## +21 ## Hazard Pointer Synchronous Reclamation Beyond Concurrency TS2 ## MAGED MICHAEL ## Basic Hazard Pointer Algorithm Hazard pointers protect access to objects that may be removed concurrently [Image](/uploads/documents/1/7/5/a/175a94ae744d634a4a2d21a0d05b187f/p2_12.jpg) Safe to delete A SAFE RECLAMATION ## Concurrency TS2 Essential Hazard Pointer Interface class hazard_pointer { template ## Asynchronous Reclamation • Asynchronous reclamation is invoked when the number of retired objects reaches some threshold: • In the Folly

ROS based application stack: |Simulator mode|ROS application mode|Determinism| |---|---|---| |Synchronous|As fast as possible|Not fully deterministic, the ROS application will be in free running mode and ROS based application stack: |Simulator mode|ROS application mode|Determinism| |---|---|---| |Synchronous|Controlled by ROS bridge using deterministic execution coordinator feature|Fully deterministic CARLA C++ ROS bridge to enable deterministic execution when the simulator is configured to run in synchronous mode: CARLA server CARLA ROS bridge with coordinator SuT (Lidar stack) while (true) { get_snapshot();

uploads, streaming large downloads, using HTTP cookies, uploading JSON data, etc... - Can send both synchronous and asynchronous requests using the same interface. - Uses PSR-7 interfaces for requests, responses ○ proxy ○ query ○ read_timeout ○ sink ○ ssl_key ○ stream ○ synchronous verify. ◦ timeout ○ version • Guzzle and PSR-7 o Headers ■ Accessing Headers ■ Complex Headers created by the client. ## Sending Requests Magic methods on the client make it easy to send synchronous requests: $response = $client->get('http://httpbin.org/get'); $response = $client-

oldptr = src_.exchange(newptr); oldptr->retire(); // Pass to hazard pointer library for safe reclamation. } ## P3135R1: Hazard Pointer Extensions (beyond C++26) P3135R1: Hazard Pointer Extensions Execution of Asynchronous Reclamation Proposed standard extensions: • Synchronous reclamation • Batch creation and destruction Hazard Pointer Execution of Asynchronous Reclamation ## Hazard Pointer Execution Execution of Asynchronous Reclamation Using C++26 Possible inline asynchronous reclamation void worker() { /* ... */ obj->retire(); // Possible inline reclamation. } Using a separate asynchronous

RCU. These extend the existing shared_ptr and the proposed atomic_shared_ptr which all have safe reclamation facilities. As such we also propose moving shared_ptr and atomic> to this new location ☐ 33.1 General Concepts ■ 33.1.1 Thread Support ■ 33.1.2 Executor Support ### • 33.2 Safe Reclamation ☐ 33.2.1 Hazard Pointers ☐ 33.2.2 RCU ○ 33.2.3 Latest/Snapshot? 33.2.4 Asymmetric fences ## in(); }; hazard_pointer_domain & hazard_pointer_default_domain() noexcept; // For synchronous reclamation void hazard_pointer_clean_up( hazard_pointer_domain & domain = hazard_pointer_default_domain())

CPU preemption and jitter suppression for online services. Additionally, its innovative memory reclamation algorithm against out of memory (OOM) allows online services to run reliably based on their higher (PUD) levels can be moved to map large memory blocks quickly. Memory reclamation against OOM: In the event of OOM, memory reclamation is preferentially performed for process groups with low priorities to performance and QoS for online services. openEuler is equipped with its own CPU scheduling and memory reclamation algorithms that are designed for hybrid deployments of cloud native services. These innovative

existing atomic are implemented (the split reference count technique) • Deferred reclamation, i.e., garbage collection in C++ ## Some assumed knowledge • You’ve heard of lock-free programming existing atomic are implemented (the split reference count technique) • Deferred reclamation, i.e., garbage collection in C++ ## Some assumed knowledge • You’ve heard of lock-free programming University ## Something new ## Solution #3 Sounds like a textbook use case of deferred reclamation ## Deferred reclamation: A quick primer ## atomic src T* protect(const std::atomic& src) - Atomically

foster further exploration by the community. Anticipatory Routing. We found that decoupling the synchronous updates of the backbone network and the routing network significantly improves training stability First, it mitigates duplicate page-cache footprints in virtualized environments and applies memory reclamation to enable safe overcommitment. Second, it alleviates spinlock contention in the container runtime

StratoVirt allocates and reclaims memory on the real-time memory demand of each workload. The memory reclamation speed can reach up to 3 GB/s through virtio-balloon. • Huge pages: On lightweight VMs, StratoVirt terms of the runtime, GC, and Just-In-Time (JIT) compiler. For example, memory allocation and reclamation in GC are optimized, and redundant memory barriers in JIT code are eliminated. In the benchmark

push(3); } // no more access to `vector`, // lock is unlocked ## crossbeam • Epoch-based memory reclamation • Easy translation of algorithms that require GC • Work stealing deque • MPMC queues ## rayon