SACR: Scheduling-Aware Cache Reconfiguration for Real-Time Embedded Systems

1. SACR: Scheduling-Aware Cache Reconfiguration for Real-Time Embedded Systems Weixun Wang and Prabhat Mishra Embedded Systems Lab Computer and Information Science and Engineering University of Florida Ann Gordon-Ross Electrical and Computer Engineering University of Florida

2. Outline • Introduction • Related Work • Scheduling-Aware Cache Reconfigurations • Phase-based Optimal Cache Selection • Scheduling-Aware Dynamic Reconfigurations • Experiments • Conclusion

3. Introduction • Real-time embedded systems • Energy constraints (battery operated) • Time constrained • Hard task deadlines • Safety-critical systems • Soft task deadlines • Gaming, multimedia, housekeeping devices have soft deadlines • Deadline miss results in temporary service/quality degradation • Dynamic cache reconfigurations • Promising for improving energy and performance • Not applicable in real-time systems • Dynamic computation is expensive • Dynamic reconfiguration leads to timing uncertainty Propose a scheduling-aware dynamic cache reconfiguration technique to generate significant energy savings in soft real-time systems by exploiting static analysis during runtime.

4. Outline • Introduction • Related Work • Scheduling-Aware Cache Reconfigurations • Phase-based Optimal Cache Selection • Scheduling-Aware Dynamic Reconfigurations • Experiments • Conclusion

5. Related Work • Energy-aware task scheduling techniques • Early Deadline First (EDF) and Rate Monotonic (RM) • Dynamic Voltage Scaling (DVS) • Jejurikar et al. [IEEE TCAD 06’], Quan et al. [ACM TECS 07’] • Caches in Real-Time Systems • Cache locking and Cache Partitioning • Puant [RTSS 02’] and Wolfe [IWRCS 1993] • Cache-related preemption delay analysis • Tan et al. [ACM TECS 2007] • Reconfigurable Cache Architectures • Reconfigurable cache architecture • Zhang et al. [ACM TECS 05’] • Application-based vs. Phase-based tuning • Gordon-Ross et al. [ISLPED 05] vs. Sherwood et al. [Micro 03]

6. Outline • Introduction • Related Work • Scheduling-Aware Cache Reconfigurations • Phase-based Optimal Cache Selection • Scheduling-Aware Dynamic Reconfigurations • Experiments • Conclusion

7. Overview Task 1 Task 2 In traditional real-time systems In our approach

8. Phase-based Optimal Cache Selection • A task is divided by n potential preemption points • A phase is the period of time between a predefined potential preemption point and task completion • Each phase has its optimal cache configuration • Performance-optimal and energy-optimal • A static profile table is generated for each task Pn-1 P2 P1 0 Task Execution Time phase n (n-1/n) Cn phase 3 (2/n) C3 phase 2 (1/n) C2 phase 1 (0/n) C1

9. Phase-based Optimal Cache Selection • Potential preemption points may not be the same as actual preemption points. • They are used for cache configuration selection. • Partition factor determines the potential preemption points and resulting phases • Large partition factor leads to large look-up table • Not feasible due to area constraints • Large partition factor may not save more energy • Partition factor around 4 to 7 is profitable

10. Outline • Introduction • Related Work • Scheduling-Aware Cache Reconfigurations • Phase-based Optimal Cache Selection • Scheduling-Aware Dynamic Reconfigurations • Statically Scheduled Systems • Dynamically Scheduled Systems • Experiments • Conclusion

11. Scheduling-Aware Cache Reconfiguration • Statically scheduled systems • Arrival times, execution times, and deadlines are known a priori for each task • Statically profile energy-optimal configurations for every execution period of each task without violating any task deadlines • Dynamically scheduled systems • Task preemption points are unknown • New tasks can enter the system at any time • Conservative approach • Aggressive approach

12. Conservative Approach • Energy-optimal cache configuration with equal or higher performance than base cache • Nearest-neighbor • Use the nearest partition point to decide which cache configuration to tune to • Static Profile table • Deadline-aware energy-optimal configurations • Task list entry • Runtime information

13. Aggressive Approach • Energy-optimal cache configuration & Performance-optimal cache configuration • Includes their execution time as well. • Ready task list (RTL) • Contains all the tasks currently in the system • Static Profile table • Energy-opt configuration • Perf.-optimal configuration • Task list entry • Runtime information

14. Outline Introduction Related Work Scheduling-Aware Cache Reconfigurations Phase-based Optimal Cache Selection Scheduling-Aware Dynamic Reconfigurations Experiments Conclusion

15. Experimental Setup • SimpleScalar to obtain simulation statistics • Used external I/O (eio) trace file, checkpointing and fastforwarding to generate static profile table • Energy model • Zhang et al. and CACTI 4.2 • Benchmarks • EEMBC • MediaBench

16. Energy Savings (Instruction Cache) 28% average energy savings using conservative approach 51% average energy saving using aggressive approach

17. Energy Savings (Data Cache) 17% average energy savings using conservative approach 22% average energy saving using aggressive approach 17

18. Hardware Overhead • Profile Table stores 18 cache configurations • Synthesized using Synopsys Design Compiler • Assumed lookup frequency of one million nanoseconds • Table lookup every 500K cycles using 500 MHz CPU • Average energy penalty is 450 nJ • Less than 0.02% of overall savings (2825563 nJ)

19. Conclusion • Dynamic cache reconfiguration is a promising approach to improve both energy consumption and overall performance. • Developed a scheduling aware dynamic cache reconfiguration technique • On average 50% reduction in overall cache energy consumption in soft real-time systems • Future work • Hard real-time systems • Multi-core and multi-processor systems

20. Thank you !

21. Aggressive Approach • When task T is the only task in the system • Always tune to energy-optimal cache if possible • When task T preempts another task • Run schedulability check • Discard the lowest priority task if absolutely necessary • Tune to energy-optimal cache • if all other tasks in RTL can meet their deadlines using their performance-optimal caches • When task T is preempted by another task • Calculate and store runtime information (RIN, CP)