $
2 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 4
MOTIVATION
The imagination driving Australia’s ICT fu
Theory: E ∝ V 2
50 100 150 200 250 300 350 400 450 500 Normalised Total Energy CPU Frequency (MHz)
3 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 5
MOTIVATION
The imagination driving Australia’s ICT fu
Practice: (PXA255 based system)
50 100 150 200 250 300 350 400 450 500 Normalised Total Energy CPU Frequency (MHz)
4 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 6
MOTIVATION
The imagination driving Australia’s ICT fu
Why?:
5 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 7
MOTIVATION
The imagination driving Australia’s ICT fu
Why?: Simple models
➜ P ∝ fV 2 ➜ T ∝ 1
f
➜ V = F(f) and F monotonically increasing
Modern systems aren’t simple!
➜ Varying number of switches (workload specific!) ➜ Multiple frequency domains ➜ Frequency independent (static) power
5-A RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 8
MOTIVATION
The imagination driving Australia’s ICT fu
Why?: Simple models
➜ P ∝ fV 2 ➜ T ∝ 1
f
➜ V = F(f) and F monotonically increasing
Modern systems aren’t simple!
➜ Varying number of switches (workload specific!) ➜ Multiple frequency domains ➜ Frequency independent (static) power
We want to be able to deal with these nuances
5-B RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 9
EXECUTION TIME MODEL
The imagination driving Australia’s ICT fu
➜ Simple execution time model: T ∝
1 fcpu
➜ i.e. Constant cycles ➜ Problem: Ignores execution time independent of CPU-clock
6 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 10
EXECUTION TIME MODEL
The imagination driving Australia’s ICT fu
➜ Simple execution time model: T ∝
1 fcpu
➜ i.e. Constant cycles ➜ Problem: Ignores execution time independent of CPU-clock
0.8 1 1.2 1.4 1.6 1.8 2 50 100 150 200 250 300 350 400 450 500 Normalised cycles CPU Frequency (MHz) bitcnt 0.8 1 1.2 1.4 1.6 1.8 2 50 100 150 200 250 300 350 400 450 500 Normalised Cycles CPU Frequency (MHz) gzip
Implicaton:
➜ Memory-bound performance is less dependent on CPU frequency
6-A RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 11
EXECUTION TIME MODEL
The imagination driving Australia’s ICT fu
➜ Task: predict the execution time of a workload in an arbitrary system configuration ➜ Low overhead, cross-architectural, dynamic applications
7 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 12
EXECUTION TIME MODEL
The imagination driving Australia’s ICT fu
➜ Task: predict the execution time of a workload in an arbitrary system configuration ➜ Low overhead, cross-architectural, dynamic applications
T = Ccpu fcpu + Cbus fbus + Cmem fmem + Cio fio + . . .
7-A RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 13
EXECUTION TIME MODEL
The imagination driving Australia’s ICT fu
➜ Task: predict the execution time of a workload in an arbitrary system configuration ➜ Low overhead, cross-architectural, dynamic applications
T = Ccpu fcpu + Cbus fbus + Cmem fmem + Cio fio + . . . Cx: characterise a-priori, or online using performance counters Cbus = α1PMC1 + α2PMC2 + . . . Cmem = β1PMC1 + β2PMC2 + . . . (Ccpu inferred from the other results)
7-B RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 14
EXECUTION TIME MODEL
The imagination driving Australia’s ICT fu
➜ Task: predict the execution time of a workload in an arbitrary system configuration ➜ Low overhead, cross-architectural, dynamic applications
T = Ccpu fcpu + Cbus fbus + Cmem fmem + Cio fio + . . . Cx: characterise a-priori, or online using performance counters Cbus = α1PMC1 + α2PMC2 + . . . Cmem = β1PMC1 + β2PMC2 + . . . (Ccpu inferred from the other results)
➜ 2-parameter: avg 1.7%, max 7% ➜ CPU frequency only: avg 10%, max 36%
