Concurrency and Scheduling Analysis of Real-time Embedded Software - PowerPoint PPT Presentation

Concurrency and Scheduling Analysis of Real-time Embedded Software on Multi-core Processors Yann-Hang Lee Young Song Duo Lu CONFIDENTIAL Center for Embedded Systems | An NSF Industry/University Cooperative Research Center

Project Overview and Description thread thread thread thread thread thread thread thread thread thread thread thread thread thread RTOS SMP-ready RTOS Single-core processor Multi-core processor � Will the application run correctly � a benign data race may become a true race � scheduling anomaly � Will the application run efficiently � cost of synchronization, context switch, and shared access � scheduling to meet deadlines 2 CONFIDENTIAL Center for Embedded Systems | An NSF Industry/University Cooperative Research Center

Two Legs to Leap the Sync Barrier When we forgot to lock: When we remember to lock: � Optimization – partitioned � Race detection – vector clock scheduling, stack-based resource based and dynamic granularity policy, and adaptive selection of for C/C++ programs locking primitives � Detection of true races -- based on task model and scheduling Correctness Performance Real-time Multicore Application Concurrency & Synchronization 3 CONFIDENTIAL Center for Embedded Systems | An NSF Industry/University Cooperative Research Center

Race Detector with Dynamic Granularity � Vector clock based data race detector for C/C++ programs � On top of FastTrack and using Intel PIN for dynamic instrumentation � Share vector clock dynamically with neighboring locations Slowdown Memory Overhead Data race detected Base Base Inspector XE Inspector XE Inspector XE Dynamic Dynamic Dynamic granularity granularity granularity Benchmark Valgrind Valgrind Valgrind time Mem. Intel Intel Intel DRD DRD DRD program (sec) (MB) facesim 6.1 288 59 128 102 2.2 6.0 4.6 8909 31 8909 ferret 6.7 146 748 87 52 2.6 5.0 8.9 108 4 2 fluidanimate 2.0 248 -- 89 81 -- 12.4 2.2 -- 7 1 raytrace 9.5 170 42 17 27 1.9 4.1 2.0 16 0 13 dedup 7.7 2682 -- -- 85 -- -- 1.0 -- -- 0 streamcluster 3.8 30 66 108 137 4.2 17.5 3.7 1067 61 1079 ffmpeg 3.0 95 120 -- 109 2.6 -- 3.1 0 -- 1 pbzip2 5.7 67 64 99 39 2.9 8.6 3.4 0 0 0 hmmsearch 26.6 23 74 64 45 4.4 21.9 4.3 1 2 1 168 85 75 3 11 4 Average 4 CONFIDENTIAL Center for Embedded Systems | An NSF Industry/University Cooperative Research Center

Project Tasks/ Deliverables Description Date Status Improve the existing prototype of race detector in 1 st Q 1 multicore environment Analyze race conditions based on scheduling 2 nd Q 2 algorithms and task models Investigate scheduling approaches to reduce the 1 st Q 3 degree of preemption and blocking Develop and test adaptive synchronization 2 nd Q 4 mechanisms 3 rd Q Benchmark evaluation and schedulability analysis 5 Deliverables � A scalable race detector for embedded C/C++ programs. � Effective classification of race conditions � Integrated design of partitioned scheduling and resource access control to minimize synchronization overhead. 5 CONFIDENTIAL Center for Embedded Systems | An NSF Industry/University Cooperative Research Center