QFuture 内存泄漏

QFuture Memoryleak

我想并行化一个函数,但遇到几个小时后我的内存超载的问题。

测试程序计算了一些简单的东西,目前运行正常。只是内存占用在不断增加。

QT 项目文件:

QT -= gui
QT += concurrent widgets
CONFIG += c++11 console
CONFIG -= app_bundle
DEFINES += QT_DEPRECATED_WARNINGS
SOURCES += main.cpp

QT程序文件:

#include <QCoreApplication>
#include <qdebug.h>
#include <qtconcurrentrun.h>

double parallel_function(int instance){
    return (double)(instance)*10.0;
}

int main(int argc, char *argv[])
{
    QCoreApplication a(argc, argv);
    int nr_of_threads = 8;
    double result_sum,temp_var;
    for(qint32 i = 0; i<100000000; i++){
      QFuture<double> * future = new QFuture<double>[nr_of_threads];

      for(int thread = 0; thread < nr_of_threads; thread++){
        future[thread] = QtConcurrent::run(parallel_function,thread);
      }
      for(int thread = 0; thread < nr_of_threads; thread++){
        future[thread].waitForFinished();
        temp_var = future[thread].result();
        qDebug()<<"result: " << temp_var;
        result_sum += temp_var;
      }
  }

  qDebug()<<"total: "<<result_sum;
  return a.exec();
}

据我观察,QtConcurrent::run(parallel_function,thread)分配内存,但在future[thread].waitForFinished()之后不释放内存。

这是怎么回事?

你有内存泄漏,因为 future 数组没有被删除。在外部 for 循环的末尾添加 delete[] future

for(qint32 i = 0; i<100000000; i++)
{
   QFuture<double> * future = new QFuture<double>[nr_of_threads];

   for(int thread = 0; thread < nr_of_threads; thread++){
     future[thread] = QtConcurrent::run(parallel_function,thread);
   }
   for(int thread = 0; thread < nr_of_threads; thread++){
      future[thread].waitForFinished();
      temp_var = future[thread].result();
      qDebug()<<"result: " << temp_var;
      result_sum += temp_var;
   }

   delete[] future; // <--
}

这看起来可能是这样的 - 请注意一切都可以简单得多!你死心塌地地进行手动内存管理:为什么?首先,QFuture 是一个值。您可以非常有效地将它存储在任何将为您管理内存的向量容器中。您可以使用 range-for 迭代这样的容器。等等

QT = concurrent   # dependencies are automatic, you don't use widgets
CONFIG += c++14 console
CONFIG -= app_bundle
SOURCES = main.cpp

尽管该示例是合成的并且 map_function 非常简单,但值得考虑如何最有效和最有表现力地做事。您的算法是典型的 map-reduce 操作,blockingMappedReduce 的开销只有手动完成所有工作的一半。

首先,让我们用 C++ 重铸原来的问题,而不是一些 C 语言的科学怪人。

// https://github.com/KubaO/Whosebugn/tree/master/questions/future-ranges-49107082
/* QtConcurrent will include QtCore as well */
#include <QtConcurrent>
#include <algorithm>
#include <iterator>

using result_type = double;

static result_type map_function(int instance){
   return instance * result_type(10);
}

static void sum_modifier(result_type &result, result_type value) {
   result += value;
}

static result_type sum_function(result_type result, result_type value) {
   return result + value;
}

result_type sum_approach1(int const N) {
   QVector<QFuture<result_type>> futures(N);
   int id = 0;
   for (auto &future : futures)
      future = QtConcurrent::run(map_function, id++);
   return std::accumulate(futures.cbegin(), futures.cend(), result_type{}, sum_function);
}

没有手动内存管理,也没有显式拆分为 "threads" - 这是毫无意义的,因为并发执行平台知道有多少线程。所以这已经更好了!

但这看起来很浪费:每个 future 内部至少分配一次 (!)。

我们可以使用 map-reduce 框架,而不是为每个结果显式使用 futures。为了生成序列,我们可以定义一个迭代器来提供我们希望处理的整数。迭代器可以是正向或双向迭代器,它的实现是 QtConcurrent 框架所需的最低限度。

#include <iterator>

template <typename tag> class num_iterator : public std::iterator<tag, int, int, const int*, int> {
   int num = 0;
   using self = num_iterator;
   using base = std::iterator<tag, int, int, const int*, int>;
public:
   explicit num_iterator(int num = 0) : num(num) {}
   self &operator++() { num ++; return *this; }
   self &operator--() { num --; return *this; }
   self &operator+=(typename base::difference_type d) { num += d; return *this; }
   friend typename base::difference_type operator-(self lhs, self rhs) { return lhs.num - rhs.num; }
   bool operator==(self o) const { return num == o.num; }
   bool operator!=(self o) const { return !(*this == o); }
   typename base::reference operator*() const { return num; }
};

using num_f_iterator = num_iterator<std::forward_iterator_tag>;

result_type sum_approach2(int const N) {
   auto results = QtConcurrent::blockingMapped<QVector<result_type>>(num_f_iterator{0}, num_f_iterator{N}, map_function);
   return std::accumulate(results.cbegin(), results.cend(), result_type{}, sum_function);
}

using num_b_iterator = num_iterator<std::bidirectional_iterator_tag>;

result_type sum_approach3(int const N) {
   auto results = QtConcurrent::blockingMapped<QVector<result_type>>(num_b_iterator{0}, num_b_iterator{N}, map_function);
   return std::accumulate(results.cbegin(), results.cend(), result_type{}, sum_function);
}

我们可以放弃 std::accumulate 并使用 blockingMappedReduced 吗?当然:

result_type sum_approach4(int const N) {
   return QtConcurrent::blockingMappedReduced(num_b_iterator{0}, num_b_iterator{N},
                                              map_function, sum_modifier);
}

我们也可以试试随机访问迭代器:

using num_r_iterator = num_iterator<std::random_access_iterator_tag>;

result_type sum_approach5(int const N) {
   return QtConcurrent::blockingMappedReduced(num_r_iterator{0}, num_r_iterator{N},
                                              map_function, sum_modifier);
}

最后,我们可以从使用范围生成迭代器切换到预先计算的范围:

#include <numeric>

result_type sum_approach6(int const N) {
   QVector<int> sequence(N);
   std::iota(sequence.begin(), sequence.end(), 0);
   return QtConcurrent::blockingMappedReduced(sequence, map_function, sum_modifier);
}

当然,我们的目的是对这一切进行基准测试:

template <typename F> void benchmark(F fun, double const N) {
   QElapsedTimer timer;
   timer.start();
   auto result = fun(N);
   qDebug() << "sum:" << fixed << result << "took" << timer.elapsed()/N << "ms/item";
}

int main() {
   const int N = 1000000;
   benchmark(sum_approach1, N);
   benchmark(sum_approach2, N);
   benchmark(sum_approach3, N);
   benchmark(sum_approach4, N);
   benchmark(sum_approach5, N);
   benchmark(sum_approach6, N);
}

在我的系统上,在发布版本中,输出是:

sum: 4999995000000.000000 took 0.015778 ms/item
sum: 4999995000000.000000 took 0.003631 ms/item
sum: 4999995000000.000000 took 0.003610 ms/item
sum: 4999995000000.000000 took 0.005414 ms/item
sum: 4999995000000.000000 took 0.000011 ms/item
sum: 4999995000000.000000 took 0.000008 ms/item

请注意,在随机可迭代序列上使用 map-reduce 比使用 QtConcurrent::run 的开销要低 3 个数量级以上,并且比非随机可迭代序列快 2 个数量级可迭代的解决方案。