Using std::chrono::steady_clock instead of system_clock in RaftKeeper

[lzydmxy]

Description

The following is flame graph of RaftKeeper instance which serving as ClickHouse metadata store. A lot of cpu time cost () on std::chrono::system::clock (because of syscall).Maybe we could optimize the std::chrono::steady_clock instead of std::chrono::steady_clock

small test

#include <iostream>
#include <chrono>

using steady_clock = std::chrono::steady_clock;
using system_clock = std::chrono::system_clock;

int main()
{
    auto t1 = steady_clock::now();
    auto t2 = steady_clock::now();
    auto t3 = steady_clock::now();
    std::cout << "chrono::steady_clock: " << (t3-t1).count() << "ns" << std::endl;


    auto t4 = steady_clock::now();
    auto t5 = system_clock::now();
    auto t6 = steady_clock::now();
    std::cout << "chrono::system_clock: " << (t6-t4).count() << "ns" << std::endl;
}

./main
chrono::steady_clock: 197ns
chrono::system_clock: 501ns

twice std::chrono::steady_clock cost 197ns
once std::chrono::steady_clock once std::chrono::system_clock cost 500ns

once std::chrono::system_clock may cost 400ns and once std::chrono::steady_clock may cost 100ns

Are you willing to submit PR?

[JackyWoo]

hi, I make a simple test with the following code snippet:

inline size_t getCurrentTimestamp()
{
    auto now = std::chrono::steady_clock::now();
    return std::chrono::duration_cast<std::chrono::nanoseconds>(now.time_since_epoch()).count();
}

int main()
{
    size_t n = 3;
    auto t1 = getCurrentTimestamp();

    for (size_t i=0; i<n; i++)
        std::chrono::steady_clock::now();

    auto t2 = getCurrentTimestamp();

    for (size_t i=0; i<n; i++)
        std::chrono::system_clock::now();

    auto t3 = getCurrentTimestamp();

    std::cout << "steady clock: " << std::to_string(t2 - t1) << "ns" << std::endl;
    std::cout << "system clock: " << std::to_string(t3 - t2) << "ns" << std::endl;
}

I found that the result are quiet different in multiple runing:

The 1st run:

steady clock: 167ns
system clock: 208ns

The 2nd run:

steady clock: 250ns
system clock: 166ns

The 3rd run:

steady clock: 291ns
system clock: 292ns

Maybe there is something we are missing.

PS: my env is MacOS(M3) clang16

[JackyWoo]

Different result in Ubuntu 20.4

Nearly same in multiple runnings

steady clock: 261ns
system clock: 577ns

[JackyWoo]

Will be fixed by #352