Reliability-Aware Energy Optimisation for
Fault-Tolerant Embedded MP-SoCs
Summary
Fault-tolerance
Faults are tolerated by using
temporal or spatial
redundancy, or a combination
of the two
Fault detection is done using
well known techniques such
as: timing and bit coding
Design optimisation tool for distributed embedded
real-time systems
Decides mapping, fault-tolerance policy and
fault-tolerant schedule
PE1
P1
PE2
P1
Replication
Deadline
Fully Transparent Scheduling
PE1
Hard real-time,
Hard reliability goal,
PE2
Static schedule for processes and messages,
Bus
Comparison of FT schemes
P1
P1
P2
P2
P3
P4
P4
P5
P3
2
1
100% E0
Fault-tolerance for k transient/soft faults
Optimise for minimal energy consumption
P5
Slack Sharing Scheduling
PE1
While considering impact of lowering voltages on
the probability of faults
P1
PE2
P4/5
P5
P3
P3
2
1
Bus
Constraint logic programming (CLP) based
implementation
P4
P2
63% E0
Conditional Scheduling
PE1
P1
P1
Fault-tolerant scheduling
PE2
Re-execution
PE1
P1
PE2
P1
Energy vs. Faults
Recent research shows that
the probability of
transient/soft faults increases
dramatically when decreasing
the voltage of a circuit
Many modern designs uses
dynamic voltage scaling (DVS)
to minimise energy
consumption
Fault-tolerant systems that
use power management
techniques may prove to be
fault-tolerant but unreliable
due to increase in faults
Relation between faults and
voltage is given by1:
1
2
PE1 PE
m1
Performance, and hence the obtainable energy savings are
greatly increased
P2
P3
P4
2
PE1
P2 40 40
P3 30 30
m2
P4 40 40
k=1
P5 40 40
P5
Energy vs. reliability
Deadline
Straightforward (SS)
PE1
The use of DVS increases the probability of faults, thus
damaging the system reliability
P1
PP2 2
PE2
P
P4
PP33
PP55
P6
1
Bus
2
R=0.999 999 987
100% E0
Energy optimisation (EO)
Considering reliability in the optimisation process allows for
finding the minimum energy schedule that meets the
reliability goal
PE1
Reliability is imposed as a constraint
P22
P
P1
PE2
P5
P4
P3
P6
1
Bus
2
R=0.999 999 878
68% E0
Reliable energy optimisation (REO)
Considering the reliability while optimising enables us to find
reliable schedules with comparable energy savings
RG imposed as hard constraint
PE1
d 1− f
1− f
min
P1
P2
P4
PE2
Bus
P3
P5
1
P6
2
R=0.999 999 920
PE1 PE
P1
P2
m1
P5
m2
P1 10 X
PE1
Voltage levels
PE1 100 66 33
PE 2 100 66 33
P4 40 X
P6
P5 X 40
P6 X 50
1
Comparison of energy savings
PE2
P3 X 40
Relation between voltage and failure rate
k=1
73% E0
2
P2 60 X
P3
P4
D. Zhu et al.: “Reliability-Aware Energy
Management for Periodic Real-Time Tasks”,
2007
PE2
P1 40 40
Reliability can be met at very little energy cost
λ f = λ 0 10
Trade-off transparency for performance
Reliability must be considered in the optimisation
process
P3
P1
System reliability is affected by use of energy
management
P5
38% E0
More complex schemes demand larger schedule
tables to be stored in the processing elements,
and more sophisticated online schedulers
P4
P2
Bus
Reliable energy management
1
P1
PE2
More complex scheduling schemes yield more
slack for energy management
Passive Replication
PE1
k=1
Comparison of system reliability
IMM / Informatics and Mathematical Modelling
eselab.imm.dtu.dk
k=1
RG=0.999 999 900