![Power Wall Problem
[Patterson & Hennessy]](https://image.slidesharecdn.com/urmiauniseminar-pdf-150721072923-lva1-app6891/85/Sara-Afshar-Scheduling-and-Resource-Sharing-in-Multiprocessor-Real-Time-Systems-29-320.jpg)
Sara Afshar: Scheduling and Resource Sharing in Multiprocessor Real-Time Systems
- 1. Resource Sharing in Real-Time Uni/Multiprocessor Embedded Systems Sara Afshar sara.afshar@mdh.se
- 2. Stockhol m Umeå Göteborg Malmö Västerås Eskilstuna About Me! • BSc. in Electrical Engineering in Tabriz University • Master in Intelligent Embedded Systems • Mälardalen University, Västerås, Sweden • Currently, PhD candidate (3rd year) • Research topic: Resource sharing in real-time multiprocessors
- 3. About MRTC! Professors: 29 Senior researchers: 52 PhD Students: 76 Research Groups: 16 Reference: www.es.mdh.se
- 4. Outline • Embedded systems • Real-time systems • Scheduling and timing analysis • Resource sharing in uniprocessors • Multiprocessors • Open problems
- 5. Daily Computers • A special-purpose computer • Computers perform the functionality
- 6. Embedded Systems A special-purpose computer that performs a few dedicated functions, usually with very specific requirements.
- 7. Real-Time Systems Embedded systems: – Specialized – Efficient But in many cases it is not enough. The system has to react to the environment at the right time instance. Timing Requirements
- 8. Real-Time System Non-Real-Time Systems Real-Time Systems The system does the right thing The system does the right thing It is on-time &
- 9. Example Airbag example (a very classical example!) time Too early Too late Collision Real-Time ≠ Fast Real-Time = Predictable
- 10. Hard vs. Soft Real-Time Each program has a deadline which should not be missed. Hard Real-Time – Missing the deadline cause catastrophe – E.g., automotive, airplane, industrial control Soft-Real-Time – Can miss some deadlines – E.g., TV, video streaming
- 11. Real-Time Tasks • Program is written by means of different tasks • On a single-core processor two tasks cannot execute in parallel – Some tasks are preempted in order that all tasks execute on time – Scheduler is responsible to schedule tasks in a processor sens Task A sens sens sens sens sens sens sens sens sens
- 12. Real-Time Tasks Periodic Tasks: repeat in periodic intervals e.g., control loops, sensor reading, etc. Execute Execute ExecuteSleeping Sleeping 𝑇𝑖 𝑇𝑖
- 13. Real-Time Tasks Aperiodic Task: tasks may arrive at any point in time e.g., alarm tasks, emergency button, etc. Execute The task executes once An interrupt event occurs, e.g., a button is pushed! Task may be trigerred again at some point in time!
- 14. Real-Time Tasks Sporadic Task: Similar to aperiodic tasks, however, minimum time for task’s next activation is known e.g., task handling keyboard input- minimum time between pressing two keys known Minimum inter arrival time = known Execute𝑇𝑖 Execute ?