7-C RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 15
POWER MODEL
The imagination driving Australia’s ICT fu
➜ Simple CMOS model: P ∝ fV 2
Problems:
➜ System power ➜ Static power and leakage ➜ Multiple frequency/voltage domains ➜ Temperature dependence ➜ Conversion inefficiencies
8 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 16
POWER MODEL
The imagination driving Australia’s ICT fu
➜ Simple CMOS model: P ∝ fV 2
Problems:
➜ System power ➜ Static power and leakage ➜ Multiple frequency/voltage domains ➜ Temperature dependence ➜ Conversion inefficiencies
A (slightly more) realistic model: P =
N
- n=0
CnfnV 2
n + Pstatic 8-A RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 17
POWER MODEL
The imagination driving Australia’s ICT fu
The interaction of run-time and static power:
➜ Dynamic energy increases as frequency increases ➜ Static energy decreases as frequency increases
Etotal = Pdyn∆t + Pstatic∆t
Energy CPU Frequency Etotal(f) Edyn(f) Estatic(f)
9 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 18
POWER MODEL
The imagination driving Australia’s ICT fu
Power/Energy model principles:
➜ Events each use an amount of energy ➜ An event may use energy in more than one voltage domain
For our system: Eevents = V 2
cpu(α0PMC 0+· · ·+αmPMC m)+β0PMC 0+· · ·+βmPMC m 10 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 19
POWER MODEL
The imagination driving Australia’s ICT fu
Power/Energy model principles:
➜ Events each use an amount of energy ➜ An event may use energy in more than one voltage domain ➜ Clocks cycles count as events
For our system: Eevents = V 2
cpu(α0PMC 0+· · ·+αmPMC m)+β0PMC 0+· · ·+βmPMC m
Efreqs = V 2
cpu(γ1fcpu+γ2fbus+γ3fmem)∆t+(γ4fcpu+γ5fbus+γ6fmem)∆t 10-A RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 20
POWER MODEL
The imagination driving Australia’s ICT fu
Power/Energy model principles:
➜ Events each use an amount of energy ➜ An event may use energy in more than one voltage domain ➜ Clocks cycles count as events ➜ Static power models power not related to events or voltages. ➜ Constant IO power for the benchmarks tested.
For our system: Eevents = V 2
cpu(α0PMC 0+· · ·+αmPMC m)+β0PMC 0+· · ·+βmPMC m
Efreqs = V 2
cpu(γ1fcpu+γ2fbus+γ3fmem)∆t+(γ4fcpu+γ5fbus+γ6fmem)∆t
Estatic = Pstatic∆t
10-B RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 21
POWER MODEL
The imagination driving Australia’s ICT fu
Parameter selection:
➜ Systematically picking the best model for N counters ➜ Least-squares regression finds the coefficients
11 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 22
EVALUATION
The imagination driving Australia’s ICT fu
➜ Typical embedded platform (PLEB 2, XScale based) ➜ Cycle counter, 2 performance counters, 13 events
12 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 23
EVALUATION
The imagination driving Australia’s ICT fu
➜ Typical embedded platform (PLEB 2, XScale based) ➜ Cycle counter, 2 performance counters, 13 events ➜ 37 benchmarks run to completion at each setpoint for all frequency settings ➜ 22 frequency setpoints with different fcpu, fbus and fmem ➜ Voltage varied to three settings for each frequency
12-A RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 24
EVALUATION
The imagination driving Australia’s ICT fu
➜ Typical embedded platform (PLEB 2, XScale based) ➜ Cycle counter, 2 performance counters, 13 events ➜ 37 benchmarks run to completion at each setpoint for all frequency settings ➜ 22 frequency setpoints with different fcpu, fbus and fmem ➜ Voltage varied to three settings for each frequency ➜ Measurements: Cycles, Frequencies, Performance counters, Energy ➜ Benchmarks were partitioned for calibration and validation
