Boost無鎖佇列
在開發接收轉發agent時,採用了多執行緒的生產者-消費者模式,用了加互斥鎖的方式來實現執行緒同步。互斥鎖會阻塞執行緒,所以壓測時,效率並不高。所以想起用無鎖佇列來實現,效能確實提升了。
首先介紹下lock-free和wait-free的區別:
阻塞演算法可能會出現整個系統都掛起的情況(佔有鎖的執行緒被中斷,無法釋放所,那麼所有試圖爭用這個鎖的執行緒會被掛起),系統中的所有執行緒全部餓死。
無鎖演算法可以保證系統中至少有一個執行緒處於工作狀態,但是還是可能有執行緒永遠搶不到資源而被餓死。
無等待演算法保證系統中的所有執行緒都能處於工作狀態,沒有執行緒會被餓死,只要時間夠,所有執行緒都能結束。相比於無鎖演算法,無等待演算法有更強的保證。
一. 用互斥鎖實現單生產者-單消費者
#include <string>
#include <sstream>
#include <list>
#include <pthread.h>
#include <iostream>
#include <time.h>
using namespace std;
int producer_count = 0;
int consumer_count = 0;
list<string> product;
list<string> consumer_list;
pthread_mutex_t mutex;
const int iterations = 10000;
//是否生產完畢標誌
bool done = false;
void* producer(void* args)
{
for (int i = 0; i != iterations; ++i) {
pthread_mutex_lock(&mutex);
int value = ++producer_count;
stringstream ss;
ss<<value;
product.push_back(ss.str());
//cout<<"list push:"<<ss.str()<<endl;
pthread_mutex_unlock(&mutex);
}
return 0;
}
//消費函式
void* consumer(void* args)
{
//當沒有生產完畢,則邊消費邊生產
while (!done) {
pthread_mutex_lock(&mutex);
if(!product.empty()){
consumer_list.splice(consumer_list.end(), product);
pthread_mutex_unlock(&mutex);
while(!consumer_list.empty()){
string value = consumer_list.front();
consumer_list.pop_front();
//cout<<"list pop:"<<value<<endl;
++consumer_count;
}
}else{
pthread_mutex_unlock(&mutex);
}
}
//如果生產完畢,則消費
while(!consumer_list.empty()){
string value = consumer_list.front();
consumer_list.pop_front();
//cout<<"list pop:"<<value<<endl;
++consumer_count;
}
return 0;
}
int main(int argc, char* argv[])
{
struct timespec time_start={0, 0},time_end={0, 0};
clock_gettime(CLOCK_REALTIME, &time_start);
pthread_t producer_tid;
pthread_t consumer_tid;
pthread_mutex_init (&mutex,NULL);
pthread_create(&producer_tid, NULL, producer, NULL);
pthread_create(&consumer_tid, NULL, consumer, NULL);
//等待生產者生產完畢
pthread_join(producer_tid, NULL);
//可以消費標誌
done = true; //主執行緒不等生產執行緒完畢就設定done標記
cout << "producer done" << endl; //輸出以觀察主執行緒和各子執行緒的執行順序
//等待消費者結束
pthread_join(consumer_tid, NULL);
clock_gettime(CLOCK_REALTIME, &time_end);
long cost = (time_end.tv_sec-time_start.tv_sec)/1000000 + (time_end.tv_nsec-time_start.tv_nsec)/1000;
cout<<"===========cost time:"<<cost<<"us==========="<<endl;
cout << "produced " << producer_count << " objects." << endl;
cout << "consumed " << consumer_count << " objects." << endl;
}
生產消費10000個string型別的資料,耗時:58185us
二. Boost庫的無鎖佇列
boost.lockfree實現了三種無鎖資料結構:
boost::lockfree::queue
alock-free multi-produced/multi-consumer queue
一個無鎖的多生產者/多消費者佇列,注意,這個queue不支援string型別,支援的資料型別要求:
- T must have a copy constructor
- T must have a trivial assignment operator
- T must have a trivial destructor
boost::lockfree::stack
alock-free multi-produced/multi-consumer stack
一個無鎖的多生產者/多消費者棧,支援的資料型別要求:
- T must have a copy constructor
boost::lockfree::spsc_queue
await-free single-producer/single-consumer queue (commonly known as ringbuffer)
一個無等待的單生產者/單消費者佇列(通常被稱為環形緩衝區),支援的資料型別要求:
- T must have a default constructor
- T must be copyable
三. Queue示例
這裡實現的還是單生產者-單消費者。
#include <pthread.h>
#include <boost/lockfree/queue.hpp>
#include <iostream>
#include <time.h>
#include <boost/atomic.hpp>
using namespace std;
//生產數量
boost::atomic_int producer_count(0);
//消費數量
boost::atomic_int consumer_count(0);
//佇列
boost::lockfree::queue<int> queue(512);
//迭代次數
const int iterations = 10000;
//生產函式
void* producer(void* args)
{
for (int i = 0; i != iterations; ++i) {
int value = ++producer_count;