- 15. Scheduling The process of deciding the execution order of real- time tasks, depends of the priority of the task. There are different mechanism to do that. Task A misses its deadline Both deadlines are met
- 17. Scheduling algorithms • Fixed Priority – RM: smaller priod higher priority – DM: smaller deadline higher priority • Dynamic Priority – EDF: earliest deadline first
- 18. Task Model 𝑎𝑖 𝐷𝑖 𝑓𝑖𝑇𝑖 • Arrival time (release time): 𝑎𝑖 = 1 • Execution time: 𝐶𝑖 = 3 • Finishing time: 𝑓𝑖 =7 • Deadline: 𝐷𝑖 = 7 • Period: 𝑇𝑖 = 8 • Response time: 𝑅𝑖 = 𝑓𝑖 − 𝑎𝑖 = 6 𝜏𝑖 𝜏𝑖 𝜏𝑖
- 19. Response Time Analysis A A A A A H H H H Response Time of Task A Ready time of A Finishing time of A W𝑅 𝐴 = 𝐶𝐴 + ∀𝑗∈ℎ𝑝(𝐴) 𝑊𝑅 𝐴 𝑇 𝑗 × 𝐶𝑗 Interference from higher priority tasks Worst-case response time
- 20. Resource Sharing • Some tasks are independent • Some tasks are aware of each other – E.g., using a shared memory – E.g., two tasks writing in a same buffer
- 21. Resource Sharing • Tasks may use hardware/software components such as a database, hard-drive, sensor etc. PROBLEM critical section = part of the task execution that access to a resource
- 22. • In real-time systems, semaphore-based locking synchronization techniques handles mutually- exclusive access to resources among tasks – Every task that wants to use a resource first has to lock the resource, use it and then unlock (release) the resource Resource Sharing critical section
- 23. Resource Sharing • Tasks may experience delay due to resource sharing – E.g., the task that needs the same resource has to wait for the resource to be released by other tasks critical section BLOCK
- 24. Resource Sharing • Blocking can endanger system correctness – Priority inversion: a high priority task (in this example task 1) is forced to wait for a lower priority task (in this example task 2) for an unbounded amount of time http://www.idt.mdh.se/kurser/CDT315/index.php?choice=contents critical section high priority task already missed its deadline due to extra waiting for middle prio. task Task 1: priority = Low Task 2 priority = Middle Task 3: priority = High deadline extra delay due to normal execution of middle prio. task which can be considerably long compared to critical sections high prio. task requests the same resource which is not available and is blocked high prio. task arrives and preempt the low priority task middle prio. task arrives and preempts low priority task low prio. task continues middle prio. task finishes low prio. task continues and then releases th resource
- 25. Resource Sharing • Blocking can endanger system correctness – Priority inversion: high priority task is forced to wait for a lower priority task for an unbounded amount of time • Mars Pathfinder – Landing on July 4, 1997 – Pathfinder experiences repeated resets after staring gathering of meteorological data. – Resets caused by timing overruns when using shared communication buss- a classical case of priority inversion problem.
