Why the IMT spend more time than sequence mode?

ROOT
1

ROOT Version:6.26.14
Platform:Ubuntu 22.04
Compiler: gcc 11.04

Problem

The time using ROOT::EnableImplicitMT() is greater than that of sequence mode?

Code

MT Mode

#include <iostream>
#include <chrono>
void mt_first() {
    int mumber_threads = 20;
    ROOT::EnableImplicitMT(number_threads);
    
    auto file = TFile::Open("./dstarmb.root");
    TTree *tree = nullptr;
    file->GetObject<TTree>("h42", tree);
   
    const auto time_0{std::chrono::steady_clock::now()};
    const auto entries = tree->GetEntries();
    for(auto i : ROOT::TSeqUL(entries)) {
       tree->GetEntry(i); 
    }
    const auto time_1{std::chrono::steady_clock::now()};
    const std::chrono::duration<double> elapsed_seconds{time_1 - time_0};

    std::cout << "Time Spend: " << elapsed_seconds.count() << " s." << "\n";
}

Seq Mode

#include <iostream>
#include <chrono>
void seq_first() {
    ROOT::DisableImplicitMT();
    
    auto file = TFile::Open("./dstarmb.root");
    TTree *tree = nullptr;
    file->GetObject<TTree>("h42", tree);
   
    const auto time_0{std::chrono::steady_clock::now()};
    const auto entries = tree->GetEntries();
    for(auto i = 0; i < entries; ++i) {
       tree->GetEntry(i); 
    }
    const auto time_1{std::chrono::steady_clock::now()};
    const std::chrono::duration<double> elapsed_seconds{time_1 - time_0};

    std::cout << "Time Spend: " << elapsed_seconds.count() << " s." << "\n";
}

Result

MT: 0.402365 s
Seq: 0.175275 s

2

Hi,

Welcome to the forum, and thanks for the interesting post.

Before diving into any investigation, could you make sure this is not a fluctuation? You can do that by running the sequential mode first and the parallelised after, by running the code several times to check the stability of the measurement…

Cheers,
Danilo

3

Hi,

Do you mean that put both codes into one file?
I ran the code about three times,and get the same results.
I guess the amount of data is small, thus parallel execution may not be as fast as sequential execution.

Best
Kun ZHU

4

That is of course one factor. It’s also true that setting up the thread pool also has some overhead. Without a profile, it’s hard to say for sure. However I still find the difference very small, especially since it’s in the subsecond range!

Thanks a lot,
D

5

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.