1. Atlanta, Georgia TiNy Threads on BlueGene/P: Exploring Many-Core Parallelisms Beyond The Traditional OS Handong Ye, Robert Pavel, Aaron Landwehr, Guang R. Gao Department of Electrical & Computer Engineering University of Delaware 2010-04-23 MTAAP’20101

2. Introduction Modern OS based upon a sequential execution model (the von Neumann model). Rapid progress of multi-core/many-core chip technology. –Parallel Computer systems now implemented on single chips. MTAAP’20102

3. Introduction Conventional OS model must adapt to the underlying changes. Further exploit the many levels of parallelism. –Hardware as well as Software We introduce a study on how to do this adaptation for the IBM BlueGene/P multi-core system. MTAAP’20103

4. Outline Introduction Contributions TNT on BlueGene/P Scheduling TNT across nodes Synchronization across nodes TNT Distributed Shared Memory Results Conclusions and Future Work MTAAP’20104

5. Contributions Isolate traditional OS functions to a single core of the BG/P multi-core chip. Ported the TiNy Thread (TNT) execution model to allow for further utilization of BG/P compute cores. Expanded the design framework to a multi-chip system designed for scalability to a large number of chips. MTAAP’20105

6. Outline Introduction Contributions TNT on BlueGene/P Scheduling TNT across nodes Synchronization across nodes TNT Distributed Shared Memory Results Conclusions and Future Work MTAAP’20106

7. TiNy Threads on BG/P TiNy Threads –Lightweight, non-preemptive, threads –API similar to POSIX Threads. –Originally presented in “TiNy Threads: A Thread Virtual Machine for the Cyclopse-64 Cellular Architecture” –Runs on IBM Cyclops64 Kernel Modifications –Alterations to the thread scheduler to allow for non-preemption MTAAP’20107

8. Outline Introduction Contributions TNT on BlueGene/P Scheduling TNT across nodes Synchronization across nodes TNT Distributed Shared Memory Results Conclusions and Future Work MTAAP’20108

9. Multinode Thread Scheduler Thread Scheduler allows TNT to run across multiple nodes. –Requests facilitated through IBM’s Deep Computing Messaging Framework’s RPCs. Multiple Scheduling Algorithms –Workload Un-Aware Random Round-Robin –Workload Aware MTAAP’20109

10. Multinode Thread Scheduling MTAAP’201010 Node ANode B tnt_create() … tnt_join() tnt_create() … tnt_join() tid … tnt_exit() … tnt_exit() … tnt_join() … tnt_join()

11. Outline Introduction Contributions TNT on BlueGene/P Scheduling TNT across nodes Synchronization across nodes TNT Distributed Shared Memory Results Conclusions and Future Work MTAAP’201011

12. Synchronization Three forms –Mutex –Thread Joining –Barrier Similar to thread scheduling –Lock requests, Join requests, and Barrier notifications sent to node responsible for said synchronization MTAAP’201012

13. Multinode Thread Scheduling MTAAP’201013 Node ANode B tid … tnt_exit() … tnt_exit() tnt_join() A tnt_exit()

14. Outline Introduction Contributions TNT on BlueGene/P Scheduling TNT across nodes Synchronization across nodes TNT Distributed Shared Memory Results Conclusions and Future Work MTAAP’201014

15. Characteristics of TDSM Provides One-Sided access to memory distributed among nodes through IBM’s DCMF. Allows for virtual address manipulation –Maps distributed memory to a single virtual address space. –Allows for array indexing and memory offsets. Scalable to a variety of applications –Size of desired global shared memory set at runtime. Mutability –Memory allocation algorithm and memory distribution algorithm can be easily altered and/or replaced. MTAAP’201015

16. 16 Example of TDSM 0 global 153045 global[15] to global[34] 0x00040012 Node 6Node 5Node 7 Local Buffer 0x0004004E to 0x0004009A Node 6: 0 to 14 and Node 7: 0 to 4 …

17. Outline Introduction Contributions TNT on BlueGene/P Scheduling TNT across nodes Synchronization across nodes TNT Distributed Shared Memory Results Conclusions and Future Work MTAAP’201017

18. Summary of Results The performance of the TNT thread system shows comparable speedup to that of Pthreads running on the same hardware. The distributed shared memory operates at 95% of the experimental peak performance of the network, with distance between nodes not being a sensitive factor. The cost of thread creation shows a linear relationship as the number of threads increase. The cost of waiting at a barrier is constant and independent of the number of threads involved. MTAAP’201018

19. Single-Node Thread System Performance Based upon Radix-2 Cooley-Tukey algorithm with the Kiss FFT library for the underlying DFT. Underlying TNT thread model performs comparably to POSIX standard when the number of threads does not exceed the number of available processor cores. MTAAP’201019

20. Memory System Performance Reads and writes of varying sizes between one and two nodes. For inter-node communications, data can be transferred at approximately 357 MB/s. Kumar et al determined experimental peak performance on BG/P to be 374 MB/s in their ICS’08 paper. MTAAP’201020

21. Memory System Performance Size of Read/Write is a function of the number of nodes across which the data is distributed. Latency linearly increases as the amount of data increases, regardless of distance between nodes. MTAAP’201021

22. Multinode Thread Creation Cost Approximately 0.2 seconds per thread Remained effectively constant MTAAP’201022

23. Synchronization Costs Performance of barrier is effectively a constant 0.2 seconds. MTAAP’201023

24. Outline Introduction Contributions TNT on BlueGene/P Scheduling TNT across nodes Synchronization across nodes TNT Distributed Shared Memory Results Conclusions and Future Work MTAAP’201024

25. Conclusions and Future Work Proven feasibility of system Benefits of Execution Model-Driven approach Room for Improvement –Improvements to kernel –More rigorous benchmarks –Improved allocation and scheduling algorithms MTAAP’201025

26. Atlanta, Georgia Thank You MTAAP’201026

27. Bibliography J. del Cuvillo, W. Zhu, Z. Hu, and G. R. Gao, “Tiny threads: A thread virtual machine for the cyclops64 cellular architecture,” Parallel and Distributed Processing Symposium, International, vol. 15, p. 265b, 2005. S. Kumar, G. Dozsa, G. Almasi et al., “The deep computing messaging framework: generalized scalable message passing on the blue gene/p supercomputer,” in ICS ’08: Proceedings of the 22nd annual international conference on Supercomputing. New York, NY, USA: ACM, 2008, pp. 94–103. M. Borgerding, “Kiss FFT.” December 2009, http://sourceforge.net/projects/kissfft/. MTAAP’201027