12-B RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 25
EVALUATION
The imagination driving Australia’s ICT fu
R2 (Intercept) v*v*fcpu v*fcpu fcpu v*v*fbus fbus v*v*fmem fmem v*v*PMC0/t v*v*PMC1/t v*v*PMC2/t v*v*PMC3/t v*v*PMC4/t v*v*PMC5/t v*v*PMC6/t v*v*PMC7/t v*v*PMC8/t v*v*PMC9/t v*v*PMC10/t v*v*PMC11/t v*v*PMC12/t v*v*PMC13/t PMC0/t PMC1/t PMC2/t PMC3/t PMC4/t PMC5/t PMC6/t PMC7/t PMC8/t PMC9/t PMC10/t PMC11/t PMC12/t PMC13/t 0.92 0.96 0.98 0.98 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 1 1 1 1 1 1 1 1
13 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 26
EVALUATION
The imagination driving Australia’s ICT fu
Power error data (using a measured run-time): Counters Param. R2 Max Err (%) Avg Err (%) 1 4 0.983 7.5 2.1 2 5 0.987 6.9 2.3 3 6 0.990 4.8 1.3 4 7 0.992 3.8 1.2 5 8 0.993 3.7 0.9 6 9 0.994 2.9 0.9 6 11 0.995 2.7 0.8
14 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 27
EVALUATION
The imagination driving Australia’s ICT fu
Combining with our performance model:
➜ Max: 4.9%, Avg: 1.5% — (Four counters)
15 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 28
CONCLUSIONS
The imagination driving Australia’s ICT fu
Conclusions:
➜ A time estimate is essential for an energy estimate ➜ Our method: Low overhead, accurate, generally applicable ➜ Parameter selection is a useful tool for analysing power ➜ Energy can be saved by changing the memory/bus frequency
16 RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 29
CONCLUSIONS
The imagination driving Australia’s ICT fu
Conclusions:
➜ A time estimate is essential for an energy estimate ➜ Our method: Low overhead, accurate, generally applicable ➜ Parameter selection is a useful tool for analysing power ➜ Energy can be saved by changing the memory/bus frequency
Future work:
➜ Temperature, IO, interrupts and DMA are unaccounted for. ➜ Investigate the effectiveness for other platforms. ➜ Develop frequency switching policies: ➜ Model frequency switching overheads, sleep state energy ➜ Operating system and application knowledge
16-A RUN TIME PREDICTION OF PERFORMANCE AND. . .
SLIDE 30
QUESTIONS?
The imagination driving Australia’s ICT fu
QUESTIONS?
David.Snowdon@nicta.com.au http://ertos.nicta.com.au
17 QUESTIONS?
SLIDE 31
AVAILABLE PMC EVENTS
The imagination driving Australia’s ICT fu PMC Description 0x0 ICache miss 0x1 ICache stall cycles 0x2 Data dependency stalls 0x3 ITLB miss 0x4 DTLB miss 0x5 Branch instruction executed 0x6 Branch mispredicted 0x7 Instruction executed 0x8 DCache buffer stall cycles 0x9 DCache buffer stall 0xa DCache access 0xb DCache miss 0xc DCache write-back 0xd Software changed the PC 18 QUESTIONS?
SLIDE 32
AVAILABLE FREQUENCY SETPOINTS
The imagination driving Australia’s ICT fu
fcpu (MHz) fbus (MHz) fmem (MHz) 1 99.531 49.766 99.531 2 117.964 58.981 117.964 3 132.71 66.354 132.71 4 149.299 49.766 99.531 5 176.946 58.981 117.964 6 199.064 49.766 99.531 7 199.064 66.354 132.71 8 199.064 99.531 99.531 9 235.929 58.981 117.964 10 235.929 117.964 117.964 11 265.420 66.354 132.710 12 265.420 132.710 132.710 13 298.598 49.766 99.531 14 298.598 99.531 99.531 15 353.894 58.981 117.964 16 353.894 117.964 117.964 17 398.131 66.354 132.71 18 398.131 99.531 99.531 19 398.131 132.71 132.71 20 398.131 199.064 99.531 21 471.858 117.964 117.964 22 471.858 235.929 117.964
19 QUESTIONS?
SLIDE 33
ERROR DISTRIBUTION
The imagination driving Australia’s ICT fu
- 5
- 4
- 3
- 2
- 1
1 2 3 4 5 50 100 150 200 250 300 Energy estimation error (%) Actual Energy (J)
20 QUESTIONS?
SLIDE 34
FIXED MEMORY FREQUENCY
The imagination driving Australia’s ICT fu
1 1.5 2 2.5 3 100 150 200 250 300 350 400 Normalised Time CPU Core Frequency (MHz) Bus: 66.4MHz, Memory: 132.7MHz
21 QUESTIONS?