- 26. PIP • Priority Inheritance Protocol (PIP) – High priority task cannot be delayed by middle priority task http://www.idt.mdh.se/kurser/CDT315/index.php?choice=contents critical section Task 1: priority = Low Low priority task inherits priotiy of high priority task Middle priority task arrives prio. = Low prio. = High prio. = Low Task 2 priority = Middle Task 3: priority = High Low priority task gets back its own priotiy :high priority task meets its deadline deadline
- 27. Synchronization Protocols • PIP: priority inversion • PCP: deadlock, chain blocking • IPIP, IPCP, SRP: blocking only in the beginning
- 28. Ready time of A • By enabling resource sharing, a blocking term is added to the worst-case response time of a task Response Time Analysis A A A A A H H H H Response Time of Task A Finishing time of A L W𝑅 𝐴 = 𝐶𝐴 + 𝐵𝐴 + ∀𝑗∈ℎ𝑝(𝐴) 𝑊𝑅 𝐴 𝑇 𝑗 × 𝐶𝑗 Blocking incurred to task A
- 29. Power Wall Problem [Patterson & Hennessy]
- 30. Multiprocessors • Integration of multiple processors on a chip • Multiprocessors platforms have become popular in the industry – Power consumption – Performance
- 31. Multiprocessors • Integration of multiple processors on a chip • Multiprocessors platforms have become popular in the industry – Power consumption – Performance • Migrating to multiprocessor technology – Immature scheduling and synchronization techniques – Over simplification
- 32. Multiprocessor Scheduling • Partitioned scheduling: – Tasks are fixed assigned to processors in design time – Each processor has its own scheduler and ready queue – Task migration among processors is not allowed in run-time local ready queues per-processor schedulers processors
- 33. Multiprocessor Scheduling • Global scheduling – Only one scheduler and one ready queue – Scheduler assign tasks to processors in run-time – Task migration among processors is allowed global ready queue global scheduler processors
- 34. Multiprocessor Scheduling • Hybrid scheduling, semi-partitioned scheduling – Combination of partitioned and global scheduling • Most tasks are fixed assigned to processor in design time • A few tasks can migrate among processors in run-time – Benefits from advantages of both approaches … task partitions local ready queues per-processor schedulers … … processors partitioned task migrating task
- 35. Resource Sharing critical section L M H P1 P2 • More complex problem in multiprocessors – Remote Blocking is added besides local blocking deadline !:high priority task misses its deadline
- 36. Response Time Analysis W𝑅𝑖 = 𝐶𝑖 + 𝐵𝑖 + ∀𝑗∈ℎ𝑝(𝑖) 𝑊𝑅 𝑖 𝑇 𝑗 × 𝐶𝑗 • Two type of blocking in multiprocessors: – Local blocking – Remote blocking 𝐵𝑖 𝐿 𝐵𝑖 𝑅
- 37. Synchronization Protocol • Various synchronization protocols: – MSRP: P-EDF – M-BWI: G-EDF – FMLP short: P/G-EDF – MrsP: P-FP – MPCP: P-RM – FMLP long: P/G-EDF – OMLP: P/G-EDF, P-FP – MSOS: P-FP
- 38. Synchronization Protocol • Queue type: access priority to the resource – FIFO, LIFO, PR-Based, Hybrid • Task behavior during waiting – Spin-based, suspension-based • Task priority changes – Inherited, boosted
- 39. Synchronization Protocol • These protocols proposed for global and partitioned scheduling • No protocol for hybrid (semi-partitioned) scheduling
