1. Meta-Scheduling Techniques for Energy-Efficient, Robust and Adaptive Time-Triggered Systems
Babak Sorkhpour, Prof. Roman Obermaisser, Ayman Murshed
Department of Electrical Engineering and Computer Science
University of Siegen Germany
Kbei, Tehran, Dec 2017

2. Agenda Introduction Basic Concepts And Related Work
MILP & MIQP
Meta-Scheduling Techniques for Energy-Efficient, Robust and Adaptive Time- Triggered Systems
The inputs models for meta-scheduling
Visualizing Scenario-based Meta-Schedules for Adaptive Time-Triggered Systems
Example Scenario And Results
Outputs and results
Conclusion

3. Introduction
This work is motivated to use of dynamic voltage frequency scaling (DVFS) for both communication and computation sides to solve the energy-efficient problem in the whole of NoC structure in MPSoCs.
In this work, an algorithm is developed, which supports the mapping and scheduling of jobs to NoC architectures for minimizing the total energy consumption considering the time constraints and adjustable frequencies .
This algorithm is intended for mixed-criticality, and safety-critical adaptive TT systems and can cover fault-tolerance requirements.

4. Basic Concepts And Related Work
Time-Triggered Embedded Systems
Worst Case Execution Time
Power-Energy Saving Technique
on-chip routers and links consume up to 18% in the Intel SCC, 20% in the Alpha processor, communication power takes 33% in RAW architectures [5]. [9] reported that 28%∼36% of the total chip power consumption depends on NoC energy (e.g., on-chip switches and links) consumption

5. MILP & MIQP
If the problem contains an objective function with no quadratic term, (a linear objective), then the problem is termed a Mixed Integer Linear Program(MILP)*.
However, if there is a quadratic term in the objective function, the problem is termed a Mixed Integer Quadratic Program (MIQP)*.
In our work, MIQP problem is using to find a global optimum solution and so the MILP & MIQP problems are solved using the IBM CPLEX optimizer.
* IBM® Knowledge Center

6. Power Models
The total execution time (ExecTimej) and energy consumption (Energy) of a job j on a CMOS-based processor can be estimated by:
ExecTime j ≈ 1 𝑓 and Power ≈ C ∙ 𝑓∙ V2
Energy= Power ∙ ExecTime
Since in DVFS systems 𝑉 varies approximately with 𝑓 (𝑉 ∝ 𝑓), the performance-energy trade-off of frequency scaling can be expressed as ExecTime j ∝ 1 𝑓 and Energy ∝ 𝑓 2 [25].
𝑉 ∝ 𝑓 →Power ≈ 𝑓3 → Energy= 𝑓3 ∙ExecTime j

7. The inputs models for meta-scheduling
AM with constants of application (e.g., number of jobs, WCET of each job, precedence constraints).
PM with constants of the platform (e.g., number of nodes, links between nodes).
SM with values for decision variables (e.g., allocation to node, start time of execution).
CM with fault and events details(e.g., event type, event execution time).

8. Static-Scheduling Collision Avoidance Constraint
Connectivity Constraints
Job Allocation
Hop Count
Path and Visited Cores

9. Meta-Scheduling Techniques for Energy-Efficient, Robust and Adaptive Time-Triggered Systems
This technique is trying to solve scheduling problems in scenario-based scheduling and quasi-static task mapping which in some papers is called super-scheduling.
In this work, a Meta-Scheduler (MES) tool to solve meta-scheduling problems developed.
We propose an energy-efficient scheduling algorithm regarding dynamic-slack and slowdown factors in multi-scenario-based systems for NoC-based MPSoC.

10. Scheduling Decision Variables
Slow Down Factors: SlowDF= j1 ⋮ ⋮ jm ∈ 𝑀𝑖𝑛,…,𝑀𝑎𝑥 ∧𝑀𝑖𝑛>0
Scheduling Constraints
Job Dependency Constraints : INJECTIME[m1] + EXECTIME[j1] ∙SLOWDF[j1]+ ( ( HOPS[m1]+1 ) ∙DUR[m1] ) <= INJECTIME[m2]
EXECTIME[j1]*SLOWDF[j1] <= INJECTIME[m2]
Message Duration : 𝐷𝑢= du1 ⋮ ⋮ dum ∈ {1,…,𝑀𝑎𝑥} 𝑚
Message Deadlines : ∀j1 ∈ 1,…,𝑗 , ∀m1 ∈ 1,…,𝑚 ,∀ℎm1 ∈ {1,…,𝑀𝑎𝑥H } : 𝑖m1+ ℎm1+1 ∙𝐷𝑢m1≤𝐷j1

11. 𝐶𝑃 job1 =(ExecTime j1∙(SlowDF j1) 2 )
Objective Function
The objective is to maximize the energy-efficiency. In other words, it is minimizing energy consumption by increasing job execution times regarding.
∀ job1 ∈ {1,…,j } :
𝐶𝑃 job1 =(ExecTime j1∙(SlowDF j1) 2 )
𝑚𝑎𝑥𝑖𝑚𝑖𝑧𝑒 ( j=1 𝑗 𝐶𝑃 )

12. Visualization of schedules
Meta-Scheduling Visualizer (MeSViz) is an advanced and applicative tool which is designed and planned to visualize and display SM contents (e.g., jobs, cores, messages, allocation, assignment.
For calculating the energy consumption of static-slack SM (EnergySM), an average of all dynamic-slack SM’s (EAvgSMdynamic) and compare the results (SavingE) MeSViz is used follows the formula:
EnergySM= ETji SlowDFji , EAvgSMdynamic= 𝑛=1 𝑛 ESMn 𝑛
SavingE=( ESMstatic − EAvgSMdynamic ESMstatic ).100

13. Example Scenario And Results
Input Model
Input Name ID
Data AM
Job
WCET=2 1
WCET=4 2
WCET=6 3
WCET=8 4
WCET=10
Message
Quantity=6 ID start=0
Deadline
All job= 1185
Slowdf
All job: min =1 & max =100
SlackEvent
All job= 50% PM
Hop (Switch)
Quantity=2 ID start=0
Core
Quantity=5 ID start=6
Link

14. Results Static 1.875 Dynamic 1.62 13.6% 0.6675 64.4% 1.0948 41.61%
SM ID
Slack Mode
Energy Consumption
Energy
Saving
Static
1.875 49
Dynamic
1.62
13.6% 5
0.6675
64.4%
Average
1.0948
41.61%
SM ID- Note
Slack Mode
Job ID ET
SLOWDF
Static 2 4 1 6 3 8 10
49 (Minimum Energy saving)
Dynamic 5
5 (Maximum Energy Saving)

15. Conclusion
The simulation results show that our dynamic-slack algorithm, compared to the static-slack, produces a maximum of 64.4% in a single schedule and 41.61% energy-saving of NoCs on average:
A novel MeS technique is used to reduce the dynamic power consumption in the schedules. MeS can be used in scenario-based (fault, safety, power-saving) scheduling and adaptivity TT systems.
further work still needs to be done in the future (e.g., energy-saving on communication links, fault-injection and more scenarios related to safety).
ACKNOWLEDGMENT
The Europian project H2020 SAFEPOWER has supported this work under the Grant Agreement No

16. Thank you for your attention. Babak. Sorkhpour@uni-siegen
Thank you for your attention.