Energy-Efficient Mapping and Scheduling for DVS Enabled Distributed Embedded Systems Marcus T. Schmitz and Bashir M. Al- Hashimi University of Southampton, United Kingdom Petru Eles Linköping University, Sweden
Mar 28, 2015
Energy-Efficient Mapping and Scheduling for DVS Enabled
Distributed Embedded Systems
Marcus T. Schmitz and Bashir M. Al-HashimiUniversity of Southampton, United Kingdom
Petru ElesLinköping University, Sweden
2Marcus T. SchmitzUniversity of Southampton
Contents• Motivation & Introduction
• Dynamic Voltage Scaling
• Co-Synthesis with DVS Consideration
• DVS optimised Scheduling
• DVS optimised Mapping
• Experimental Results
• Conclusions
3Marcus T. SchmitzUniversity of Southampton
MotivationLow Energy:
• Portable Applications
• Autonomous Systems
• Feasibilty Issues (SoC - heat)
• Operational Cost and Environmental Reasons
System Level Co-Design:
• Shrinking Time-To-Market Windows
• Reducing Production Cost
• High Degree of Optimisation Freedom
4Marcus T. SchmitzUniversity of Southampton
Introduction
Dynamic Voltage Scaling
System Level Co-Synthesis
Energy-Efficient Co-Synthesis for
DVS Sytems
5Marcus T. SchmitzUniversity of Southampton
Dynamic Voltage Scaling (DVS)
f Reg.
DVS Processor
0
0.2
0.4
0.6
0.8
1
1.2
1 1.5 2 2.5 3 3.5 4 4.5 5
Energy vs. Speed
Voltage/Frequency
Frequency
VR
Available from: Transmeta, AMD, Intel
1/Speed
En
erg
y
2ddVkE
6Marcus T. SchmitzUniversity of Southampton
Co-Synthesis for DVS Systems
Allocation
Mapping
Scheduling
Voltage Scaling
Evaluation
EE
-GL
SA
EE
-GM
A
De
sig
ne
r d
riv
en
System Specification, Technology Lib.
7Marcus T. SchmitzUniversity of Southampton
DVS in Distributed Systems [23]
PE0
PE1
CL0
P
td
PE0
PE1
CL0
P
td
@ Vmax @ dyn. V
Input:Scheduling (mapping)Power profile
Output:scaled voltage for each DVS task
Emax Esc < Emax
Slack
2.3V 2.4V3.3V
Voltage Scaling
8Marcus T. SchmitzUniversity of Southampton
Energy-Efficient Scheduling
Two objectives:
• Timing feasibility
• Garantee deadlines
• Low energy dissipation
• Optimisation DVS usability – Slack time
Problem due to power variations:
• Simply increase deadline slack leads to sub-optimal solutions!
Traditional scheduling technique focus mainly on timing feasibility!
9Marcus T. SchmitzUniversity of Southampton
Energy-Efficient Scheduling
0
4 5
12
36
E=71J
4 5
01 2
36
4 5
01 2
3 6
012
36
4 5
E=71J
E=53.9J
E=65.6J
Slack Savings
Slack Savings
S1:
S2:
DVS
DVS
Slack
Slack
PE0
PE1
PE2
PE0
PE1
PE2
P
t t
tt
P
P
P
10Marcus T. SchmitzUniversity of Southampton
Energy-Efficient Scheduling• Based on Genetic List Scheduling Algorithm [6,10]
• Task priorities are encoded into priorities strings
List Scheduler
4 3 9 7 2
PS
Duties of the Scheduler:1. Select ready task with highest
priority2. Schedule selected task3. Update schedule and ready list4. Repeat until no un-scheduled
task is left
Schedule
11Marcus T. SchmitzUniversity of Southampton
EE-GLSA
List Scheduler DVS
Assign fitness
Rank individuals
Selection
Mutation
Mating
InsertionIniti
al P
opul
atio
n
Opt
imis
ed P
opul
atio
n
GA
low high
Timing, Energy
3
7
8
1
2
3
2
1
3
2
No Hole Filling!No Mapping!
12Marcus T. SchmitzUniversity of Southampton
Advantages
• Optimisation can be based on an arbitrary complex
fitness function, including:
• Timing
• Energy (DVS technique)
• Enlarged search space (|T+C|! different schedules)
• Trade-off freedom: Synthesis time <-> quality
• Easily adaptable to computing clusters
• Multiple populations with immigration scheme
13Marcus T. SchmitzUniversity of Southampton
Hole Filling Problem
d2
d4
d3
7
6
4
4
1
d2 d3,4
Hole filling
Therefore, priorities decide solely upon execution order!
