workspace是基于C++11的轻量级异步执行框架,支持:通用任务异步执行、优先级任务调度、自适应动态线程池、高效静态线程池、异常处理机制等。
std::function
的小任务优化减少内存碎片、拥有良好的异步线程异常处理机制。workbranch(工作分支)是动态线程池的抽象,内置了一条线程安全的任务队列用于同步任务。其管理的每一条异步工作线程被称为worker,负责从任务队列不断获取任务并执行。(以下示例按顺序置于workspace/example/
)
让我们先简单地提交一点任务,当你的任务带有返回值时,workbranch会返回一个std::future,否则返回void。
#include <workspace/workspace.hpp>
int main() {
// 2 threads
wsp::workbranch br(2);
// return void
br.submit([]{ std::cout<<"hello world"<<std::endl; });
// return std::future<int>
auto result = br.submit([]{ return 2023; });
std::cout<<"Got "<<result.get()<<std::endl;
// wait for tasks done (timeout: 1000 milliseconds)
br.wait_tasks(1000);
}
由于返回一个std::future会带来一定的开销,如果你不需要返回值并且希望程序跑得更快,那么你的任务应该是void()
类型的。
当你有一个任务并且你希望它能尽快被执行时,你可以指定该任务的类型为urgent
,如下:
#include <workspace/workspace.hpp>
int main() {
// 1 threads
wsp::workbranch br;
br.submit<wsp::task::nor>([]{ std::cout<<"task B done\n";}); // normal task
br.submit<wsp::task::urg>([]{ std::cout<<"task A done\n";}); // urgent task
br.wait_tasks(); // wait for tasks done (timeout: no limit)
}
在这里我们通过指定任务类型为wsp::task::urg
,来提高任务的优先级。最终
在我的机器上:
jack@xxx:~/workspace/example/build$ ./e2
task A done
task B done
在这里我们不能保证task A
一定会被先执行,因为当我们提交task A
的时候,task B
可能已经在执行中了。urgent
标签可以让任务被插入到队列头部,但无法改变已经在执行的任务。
假如你有几个轻量异步任务,执行他们只需要非常短暂的时间。同时,按照顺序执行它们对你来说没有影响,甚至正中你下怀。那么你可以把任务类型指定为sequence
,以便提交一个任务序列。这个任务序列会被单个线程顺序执行:
#include <workspace/workspace.hpp>
int main() {
wsp::workbranch br;
// sequence tasks
br.submit<wsp::task::seq>([]{std::cout<<"task 1 done\n";},
[]{std::cout<<"task 2 done\n";},
[]{std::cout<<"task 3 done\n";},
[]{std::cout<<"task 4 done\n";});
// wait for tasks done (timeout: no limit)
br.wait_tasks();
}
任务序列会被打包成一个较大的任务,以此来减轻框架同步任务的负担,提高整体的并发性能。
当任务中抛出了一个异常,workbranch有两种处理方式:A-将其捕获并输出到终端 B-将其捕获并通过std::future传递到主线程。第二种需要你提交一个带返回值的任务。
#include <workspace/workspace.hpp>
// self-defined
class excep: public std::exception {
const char* err;
public:
excep(const char* err): err(err) {}
const char* what() const noexcept override {
return err;
}
};
int main() {
wsp::workbranch wbr;
wbr.submit([]{ throw std::logic_error("A logic error"); }); // log error
wbr.submit([]{ throw std::runtime_error("A runtime error"); }); // log error
wbr.submit([]{ throw excep("XXXX");}); // log error
auto future1 = wbr.submit([]{ throw std::bad_alloc(); return 1; }); // catch error
auto future2 = wbr.submit([]{ throw excep("YYYY"); return 2; }); // catch error
try {
future1.get();
} catch (std::exception& e) {
std::cerr<<"Caught error: "<<e.what()<<std::endl;
}
try {
future2.get();
} catch (std::exception& e) {
std::cerr<<"Caught error: "<<e.what()<<std::endl;
}
}
在我的机器上:
jack@xxx:~/workspace/test/build$ ./test_exception
workspace: worker[140509071521536] caught exception:
what(): A logic error
workspace: worker[140509071521536] caught exception:
what(): A runtime error
workspace: worker[140509071521536] caught exception:
what(): XXXX
Caught error: std::bad_alloc
Caught error: YYYY
supervisor是异步管理者线程的抽象,负责监控workbranch的负载情况并进行动态调整。它允许你在每一次调控workbranch之后执行一个小任务,你可以用来写日志或者做一些其它调控等。
每一个supervisor可以管理多个workbranch。此时workbranch之间共享supervisor的所有设定。
#include <workspace/workspace.hpp>
int main() {
wsp::workbranch br1(2);
wsp::workbranch br2(2);
// 2 <= thread number <= 4
// time interval: 1000 ms
wsp::supervisor sp(2, 4, 1000);
sp.set_tick_cb([&br1, &br2]{
auto now = std::chrono::system_clock::now();
std::time_t timestamp = std::chrono::system_clock::to_time_t(now);
std::tm local_time = *std::localtime(×tamp);
static char buffer[40];
std::strftime(buffer, sizeof(buffer), "%Y-%m-%d %H:%M:%S", &local_time);
std::cout<<"["<<buffer<<"] "<<"br1: [workers] "<<br1.num_workers()<<" | [blocking-tasks] "<<br1.num_tasks()<<'\n';
std::cout<<"["<<buffer<<"] "<<"br2: [workers] "<<br2.num_workers()<<" | [blocking-tasks] "<<br2.num_tasks()<<'\n';
});
sp.supervise(br1); // start supervising
sp.supervise(br2); // start supervising
for (int i = 0; i < 1000; ++i) {
br1.submit([]{std::this_thread::sleep_for(std::chrono::milliseconds(10));});
br2.submit([]{std::this_thread::sleep_for(std::chrono::milliseconds(20));});
}
br1.wait_tasks();
br2.wait_tasks();
}
在我的机器上,输出如下:
jack@xxx:~/workspace/example/build$ ./e4
[2023-06-13 12:24:31] br1: [workers] 4 | [blocking-tasks] 606
[2023-06-13 12:24:31] br2: [workers] 4 | [blocking-tasks] 800
[2023-06-13 12:24:32] br1: [workers] 4 | [blocking-tasks] 213
[2023-06-13 12:24:32] br2: [workers] 4 | [blocking-tasks] 600
[2023-06-13 12:24:33] br1: [workers] 4 | [blocking-tasks] 0
[2023-06-13 12:24:33] br2: [workers] 4 | [blocking-tasks] 404
[2023-06-13 12:24:34] br1: [workers] 3 | [blocking-tasks] 0
[2023-06-13 12:24:34] br2: [workers] 4 | [blocking-tasks] 204
[2023-06-13 12:24:35] br1: [workers] 2 | [blocking-tasks] 0
[2023-06-13 12:24:35] br2: [workers] 4 | [blocking-tasks] 4
[2023-06-13 12:24:35] br1: [workers] 2 | [blocking-tasks] 0
[2023-06-13 12:24:35] br2: [workers] 4 | [blocking-tasks] 0
workspace是一个托管器/任务分发器,你可以将workbranch和supervisor托管给它,并用workspace分配的组件专属ID来访问它们。将组件托管至workspace至少有以下几点好处:
我们可以通过workspace自带的任务分发机制来异步执行任务(调用submit
)。
#include <workspace/workspace.hpp>
int main() {
wsp::workspace spc;
auto bid1 = spc.attach(new wsp::workbranch);
auto bid2 = spc.attach(new wsp::workbranch);
auto sid1 = spc.attach(new wsp::supervisor(2, 4));
auto sid2 = spc.attach(new wsp::supervisor(2, 4));
spc[sid1].supervise(spc[bid1]); // start supervising
spc[sid2].supervise(spc[bid2]); // start supervising
// Automatic assignment
spc.submit([]{std::cout<<std::this_thread::get_id()<<" executed task"<<std::endl;});
spc.submit([]{std::cout<<std::this_thread::get_id()<<" executed task"<<std::endl;});
spc.for_each([](wsp::workbranch& each){each.wait_tasks();});
}
当我们需要等待任务执行完毕的时候,我们可以调用for_each
+wait_tasks
,并为每一个workbranch指定等待时间,单位是毫秒。
(更多详细接口见workspace/test/
)
