![Non-blocking QueueTB 1TB 1HeadTB 2TB 2T1T2T3T4TailTB nReferenceP. Tsigas and Y. Zhang, A simple, fast and scalable non-blocking concurrent FIFO queue for shared memory multiprocessor systems[SPAA01]](https://image.slidesharecdn.com/cedermanondynamicloadbalancingongraphicsprocessors-12516516004696-phpapp02/85/Dynamic-Load-balancing-On-Graphics-Processors-19-320.jpg)
![Task stealingT1TB 1T3T2TB 2TB nReferenceArora N. S., Blumofe R. D., Plaxton C. G. , Thread Scheduling for Multiprogrammed Multiprocessors [SPAA 98]](https://image.slidesharecdn.com/cedermanondynamicloadbalancingongraphicsprocessors-12516516004696-phpapp02/85/Dynamic-Load-balancing-On-Graphics-Processors-25-320.jpg)
Dynamic Load-balancing On Graphics Processors
- 1. On Dynamic Load Balancing on Graphics ProcessorsDaniel Cederman and Philippas TsigasChalmers University of Technology
- 10. Task sharingCheck conditionWork done?DoneTask SetAcquire TaskTry to get taskTaskGot task?No, retryTaskTaskPerform taskTaskNew tasks?No, continueAdd TaskTaskAdd task
- 11. System ModelCUDAGlobal MemoryGather and scatterCompare-And-SwapFetch-And-IncMultiprocessorsMaximum number ofconcurrent thread blocksGlobal MemoryMulti-processorMulti-processorMulti-processorThread BlockThread BlockThread BlockThread BlockThread BlockThread BlockThread BlockThread BlockThread Block
- 12. SynchronizationBlockingUses mutual exclusion to only allow one process at a time to access the object. LockfreeMultiple processes can access the object concurrently. At least one operation in a set of concurrent operations finishes in a finite number of its own steps.WaitfreeMultiple processes can access the object concurrently. Every operation finishes in a finite number of its own steps.
- 13. Load Balancing MethodsBlocking Task QueueNon-blocking Task QueueTask StealingStatic Task List
- 14. Blocking queueFreeTB 1HeadTB 2TailTB n
- 15. Blocking queueFreeTB 1HeadTB 2TailTB n
- 16. Blocking queueFreeTB 1HeadTB 2T1TailTB n
- 17. Blocking queueFreeTB 1HeadTB 2T1TailTB n
- 18. Blocking queueFreeTB 1HeadTB 2T1TailTB n
- 19. Non-blocking QueueTB 1TB 1HeadTB 2TB 2T1T2T3T4TailTB nReferenceP. Tsigas and Y. Zhang, A simple, fast and scalable non-blocking concurrent FIFO queue for shared memory multiprocessor systems[SPAA01]
- 20. Non-blocking QueueTB 1TB 1HeadTB 2TB 2T1T2T3T4TailTB n
- 21. Non-blocking QueueTB 1TB 1HeadTB 2TB 2T1T2T3T4TailTB n
- 22. Non-blocking QueueTB 1TB 1HeadTB 2TB 2T1T2T3T4TailTB n
- 23. Non-blocking QueueTB 1TB 1HeadTB 2TB 2T1T2T3T4T5TailTB n
- 24. Non-blocking QueueTB 1TB 1HeadTB 2TB 2T1T2T3T4T5TailTB n
- 25. Task stealingT1TB 1T3T2TB 2TB nReferenceArora N. S., Blumofe R. D., Plaxton C. G. , Thread Scheduling for Multiprogrammed Multiprocessors [SPAA 98]
- 26. Task stealingT1T4TB 1T3T2TB 2TB n
- 27. Task stealingT1T4T5TB 1T3T2TB 2TB n
- 28. Task stealingT1T4TB 1T3T2TB 2TB n
- 29. Task stealingT1TB 1T3T2TB 2TB n
- 30. Task stealingTB 1T3T2TB 2TB n
- 31. Task stealingTB 1T2TB 2TB n
- 33. Static Task ListInT1TB 1T2TB 2T3TB 3T4TB 4
- 34. Static Task ListInOutT1TB 1T2TB 2T3TB 3T4TB 4
- 35. Static Task ListInOutT1T5TB 1T2TB 2T3TB 3T4TB 4
- 36. Static Task ListInOutT1T5TB 1T2T6TB 2T3TB 3T4TB 4
- 37. Static Task ListInOutT1T5TB 1T2T6TB 2T3T7TB 3T4TB 4
- 43. Graphics Processors8800GT14 Multiprocessors57 GB/sec bandwidth9600GT8 Multiprocessors57 GB/sec bandwidth
- 44. Blocking Queue – Octree/9600GT
- 45. Blocking Queue – Octree/8800GT
- 46. Blocking Queue – Four-in-a-row
- 47. Non-blocking Queue – Octree/9600GT
- 48. Non-blocking Queue – Octree/8800GT
- 49. Non-blocking Queue - Four-in-a-row
- 50. Task stealing – Octree/9600GT
- 51. Task stealing – Octree/8800GT
- 52. Task stealing – Four-in-a-row
- 53. Static List
- 55. Previous workKorch M., Raubert T., A comparison of task pools for dynamic load balancing of irregular algorithms, Concurrency and Computation: Practice & Experience, 16, 2003Heirich A., Arvo J., A competetive analysis of load balancing strategies for parallel ray tracing, Journal of Supercomputing, 12, 1998Foley T., Sugerman J., KD-tree acceleration structures for a GPU raytracer, Graphics Hardware 2005
- 56. ConclusionSynchronization plays a significant role in dynamic load-balancingLock-free data structures/synchronization scales well and looks promising also in the GPU general purpose programmingLocks perform poorly It is good that operations such as CAS and FAA have been introduced in the new GPUsWork stealing could outperform static load balancing