mine, all mine! You can’t have them. Agenda 1. Current status of C++ safety: MISRA and C++ CG 2. Parallel Safety rules 3. Automotive Safety case Safety Critical API Evolution minimize API surface area no harm What is still missing? So far most only deal with Sequential code Very few deal with Parallel code Even fewer deal with Concurrent, event driven code None deal with Heterogeneous dispatch rules pulled from • C++CG • HIC++ • REphrase H2020 project • CERT C++ • JSF++ (no parallel rules) • WG23 (no parallel rules) • Added some from our own contributions • Many joined, average 5-8 per meeting

programs using parallel algorithms and executors Mark Hoemmen* mhoemmen@stellarscience.com CppCon 2020 * hoʊ’mən, or hœm’mən; he/himWho am I? • > 10 years post-PhD experience writing parallel C++ for for science and engineering • Background: Parallel algorithms for big linear algebra problems • 1st WG21: Nov 2017 • Started new job at Stellar Science in March Eschew raw pointersOutline • Parallelism Parallelism makes error handling harder… • …C++ parallel algorithms and tasks specifically • Message Passing Interface (MPI): 3 decades of distributed-memory parallel programming • MPI teaches design patterns

std::simd how to express inherent parallelism efficiently via data-parallel types Dr. Matthias Kretz GSI Helmholtz Center for Heavy Ion Research CppCon ’23 @mkretz@floss.social github.com/mattkretzMotivation g++ -O3 -DNDEBUG -march=native scalar 0.25 FLOP/cycle 0.5 FLOP/cycle data-parallel 8 FLOP/cycle 64 FLOP/cycle A data-parallel type wider than the default can increase ILP! 64 0.25 = 256 times faster g++ -O3 -DNDEBUG -march=native scalar 0.25 FLOP/cycle 0.5 FLOP/cycle data-parallel 8 FLOP/cycle 64 FLOP/cycle A data-parallel type wider than the default can increase ILP! 64 0.25 = 256 times faster

pointer extensions (P3135; Maged) 3. Pointer tagging (P3125; Maged) 4. Parallel Range algorithms (P3179; Michael), may be Parallel Algorithms (P2500)C++26: Atomic Min/MaxC++26: Improving C++20 Concurrency applications, including reductions in data-parallel algorithms, statistics collection, and optimization processes. Useful for: ● Lock-free data structures ● Parallel reductions (OpenMP) ● Optimization algorithms Concurrency++ RCU, HP [MichaelW, Gonzalo, Maged, Paul M] SIMD [Matthias, Daniel, Ruslan] C++ 2017 Parallel Algorithms [many] Concurrency++ Memory Model++ [MichaelW, Maged, Paul M] Forward Progress C++

licensed under a Creative Commons Attribution 4.0 International License SYCL Single Source C++ Parallel Programming GPU FPGA DSP Custom Hardware GPU CPU CPU CPU Standard C++ Application Code Compiler CPU Compiler CPU One-MKL One-DNN OneDPC SYCL-BLAS SYCL-Eigen SYCL-DNN SYCL Parallel STL ... C++ templates and lambda functions separate host & accelerated device code Accelerated CPU CPU AI/Tensor HW Other BackendsSYCL 2020 is here! Open Standard for Single Source C++ Parallel Heterogeneous Programming SYCL 2020 is released after 3 years of intense work Significant adoption

Basics: Concurrency CONCURRENCY MADE EASY C++ parallel algorithms Let the compiler do all the hard work for you Bryce Adelstein Lelbach, “The C++17 Parallel Algorithms Library and Beyond,” CppCon 2016 it’s parallelism, not concurrencyCppCon 2023 23 David Olsen - Back to Basics: Concurrency C++ PARALLEL ALGORITHMS Add execution policy as first argument to algorithm call std::execution::seq Run sequentially std::execution::par, std::execution::par_unseq Request to compiler to run in parallel Promise by user that code is safe to run in parallel; no data races std::execution::unseq, std::execution::par_unseq Request

Transactions 174 Managing XIDs and the Frozen XID 175 Example of XID Modulo Calculations 176 About Parallel Data Loading 177 About Redundancy and Failover in Greenplum Database 178 About Segment Mirroring External Tables with XML Transformations 446 Accessing External Data with PXF 446 Using the Greenplum Parallel File Server (gpfdist) 447 About gpfdist and External Tables 447 About gpfdist Setup and Performance Understanding Query Planning and Dispatch 566 Understanding Greenplum Query Plans 568 Understanding Parallel Query Execution 569 About GPORCA 570 Overview of GPORCA 571 Activating and Deactivating GPORCA

Transactions 323 Managing XIDs and the Frozen XID 324 Example of XID Modulo Calculations 325 About Parallel Data Loading 326 About Redundancy and Failover in Greenplum Database 327 About Segment Mirroring Foreign Data Wrapper Extension with PGXS 649 Deployment Considerations 651 Using the Greenplum Parallel File Server (gpfdist) 651 About gpfdist and External Tables 651 About gpfdist Setup and Performance Understanding Query Planning and Dispatch 772 Understanding Greenplum Query Plans 773 Understanding Parallel Query Execution 774 About GPORCA 775 Overview of GPORCA 776 Activating and Deactivating GPORCA

linewise Default: False 强制以行为缓冲区单位，以替换字符／比特，通常用在并行模式下。可以使用 --linewise 参数来激活。env.parallel 模式隐含了这项设置——即使 linewise 为 False，parallel 如果为 True 就会引发行级输出。 参见 行级输出 vs 比特级输出 1.3 新版功能. local_user 一个包含本地系统用户名的只读值。该值即 目录加载密钥。（当然，你可以显 示地使用 fab -i 来指定密钥。） 0.9.1 新版功能. 参见 -k parallel Default: False 如为 True，强制所有任务并行执行。隐式指定 env.linewise。 1.3 新版功能. 参见 --parallel、并行执行 password Default: None SSH 层连接原程主机 以及／或者 sudo 时使用的默认密码。 Execution strategy Fabric defaults to a single, serial execution method, though there is an alternative parallel mode available as of Fabric 1.3 (see 并行执行). This default behavior is as follows: A list of tasks

Transactions 309 Managing XIDs and the Frozen XID 310 Example of XID Modulo Calculations 311 About Parallel Data Loading 312 About Redundancy and Failover in Greenplum Database 313 About Segment Mirroring Foreign Data Wrapper Extension with PGXS 620 Deployment Considerations 622 Using the Greenplum Parallel File Server (gpfdist) 622 About gpfdist and External Tables 622 About gpfdist Setup and Performance Understanding Query Planning and Dispatch 742 Understanding Greenplum Query Plans 744 Understanding Parallel Query Execution 745 About GPORCA 746 Overview of GPORCA 747 Activating and Deactivating GPORCA