C++ multi-thread slower than single thread

Question

I have some code like this:

class MyTask {
public:
    run(size_t pool_size) {
        ... // do some pre things
        std::vector<std::string> name_list=read_a_list(); // read task list

        std::vector<std::pair<std::string, double>> result_list; // name & time

        boost::thread_pool pool(pool_size); // "pool_size" threads in pool
        size_t max_task=2*pool_size;        // max "2*pool_size" tasks in queue
        size_t task_number=0;               // using task_number to limit the number of tasks in queue
        boost::mutex task_number_mu;
        boost::condition_variable_any task_number_condition;

        for(size_t idx=0;idx<name_list.size();++idx){
             boost::unique_lock<boost::mutex> out_lock(task_number_mu);
             task_number_condition.wait(out_lock, [&] {
                 return task_number < max_task;
                 });
             ++task_number;
             boost::asio::post(pool,
                  [&,idx] {
                      {
                          boost::unique_lock<boost::mutex> in_lock(task_number_mu);
                          --task_number;
                          task_number_condition.notify_one();
                      }
                      std::string name=name_list[idx];
                      Timer timer; // a class using std::chrono to collect time
                      timer.Start();

                      A a=read_A_data(name+"_a.csv"); // one file
                      timer.Stop();
                      double time_a=timer.Elapsed();

                      B b=read_B_data(name+"_b"); // many files in "name_b" directory
                      timer.Stop();
                      double time_b=timer.Elapsed();

                      result_type result=do_some_selection(a,b); // very expensive function
                      timer.Stop();
                      double time_r=timer.Elapsed();

                      write_result(result,name+"_result.csv"); // one file
                      timer.Stop();
                      double time_w=timer.Elapsed();

                      ... // output idx, time_{a,b,r,w} by boost log

                      {
                           boost::lock_guard<boost::mutex> lock(result_list_mu);
                           result_list.emplace_back(std::make_pair(name,time_w));
                      }
                });//post
           }//for
      pool.join();
      ... // do some other things
   } //run

public :
   static A read_A_data(const std::string& name_a){
         ... // read "name_a" file, less than 1.5M 
   }
   static B read_B_data(const std::string& name_b){
         ... // read files in "name_b" directory, more than 10 files, 1M~10M per file
   }
   static result_type do_some_selection(A a,B b){
         result_type result;
         for(const auto& i:b){
              for(const auto& j:a){
                   if(b_fit_in_a(i,j)){ //b_fit_in_a() does not have loops
                       result.emplace_back(i);
                   }//if
              }//for j
         }//for i
         return result;
   }
   static void write_result(const result_type& result, const std::string& name_r){
         ... // write result to "name_r", about 2M~15M
   }
}

When I set pool_size to 1 (single thread), the time output is like:

1 14.7845 471.214 1491.16 1927.86
2 4.247 649.694 1327 1523.7
3 5.4375 924.407 2852.44 3276.1
4 4.1798 754.361 1078.97 1187.15
5 5.4944 1284.37 2935.02 3336.19
6 5.515 694.369 2825.79 3380.3
...

I have a Xeon-W which is 16C32T, so set pool_size to 8, and:

1 14.7919 2685.21 6600.4 7306.15
2 16.0127 2311.94 10517.2 12044.3
3 7.4403 2111.83 6210.49 7014.61
4 9.0292 2165.12 10482.5 11893
5 16.6851 1664.2 17282.7 20489.9
6 32.9876 6488.17 25730.6 25744.7
...

set 16, and:

1 22.5189 5324.67 18018.6 20386
2 17.1096 8670.3 21245.8 23229.1
3 17.9065 10930.7 27335.3 29961.55
4 20.6321 5227.19 30733 34926
5 25.104 2372.04 13810.9 15916.7
6 39.6723 18734.3 79300.1 79393.5
...

set 32, and:

1 39.3981 19159.7 43451.7 44527.1
2 51.1908 5693.48 43391.3 50314.4
3 42.4458 18068.6 59520.6 67359.4
4 44.1195 29214.7 70312.4 76902
5 64.1733 23071.1 86055.2 86146.7
6 44.1062 36277.5 89474.4 98104.7
...

I understand that multithreaded programs often have disk read/write problems, which explains the increase in time_a , time_b and time_w . But what confused me is that time_r increased a lot as well. do_some_selection is a static member function, so I don't think the threads will interact, but it seems that the more threads I use, the more time one task will take. What did I do wrong? How can I make these kind of tasks parallel?

Answer 1

First, you should display data in a sensible manner. As it is - it is hard to make any assessments. Like print time difference - so we can easily see how much time each task took instead of "how much time passed from the tasks' beginning".

Second, the tasks you run are mostly disk read/write and it is not quite parallelizable. So total execution time will not change by much. As you schedule several unrelated tasks - they will all finish up at about the same time were it a single thread. However, since you run multiple threads each task will compete for resources - thus delaying each tasks' completion till most tasks are done.

About why "unrelated computation-only" is slowed down. This depends a lot on the computation you perform. Cannot say too much as it now aside from some generic could-be reasons. From the looks of it, you perform some memory manipulation. RAM memory access is restricted by memory bus and is generally slow. In single-threaded case a lot of the data could be still stored in the processor's memory cache speeding up considerably the amount of time it takes to process it. But this is just a general guess of what the reason could be. You ought to make a deeper analysis to find the bottleneck - on PCs processors memory bus should be more than sufficient for multiple threads.