1. BFSBFS by Con KolivasCon Kolivas Guruprasad Aphale. Real Time Lunch, 10/21/2009 1 Guruprasad Aphale.

2. Outline What is BFS? CFS vs. BFS Design of BFS ▫ Design Details ▫ Design Decisions Performance Evaluation ▫ Comparison with CFS Questions?? 2 Guruprasad Aphale.

3. Why BFS?? [1] 3 Guruprasad Aphale.

4. What is BFS? A “Fair” Scheduler by Con Kolivas. Simple, Basic Design Forward Looking Only Time slice independent of sleep time ▫ Fixed Time slice length Desktop Oriented Scheduler ▫ Meant for low spec machines ▫ Not supposed to be scaled for large hardware ▫ NUMA unaware 4 Guruprasad Aphale.

5. FeatureCFSBFS Scheduling Complexity O(log n)O(n) Ideal ForBig MachinesDesktop to mid-size (No. of Logical CPUs < 16) Tunable Knobs8 [2] 1 Run QueuesMultipleSingle Run Queue Implementation Red Black TreesDoubly Linked Lists Design ComplexityComplexSimple, Basic Code ComplexityComplexLess than CFS (Number of lines also reduced by more than 4000) Comparison of CFS and BFS 5 Guruprasad Aphale.

6. Design Details Virtual Deadline Task Insertions Task Selection Locking 6 Guruprasad Aphale.

7. Virtual Deadline “round robin interval” (rr_interval) ▫ Length of time slice in ms ▫ Configurable ▫ Modified based on number of CPUs  rr_interval *= 1 + lg(num_online_cpus) When a task is created, it is given a time slice and a virtual deadline Virtual Deadline = jiffies + (user_priority * rr_interval) user_priority : nice value + 20 (in range 0..39) 7 Guruprasad Aphale.

8. Virtual Deadline Continued.. The deadline is a virtual one ▫ No guarantee that a task will be scheduled by this time. ▫ Used to compare which task should go next Events causing rescheduling ▫ time_slice Exhausted ▫ sleep ▫ preemption 8 Guruprasad Aphale.

9. Virtual Deadline Continued… After swapping a task out ▫ put back on queue ▫ task with earliest deadline is chosen from queued but not running tasks. A problem with this approach ▫ Task’s deadline may already pass before it is chosen for execution. ▫ Indication that the task needs CPU time ahead of all later deadlines. 9 Guruprasad Aphale.

10. Task Insertion Has only one global queue which contains sub-queuesglobal queue Inserts task into relevant queue ▫ Each insertion takes O(1) time as inserting into doubly linked list. Check if new task can run on any idle CPU or can preempt any other running low priority task. ▫ Don’t wait for schedule to be called if a cpu is idle or low priority task is running Worst case time O(M), where M is number of CPUs. 10 Guruprasad Aphale.

11. 0001010 Bitmap Array of sub queue heads 1-100 RT tasks SCHED_NORMAL SCHED_ISO SCHED_IDLEPRIO t1t1 titi tjtj tntn Insertion t n+1 New Task to be inserted. SCHED_ISO 1 t n+2 New Task to be inserted. SCHED_RR (3 rd Sub-queue) 11 Guruprasad Aphale.

12. Task Selection Check bitmaps to find first set bit. If the first bit corresponds to a RT sub queue, select first task in the queue If no bit set in RT queue, then search for next task based on EVDF and suitable CPU affinity. ▫ If a task with expired deadline found, return that task. In most cases, entire queue lookup is not required. 12 Guruprasad Aphale.

13. 0000010 Bitmap Array of sub queue heads 1-100 RT tasks SCHED_NORMAL SCHED_ISO SCHED_IDLEPRIO d=7 d =5 Selection Task removed from queue and assigned to cpu 13 Guruprasad Aphale.

14. Locking Only one queue, thus only one lock protecting the process data So for every operation that modifies the queue, global lock has to be acquired. Once a task is assigned to a CPU, CPU removes it from queue and creates a local copy of task’s data structure. Thus updates are lockless Max number of tasks in global queue = number of tasks remaining – number of CPUs + 1 14 Guruprasad Aphale.

15. Design Decisions Single Run Queue ( O(n) lookup) ▫ Refers to queued but not running processes ▫ Absence of multiple queues results in no complex interactions among them Scheduling Decisions based only on CPU Usage ▫ No sleep time taken into consideration during scheduling decisions ▫ No interactivity estimator present 15 Guruprasad Aphale.

16. Design Decisions Interactivity in BFS ▫ The tasks that are waking up try to preempt same priority tasks. ▫ Even if they do not preempt, the waiting time for a task is still bounded(because of rr_interval), thus it is scheduled within a timeframe. 16 Guruprasad Aphale.

17. Performance Evaluation [3] TestCFSBFS Timed Apache Compilation 131.48 Seconds122.76 Seconds Timed PHP Compilation197.90 Seconds197.21 Seconds 7-Zip Compression2180.66 MIPS2104.00 MIPS Apache Benchmark1902.77 Requests Per Second 3140.67 Requests Per Seconds Threaded I/O Tester (Read Filesize Per Thread : 16 MB, Thread Count : 8) 14.66 ms16.00 ms System Configuration Dual-core Intel Atom 330 CPU with Hyper-Threading clocked at 2.10GHz Good for Interactive Desktops? People have been doing a lot of benchmarking. But the results are not conclusive. 17 Guruprasad Aphale.

18. Thank You Questions??? 18 Guruprasad Aphale.

19. References 1.http://www.linux- magazine.com/Online/News/Con-Kolivas- Introduces-New-BFS-Schedulerhttp://www.linux- magazine.com/Online/News/Con-Kolivas- Introduces-New-BFS-Scheduler 2.http://www.ibm.com/developerworks/linux/libra ry/l-cfs/http://www.ibm.com/developerworks/linux/libra ry/l-cfs/ 3.http://www.phoronix.com/scan.php?page=article &item=bfs_scheduler_benchmarks&num=1http://www.phoronix.com/scan.php?page=article &item=bfs_scheduler_benchmarks&num=1 19 Guruprasad Aphale.

20. Global Queue Structure Contains 103 subqueues ▫ 100 RT queues ▫ SCHED_ISO queue ▫ SCHED_NORMAL queue ▫ SCHED_IDLEPRIO queue ▫ Each sub queue is a doubly linked list. While setting up a new task, a bitmap of running priorities is set. Each bit corresponds to a sub queue inside global queue. This shows which priorities have waiting tasks. 20 Guruprasad Aphale.

21. 0001010 Bitmap Array of sub queue heads 1-100 RT tasks SCHED_NORMAL SCHED_ISO SCHED_IDLEPRIO t1t1 titi tjtj tntn Task Sub queues Global Run Queue Structure (Only Queues) 21 Guruprasad Aphale.