- 41. Semi-Partitioned Scheduling • Partitioned tasks: tasks that are fixed assigned to processors and execute only on those processors, i.e., they fit (utilization-wise) on processors during partitioning (𝜏1, … , 𝜏8) • Migrating tasks: Task(s) that do(es) not fit on any processor (𝜏9) 𝜏1 𝜏2 𝜏3 𝜏4 𝜏5 𝜏6 𝜏7 𝑃1 𝑃2 𝑃3 𝜏8 𝜏9
- 42. Semi-Partitioned Scheduling • Migrating tasks split among processors which can provide capacity remained from partitioning (𝜏9 splits among processors 1 to 3) 𝜏1 𝜏2 𝜏3 𝜏4 𝜏5 𝜏6 𝜏7 𝑃1 𝑃2 𝑃3 𝜏8 𝜏9
- 43. Resource Sharing • Variation in execution time of tasks • May cause critical sections to happen at any point during task execution • In case of semi-partitioned scheduling, critical sections may happen in any part of a split task and therefore on any processor that the task is split over cs1 P1 P2 cs2 P3 cs2 P1 P2 cs1 P3 P1 P2 cs1 P3 cs2 cs1 Case 1 Case 2 Case 3
- 44. Resource Sharing • Therefore, in semi-partitioned scheduling, existing synchronization protocols cannot be used directly cs1 P1 P2 cs2 P3 cs2 P1 P2 cs1 P3 P1 P2 cs1 P3 cs2 cs1 Case 1 Case 2 Case 3
- 45. Centralized Solution • Critical sections migrate to marked processor • Advantages: – Centralize resource access – Remote blocking on marked processor • Disadvantages: – Extra migration overhead P1 P2 Rs 1 i 2 i marked processor 2 i 2 i 2 i P2 P1 migration overhead non-split tasks split task
- 46. Decentralized Solution • Critical sections served where they occur • Advantages: – Decreased migration overhead • Disadvantages: – Introduced blocking to local tasks – Increases remote blocking non-split tasks split task P1 P2 Rs 1 i 2 i 2 i 2 i P2 P1Rs 1 i 1 i
- 47. Analysis • Local blocking due to local resources 𝐵𝑖,1 = min{𝑛𝑖 𝐺 + 1, 𝜌 𝑖<𝜌 𝑗 𝑇𝑖 𝑇𝑗 + 1 𝑛𝑗 𝐿 (𝜏𝑖)} max 𝜌 𝑖<𝜌 𝑗 𝜋 𝑖,𝜋 𝑗∈𝑃 𝑘 𝑅 𝑙∈𝑅 𝑃 𝑘 𝐿 𝜌 𝑖≤𝑐𝑒𝑖𝑙(𝑅 𝑙) {𝐶𝑠𝑗,𝑙} 𝑐𝑒𝑖𝑙 𝑅𝑙 = max{𝜌2 𝜏2 ∈ 𝜏𝑙,1}
- 48. Analysis • Local blocking due to global resources 𝐵𝑖,2 = ∀𝜌 𝑗<𝜌𝑖 𝜏 𝑖,𝜏 𝑗 𝜖𝑃 𝑘 min{𝑛𝑖 𝐺 + 1, ( 𝑇𝑖 𝑇𝑗 + 1)𝑛𝑗 𝐺 } max 𝑅 𝑞∈𝑅 𝑃 𝑘 𝐺 {𝐶𝑠𝑗,𝑞}
- 49. Analysis • Remote blocking due to lower priority tasks 𝐵𝑖,3 = ∀𝑅 𝑞∈𝑅 𝑃 𝑘 𝐺 𝜏 𝑖 𝜖𝜏 𝑞,𝑘 𝑛𝑖,𝑞 𝐺 max ∀𝜌 𝑗<𝜌𝑖 𝜏 𝑗 𝜖𝜏 𝑞,𝑟 𝑘≠𝑟 {𝐶𝑠𝑗,𝑞 + 𝜌ℎ,𝑗 𝑅 𝑞 ′ max 𝜏 𝑡∈𝑃𝑟 𝜌 𝑡>𝜌 𝑗 𝑅 𝑠∈𝑅 𝑟 𝐺 𝑠≠𝑞 {𝐶𝑠𝑡,𝑠}}
- 50. Analysis • Remote blocking due to higher priority tasks 𝐵𝑖,4 = ∀𝑅 𝑞∈𝑅 𝑃 𝑘 𝐺 𝜏 𝑖 𝜖𝜏 𝑞,𝑘 ∀𝜌 𝑗>𝜌 𝑖 𝜏 𝑗 𝜖𝜏 𝑞,𝑟 𝑘≠𝑟 𝑛𝑗,𝑞 𝐺 ( 𝑇𝑖 𝑇𝑗 + 1)(𝐶𝑠𝑗,𝑞 + 𝜌ℎ,𝑗 𝑅 𝑞 ′ max 𝜏 𝑡∈𝑃𝑟 𝜌 𝑡>𝜌 𝑡𝑗 𝑅 𝑠∈𝑅 𝑟 𝐺 𝑠≠𝑞 {𝐶𝑠𝑡,𝑠})
- 51. Evaluation results (a) Overhead = 0 µs (b) Overhead = 140 µs 0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 5 25 45 65 85 105 125 145 165 185 205 Schedulability Critical Section Length (µs) MLPS NMLPS 0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 5 25 45 65 85 105 125 145 165 185 205 Schedulability Critical Section Length (µs) MLPS NMLPS : centralized solution : decentralized solution : centralized solution : decentralized solution
- 52. Details in Paper “A Resource Sharing under Multiprocessor Semi- Partitioned Scheduling.” Sara Afshar, Farhang Nemati, Thomas Nolte. In proceedings of the 18th IEEE International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA), 2012, August.
- 53. Open Problems • Improving synchronization techniques – Improving of analysis – Improving of protocols • Blocking aware partitioning • Compositional scheduling
- 54. THANK YOU! 54