wsp::futures是一个std::future收集器(collector),可以缓存同类型的std::future,并进行批量操作。一个简单的操作如下:
#include <workspace/workspace.hpp>
int main() {
wsp::futures<int> futures;
wsp::workspace spc;
spc.attach(new wsp::workbranch("br", 2));
futures.add_back(spc.submit([]{return 1;}));
futures.add_back(spc.submit([]{return 2;}));
futures.wait();
auto res = futures.get();
for (auto& each: res) {
std::cout<<"got "<<each<<std::endl;
}
}
这里futures.get()
返回的是一个std::vector<int>
,里面保存了所有任务的返回值。
测试原理:通过快速提交大量的空任务以考察框架同步任务的开销。
测试环境:Ubuntu20.04 : 16核 : AMD Ryzen 7 5800H with Radeon Graphics 3.20 GHz
<测试1>
在测试1中我们调用了submit<wsp::task::seq>
,每次打包10个空任务并提交到workbranch中执行。结果如下:(代码见workspace/benchmark/bench1.cc
)
threads: 1 | tasks: 100000000 | time-cost: 2.68801 (s)
threads: 2 | tasks: 100000000 | time-cost: 3.53964 (s)
threads: 3 | tasks: 100000000 | time-cost: 3.99903 (s)
threads: 4 | tasks: 100000000 | time-cost: 5.26045 (s)
threads: 5 | tasks: 100000000 | time-cost: 6.65157 (s)
threads: 6 | tasks: 100000000 | time-cost: 8.40907 (s)
threads: 7 | tasks: 100000000 | time-cost: 10.5967 (s)
threads: 8 | tasks: 100000000 | time-cost: 13.2523 (s)
<测试2>
在测试2中我们同样将10个任务打成一包,但是是将任务提交到workspace中,利用workspace进行任务分发,且在workspace托管的workbranch只拥有 1条 线程。结果如下:(代码见workspace/benchmark/bench2.cc
)
threads: 1 | tasks: 100000000 | time-cost: 4.38221 (s)
threads: 2 | tasks: 100000000 | time-cost: 4.01103 (s)
threads: 3 | tasks: 100000000 | time-cost: 3.6797 (s)
threads: 4 | tasks: 100000000 | time-cost: 3.39314 (s)
threads: 5 | tasks: 100000000 | time-cost: 3.03324 (s)
threads: 6 | tasks: 100000000 | time-cost: 3.16079 (s)
threads: 7 | tasks: 100000000 | time-cost: 3.04612 (s)
threads: 8 | tasks: 100000000 | time-cost: 3.11893 (s)
<测试3>
在测试3中我们同样将10个任务打成一包,并且将任务提交到workspace中,但是workspace管理的每个workbranch中都拥有 2条 线程。结果如下:(代码见workspace/benchmark/bench3.cc
)
threads: 2 | tasks: 100000000 | time-cost: 4.53911 (s)
threads: 4 | tasks: 100000000 | time-cost: 7.0178 (s)
threads: 6 | tasks: 100000000 | time-cost: 6.00101 (s)
threads: 8 | tasks: 100000000 | time-cost: 5.97501 (s)
threads: 10 | tasks: 100000000 | time-cost: 5.63834 (s)
threads: 12 | tasks: 100000000 | time-cost: 5.17316 (s)
总结:利用workspace进行任务分发,且workbranch线程数为1的情况下,整个任务同步框架是静态的,任务同步开销最小。当workbranch内的线程数越多,面对大量空任务时对任务队列的竞争越激烈,框架开销越大。(更加详尽的测试结果见bench.md
,测试代码于workspace/bench
)
请提前安装doxygen
# 与workspace同级目录中(Linux)
doxygen ./doxygen.conf
生成的文档在workspace/docs/
中,可以在浏览器中打开workspace/docs/html/index.html
并查看接口。
# 项目代码与workspace同级(Linux)
g++ -I workspace/include xxx.cc -lpthread && ./a.out
其它平台可能需要链接不同的线程库,且可执行文件后缀不同。
# 在"workspace/example"中
cmake -B build
cd build
make
./e1
# 在"workspace/"中
cmake -B build
cd build
sudo make install
submit([&br]{br.wait_tasks();});
会阻塞线程 set_tick_cb([&sp]{sp.suspend();});
组件接口 | 是否线程安全 |
---|---|
workspace | 否 |
workbranch | 是 |
supervisor | 是 |
futures | 否 |
workspace有关时间的接口单位都是 -> 毫秒(ms)
《C++并发编程》
邮箱: 1848395727@qq.com