//原子計數————多執行緒不存在計數不上的情況
//若沒有進入佇列,則重複推送
while(!queue.push(value));
//cout<<"queue push:"<<value<<endl;
}
return 0;
}
//是否生產完畢標誌
boost::atomic<bool> done (false);
//消費函式
void* consumer(void* args)
{
int value;
//當沒有生產完畢,則邊消費邊生產
while (!done) {
//只要能彈出元素,就消費
while (queue.pop(value)) {
//cout<<"queue pop:"<<value<<endl;
++consumer_count;
}
}
//如果生產完畢,則消費
while (queue.pop(value)){
//cout<<"queue pop:"<<value<<endl;
++consumer_count;
}
return 0;
}
int main(int argc, char* argv[])
{
cout << "boost::lockfree::queue is ";
if (!queue.is_lock_free())
cout << "not ";
cout << "lockfree" << endl;
struct timespec time_start={0, 0},time_end={0, 0};
clock_gettime(CLOCK_REALTIME, &time_start);
pthread_t producer_tid;
pthread_t consumer_tid;
pthread_create(&producer_tid, NULL, producer, NULL);
pthread_create(&consumer_tid, NULL, consumer, NULL);
//等待生產者生產完畢
pthread_join(producer_tid, NULL);
//可以消費標誌
done = true; //主執行緒不等生產執行緒完畢就設定done標記
cout << "producer done" << endl; //輸出以觀察主執行緒和各子執行緒的執行順序
//等待消費者結束
pthread_join(consumer_tid, NULL);
clock_gettime(CLOCK_REALTIME, &time_end);
long cost = (time_end.tv_sec-time_start.tv_sec)/1000000 + (time_end.tv_nsec-time_start.tv_nsec)/1000;
cout<<"===========cost time:"<<cost<<"us==========="<<endl;
//輸出生產和消費數量
cout << "produced " << producer_count << " objects." << endl;
cout << "consumed " << consumer_count << " objects." << endl;
return 0;
}
生產消費10000個int型別的資料,耗時:3963us
stack與queue類似,只不過是先進後出。
四. Waitfree Single-Producer/Single-Consumer Queue無等待單生產者/單消費者佇列
#include <pthread.h>
#include <iostream>
#include <time.h>
#include <boost/lockfree/spsc_queue.hpp>
#include <boost/atomic.hpp>
using namespace std;
int producer_count = 0;
boost::atomic_int consumer_count (0);
boost::lockfree::spsc_queue<int, boost::lockfree::capacity<1024> > spsc_queue;
const int iterations = 10000;
void* producer(void* args)
{
for (int i = 0; i != iterations; ++i) {
int value = ++producer_count;
while(!spsc_queue.push(value));
//cout<<"queue push:"<<value<<endl;
}
return 0;
}
//是否生產完畢標誌
boost::atomic<bool> done (false);
//消費函式
void* consumer(void* args)
{
int value;
//當沒有生產完畢,則邊消費邊生產
while (!done) {
//只要能彈出元素,就消費
while (spsc_queue.pop(value)) {
//cout<<"queue pop:"<<value<<endl;
++consumer_count;
}
}
//如果生產完畢,則消費
while (spsc_queue.pop(value)){
//cout<<"queue pop:"<<value<<endl;
++consumer_count;
}
return 0;
}
int main(int argc, char* argv[])
{
using namespace std;
cout << "boost::lockfree::queue is ";
if (!spsc_queue.is_lock_free())
cout << "not ";
cout << "lockfree" << endl;
struct timespec time_start={0, 0},time_end={0, 0};
clock_gettime(CLOCK_REALTIME, &time_start);
pthread_t producer_tid;
pthread_t consumer_tid;
pthread_create(&producer_tid, NULL, producer, NULL);
pthread_create(&consumer_tid, NULL, consumer, NULL);
//等待生產者生產完畢
pthread_join(producer_tid, NULL);
//可以消費標誌
done = true; //主執行緒不等生產執行緒完畢就設定done標記
cout << "producer done" << endl; //輸出以觀察主執行緒和各子執行緒的執行順序
//等待消費者結束
pthread_join(consumer_tid, NULL);
clock_gettime(CLOCK_REALTIME, &time_end);
long cost = (time_end.tv_sec-time_start.tv_sec)/1000000 + (time_end.tv_nsec-time_start.tv_nsec)/1000;
cout<<"===========cost time:"<<cost<<"us==========="<<endl;
cout << "produced " << producer_count << " objects." << endl;
cout << "consumed " << consumer_count << " objects." << endl;
}
生產消費10000個int型別的資料,耗時:1832us
如果把int改為string型別,耗時:28788us
五.效能對比
從上面可以看出在單生產者-單消費者模式下,spsc_queue比queue效能好,無鎖佇列比互斥鎖的方式效能也要好。