shared_ptr s1: control_block: atomic ref_count = 1; … control_block* ctrl; T12 Daniel Anderson -- danielanderson.net std::shared_ptr (the basics) T* ptr; shared_ptr s1: control_block: control_block* shared_ptr s2: control_block* ctrl; atomic ref_count = 2; …13 Daniel Anderson -- danielanderson.net std::shared_ptr (the basics) control_block: T T* ptr; control_block* ctrl; atomic … shared_ptr s2:14 Daniel Anderson -- danielanderson.net std::shared_ptr (the basics) control_block: T atomic ref_count = 0; …15 Daniel Anderson -- danielanderson.net std::atomic16

• Introduction • Memory • Host vs Device Functions • Return on Investment • Concluding remarks Outline 2 |• I work the RiskLab team at CSIRO on applied mathematics for Financial Risk. • The aim of … add_gpu<<>>(N, x, y); // … } Ok, about the kernel parameters 10 |Memory“In a typical PC or cluster node today, the memories of the CPU and GPU are physically distinct and https://developer.nvidia.com/blog/unified-memory-in-cuda- 6/ CPU vs GPU Memory System Memory GPU Memory 12 |“Unified Memory creates a pool of managed memory that is shared between the CPU and GPU,

Windows, sizeof(wchar_t) == 2 • Non-Windows, sizeof(wchar_t) == 4 std::basic_string has memory implications. More on that later.What assumptions are being made? void get_size(size_t dev, long* returns the struct in registers, but the get_data_from() member function returns in caller provided memory. This is often unexpected but occurs using the MSVC compiler for x86 with stdcall (callee cleanup) cleanup) or cdecl (caller cleanup). For non-MSVC, data_t is always returned in a caller provided memory.What else isn’t being declared? struct data_t { int a; int b; }; /* Get data from device ‘dev’

Docker CP vulnerability Pod Isolation Challenges • Noisy neighbor −Impact performance on CPU, Memory, Bandwidth, Buffer IO, PIDs, File descriptors # kubectl run --rm -it bb --image=busybox sh / # with SELinux, AppArmor, Seccomp, cgroup VM-based Sandbox Kata-container BareMetal Only Heavy control logic Application kernel based Sandbox gVisor Compatibility problem, Bottleneck in sentry mVMd abstracts the common virtualization components which implements a Rust-based VMM. • Written in Rust: Memory-safe language • Secure: Minimal hardware emulation • Flexible: Easy to customize to fit various network

virtual functions  constexpr functions can now:  use dynamic_cast() and typeid  do dynamic memory allocations, new / delete  contain try/catch blocks  But still cannot throw exceptions33 constexpr September 15 • 13:30Concurrency Changes35 Atomic Smart Pointers  Is shared_ptr thread safe?  Yes: control block manipulation thread safe  guarantees object is deleted exactly once  No: accessing pointer attribute can now include a reason  Example: [[nodiscard("Ignoring the return value will result in memory leaks.")]] void* GetData() { /* ... */ }80 Bit Operations  Header  Set of global

AppInspector iOS Chrome DevTools VSCode Extension Debugger Console Resources Network Elements Memory Profiling Timeline/CPU Prof. Anything Else? Two easy ways to use NativeScript 1) 2) NativeScript iOS/Android (requires Mac for iOS) o Integrates with Visual Studio Code (via plugin) WHY: More control, Free, Integrate with existing IDEs/code editors WHO: More technical developers used to using

microbenchmarking. • 2020: One user: High frequency of retiring objects. Unsharded list grew out of control. • 2020: Sharded cohorts without reclamation under lock. Fast. Scalable. Robust. Quasi-Synchronous Working Draft, Extensions to C++ for Concurrency Version 2 (wg21.link/n4895).  Hazard Pointers: Safe Memory Reclamation for Lock-Free Objects, IEEE Transactions on Parallel and Distributed Systems. 15 (8):

implement it • We read post-2018 leap seconds from a Windows registry • SYSTEM\CurrentControlSet\Control\LeapSecondInformation • For pre-2018 leap seconds, we maintain a static constexpr table to pull (Google) • Static Analysis and Program Safety in C+ +: Making it Real – Sunny Chatterjee • In-memory and Persistent Representations of C++ – Gabriel Dos Reis (online 27th) • Extending and Simplifying

会自动进行同步操作 ，即和 cudaDeviceSynchronize() 等价！ 因此前面的 cudaDeviceSynchronize() 实 际上可以删掉了。 统一内存地址技术（ Unified Memory ） • 还有一种在比较新的显卡上支持的特性， 那就是统一内存 (managed) ，只需把 cudaMalloc 换成 cudaMallocManaged 即可，释放时也是通过 cudaFree ）组成的。每个 SM 可以处理一个或多个板块。 • SM 又由多个流式单处理器（ SP ）组成。每个 SP 可以处理一个或多个线程。 • 每个 SM 都有自己的一块共享内存（ shared memory ），他的性质类似于 CPU 中的缓 存——和主存相比很小，但是很快，用于缓冲临时数据。还有点特殊的性质，我们稍后会 讲。 • 通常板块数量总是大于 SM 的数量，这时英伟达驱动就会在多个 SM 循环内部都是没有数据依赖，从而是可以 并行的（对 CPU 而言是 SIMD 和指令级并行，虽 然 GPU 没有，但为了引出共享内存的概念我才这样 改）。 板块的共享内存（ shared memory ） • 刚刚已经实现了无数据依赖可以并行的 for ，那么如何把 他真正变成并行的呢？这就是板块的作用了，我们可以把 刚刚的线程升级为板块，刚刚的 for 升级为线程，然后把 刚刚 local_sum