all those things are structs, dctors are available for customization. 4344 44 JOB SYSTEM • Scheduler + User defined workload • Optimize for CPU throughput Okay, finally, we are ready to talk about which is spared time frames on the CPU timeline. 4445 COROUTINE JOB SYSTEM isa::await_suspend Scheduler Worker Worker Worker Worker Job Queue Job Queue Job Queue Job coroutine_handle Job::Job coroutine_handle Job::Job final_suspend And voila! This is our system. 4546 isa::await_suspend Scheduler Worker Worker Worker Worker Job Queue Job Queue Job Queue Job coroutine_handle Job::Job

() { IOAwaitable awaitable = async_io(scheduler, ...); auto data = co_await awaitable; co_return IoResult :: from_io_data(data ); }42/55 Where does the scheduler come from? Async inner_function async_io(scheduler , ...); /* ... */ } int main () { Scheduler scheduler; spawn_task(scheduler); }42/55 Where does the scheduler come from? Async inner_function (Scheduler& scheduler) { auto auto data = co_await async_io(scheduler , ...); /* ... */ } int main () { Scheduler scheduler; spawn_task(scheduler ); }43/55 Passing data from the outside in template struct Async { /*

What is Task Graph-based Programming System (TGPS) Code A B C D B A C D Scheduler 7Why TGPS 1 Scheduler sched; 2 task_a = sched.emplace([](&){ 3 // Code block A; 4 }); 5 task_b = sched decomposition strategies 8Existing TGPSs on Heterogenous Computing - Challenge A C D B! B" 1 Scheduler sched; 2 task_a = sched.emplace([](&){ 3 // Code block A; 4 }); 5 task_b = sched.emplace([](&){ suspension point behavior co_await std::suspend_always() Symmetric coroutine transfer Implement a scheduler 22A new heterogeneous TGPS using C++ Coroutine? Abstract away C++ Coroutine detailsA new heterogeneous

STD::EXECUTION Proposes: • A set of concepts that represent: • A handle to a compute resource (aka, scheduler) • A unit of lazy async work (aka, sender) • A completion handler (aka, receiver) • A small, compute_intensive(int); int main() { unifex::static_thread_pool pool{8}; ex::scheduler auto sched = pool.get_scheduler(); ex::sender auto work = ex::when_all( ex::then(ex::schedule(sched), compute_intensive(int); int main() { unifex::static_thread_pool pool{8}; ex::scheduler auto sched = pool.get_scheduler(); ex::sender auto work = ex::when_all( ex::then(ex::schedule(sched),

return process_file(std::move(file)); }); auto stats = co_await unifex::on( pool.get_scheduler(), unifex::when_all_range(jobs.begin(), jobs.end())); for (std::size_t i = 0; i < stats.size(); return process_file(std::move(file)); }); auto stats = co_await unifex::on( pool.get_scheduler(), unifex::when_all_range(jobs.begin(), jobs.end())); for (std::size_t i = 0; i < stats.size(); return process_file(std::move(file)); }); auto stats = co_await unifex::on( pool.get_scheduler(), unifex::when_all_range(jobs.begin(), jobs.end())); for (std::size_t i = 0; i < stats.size();

offer parallelism, but lack control over execution hardware ("where"). ● P2300 introduces the "scheduler" concept, representing execution contexts, addressing the "where." ● The integration of these proposes extending C++ algorithms to accept a "policy-aware scheduler." ● This scheduler combines an execution policy ("how") and a scheduler ("where"). ● The execute_on function facilitates the creation algorithms for optimal performance. ● Key Features: ● Combining scheduler and policy. ● Minimal, incremental API changes.Combining Scheduler with Policy Why Use Schedulers? ● Schedulers represent execution

TidbInitializer - TidbMonitor - Backup/Restore - Custom Controller - Scheduler Extender (optional) - Admission Webhook (optional) - …Scheduler framework, Aggregated APIServer PingCAP.com Custom Resource Kubernetes API PingCAP.com Scheduler Extender kube-scheduler kube-scheduler tidb-scheduler other Pod other Pod other Pod PD Pod TiKV Pod TiDB Pod PingCAP.com Scheduler Extender 3 TiKV ❌ Node-A Node-B Operator Pattern: Extending Kubernetes - 扩展 Kubernetes 的几种方式 - Custom Resource - Custom Controller - Scheduler Extender - Admission Webhook - TiDB Operator 实现 - the tidb-controller-manager - Full lifecycle management

statement count : 单位时间内不同类型语句执行的数目 Query Duration 99th percentile : 99% 的 query 时间 TiKV 99% & 99.99% scheduler command duration : 99% & 99.99% 命令执行的时间 99% 小于 50ms；99.99% 小于 100ms 95% & 99% storage async_request data，如果有了监控值，说明对应 TiKV 节点的消息处理不过来 了 server report failures：正常情况显示 No data，如果出现了 Unreachable，说明 TiKV 之 间通信有问题 scheduler pending commands：写入堆积的数量，偶尔出现峰值属于正常现象 coprocessor pending requests：正常情况监控为 0 或者数量很少 coprocessor 监控 TiDB Server PD Server TiKV Server 组件状态 API 监控 - 261 - 本文档使用 书栈(BookStack.CN) 构建 执行 KV 命令的总次数 Scheduler 执行命令的耗时统计 Raft propose 命令的总次数 Raft 执行命令的耗时统计 Raft 执行命令失败的总次数 Raft 处理 ready 状态的总次数 整个架构如下图所示，在 TiDB/PD/TiKV

yaml文件能搞定的 平台侧自身复杂度变高， 监控难度加大 从 Kubernetes 架构来 看要监控的组件 Kubernetes架构 l 服务端组件，控制面：API Server、Scheduler、 Controller-Manager、ETCD l 工作负载节点，最核心就是监控Pod容器和节点本 身，也要关注 kubelet 和 kube-proxy l 业务程序，即部署在容器中的业务程序的监控，这 process_cpu_seconds_total 进程使用的CPU时间的总量，rate 之后就是 CPU 使用率 Kubernetes控制面 scheduler的监控 • scheduler 通过 /metrics 接口暴露监控数据，直接拉取即 可 • scheduler 在 Kubernetes 架构中，是负责调度对象到合 适的node上，会有一系列的规则计算和筛选。重点关注调 度这个动作的相关指标 • categraf 仓库的 k8s/deployment.yaml，大盘可以参考 k8s/scheduler- dash.json • rest_client_request_duration_seconds 请求 apiserver 的耗时分布，histogram类型，按照 url + verb 统计 • scheduler_framework_extension_point_duration_s econds

Each object owns its own thread. ▪ Each method invocation is stored in an activation list. ▪ A scheduler triggers the method execution.Active Object Proxy: ▪ Proxy for the member functions on the active ▪ Has the pending requests objects. ▪ Decouples the client from the Active Object thread. Scheduler: ▪ Runs in the thread of the Active Object. ▪ Decides with request from the Activation List is to the scheduler. ▪ The scheduler enqueues the request on the activation list and returns a future to the client if the request returns something. 2. Member function execution ▪ The scheduler determines