PE0
PE1
14Marcus T. SchmitzUniversity of Southampton
Task Mapping
Why seperation from the list scheduling?• Regardless of priorties, greedy mapping
LS
d2
7
4
5
d1
d1,2
PE0
PE1
P
t
15Marcus T. SchmitzUniversity of Southampton
Task Mapping
Make greedy mapping decision based on:• Timing• Energy
LS
d2
7
4
5
d1
d1,2
?
?PE0
PE1
P
t
16Marcus T. SchmitzUniversity of Southampton
Task Mapping
Make mapping decision based on:• Timing• Energy
LS
d2
7
4
5
d1
d1,2
PE0
PE1
P
t
17Marcus T. SchmitzUniversity of Southampton
Task Mapping
Make mapping decision based on:• Timing• Energy
LS
d2
7
4
5
d1
d1,2
?
?
PE0
PE1
P
t
18Marcus T. SchmitzUniversity of Southampton
Task Mapping
Make mapping decision based on:• Timing• Energy
LS
d2
7
4
5
d1
d1,2
PE0
PE1
P
t
19Marcus T. SchmitzUniversity of Southampton
Task Mapping
Make mapping decision based on:• Timing• Energy
LS
d2
7
4
5
d1
d1,2
PE0
PE1
P
t
20Marcus T. SchmitzUniversity of Southampton
Task Mapping
Make mapping decision based on:• Timing• Energy
LS
d2
7
4
5
d1
d1,2
PE0
PE1
P
t
21Marcus T. SchmitzUniversity of Southampton
Task Mapping
Make mapping decision based on:• Timing• Energy
LS
d2
7
4
5
d1
d1,2
PE0
PE1
P
t
22Marcus T. SchmitzUniversity of Southampton
Genetic Mapping Algorithm [8]
CPU DVS-CPU
ASIC
01
2d
d
5
3
6
4
0
1 2
task PE
0 1
1 0
2 2
3 1
4 1
5 0
6 0
Chromosome
Task mapping are encoded into mapping strings
23Marcus T. SchmitzUniversity of Southampton
EE-GMA
EE-GLSA
Assign fitness
Rank individuals
Selection
Mutation
Mating
Insertion
Initi
al P
opul
atio
n
Opt
imis
ed P
opul
atio
n
GA
low high
Timing, Energy + Area
Including DVS
24Marcus T. SchmitzUniversity of Southampton
Experimental Results• 4 Benchmark Sets:
• 27 generated by TGFF [7]
– 8 to 100 tasks: Power variations 2.6
• 2 Hou examples taken from [13]
– 8 to 20 tasks: Power variations 11
• TG1 and TG2 taken from [11]
– 60 examples with 30 tasks, each: No power variations
• Measurement application taken from [3]
– 12 tasks: No power profile is provided
• Power and time overhead for DVS is neglected
• Average results of 5 optimisation runs
25Marcus T. SchmitzUniversity of Southampton
Schedule Optimisation
0
10
20
30
40
50
60
70
80
Tgff1 Tgff2 Tgff3 Tgff4 Tgff5 Tgff6 Tgff7 Tgff8 Tgff9 Tgff10
Example
Red
uct
ion
(%
)
EVEN-DVS[18]
GLSA+EVEN
EE-GLSA
26Marcus T. SchmitzUniversity of Southampton
Schedule Optimisation
0
5
10
15
20
25
30
35
40
TG1 TG2
Benchmark set
Re
du
cti
on
(%
)
LEneS [11]
EE-GLSA
27Marcus T. SchmitzUniversity of Southampton
Mapping Optimisation
0
10
20
30
40
50
60
70
80
90
Tgff1 Tgff2 Tgff3 Tgff4 Tgff5 Tgff6 Tgff7 Tgff8 Tgff9 Tgff10
Example
Red
uct
ion
(%
)
EVEN-DVS
EE-GMA
28Marcus T. SchmitzUniversity of Southampton
Conclusions
• DVS capability can achieve high energy savings in distributed embedded systems
• Proposed a new energy-efficient two-step mapping and scheduling approach
• Iterative improvement provides high savings / ad hoc constructive techniques are not suitable
• Optimisation times are reasonable
• Additional objectives can be easily included
• Consideration of power profile information leads to further energy reductions