General Exploration - Store Bound metric always at 0

[Milian_W_]

Hey all,

I'm trying a simple text-book example on false sharing where four threads continuously write to distinct data on the same shared cache line. Using the General Exploration analysis, I used to see the "Store Bound" value flagged with a high Store Latency and False Sharing child metric.

Now, it looks like this:

Elapsed Time:	0.828s
    Clockticks:	10,659,000,000
    Instructions Retired:	5,616,800,000
    CPI Rate:	1.898
    MUX Reliability:	0.998
    Front-End Bound:	5.0%
    Bad Speculation:	6.1%
    Back-End Bound:	72.5%
        Memory Bound:	21.3%
            L1 Bound:	24.2%
            L2 Bound:	N/A with HT on
            L3 Bound:	N/A with HT on
            DRAM Bound:	N/A with HT on
            Store Bound:	0.0%
        Core Bound:	51.3%
            Divider:	32.1%
            Port Utilization:	65.2%
                Cycles of 0 Ports Utilized:	51.3%
                Cycles of 1 Port Utilized:	41.0%
                Cycles of 2 Ports Utilized:	21.7%
                Cycles of 3+ Ports Utilized:	5.0%
    Retiring:	16.4%
    Total Thread Count:	5
    Paused Time:	0s

Note how Store Bound got 0.0%... I can increase the workload but that doesn't change anything. The child-metrics also look OK, which makes this even more odd... Store Latency got 76.6% but is supposedly unreliably due to MUX issues or lack of PMU events. Fale Sharing metric is at 13.9% and not unreliable.

Can anyone explain what's going on here? The code for my test case is the following:

#include <cmath>
#include <iostream>
#include <thread>
#include <future>

struct Results
{
    virtual ~Results() = default;
    virtual unsigned int* data1() = 0;
    virtual unsigned int* data2() = 0;
    virtual unsigned int* data3() = 0;
    virtual unsigned int* data4() = 0;
};

// the two data elements are distinct but share a single cache line
struct Unaligned : Results
{
    unsigned int* data1() override
    {
        return &m_data1;
    }
    unsigned int* data2() override
    {
        return &m_data2;
    }
    unsigned int* data3() override
    {
        return &m_data3;
    }
    unsigned int* data4() override
    {
        return &m_data4;
    }
    unsigned int m_data1 = 0;
    unsigned int m_data2 = 0;
    unsigned int m_data3 = 0;
    unsigned int m_data4 = 0;
};

// the two data elements are distinct and live on separate cache lines
struct Aligned : Results
{
    unsigned int* data1() override
    {
        return &m_data1;
    }
    unsigned int* data2() override
    {
        return &m_data2;
    }
    unsigned int* data3() override
    {
        return &m_data3;
    }
    unsigned int* data4() override
    {
        return &m_data4;
    }
    unsigned int m_data1 = 0;
    alignas(64) unsigned int m_data2 = 0;
    alignas(64) unsigned int m_data3 = 0;
    alignas(64) unsigned int m_data4 = 0;
};

void do_something(volatile unsigned int* result)
{
    for (size_t i = 0; i < 100000000; ++i) {
        *result += sqrt(i);
    }
}

int main(int argc, char** /*argv*/)
{
    // when any command line argument is given, use the aligned data, otherwise
    // use the unaligned data to show the effect of false sharing
    std::unique_ptr<Results> results(argc > 1
        ? static_cast<Results*>(new Aligned)
        : static_cast<Results*>(new Unaligned));
    {
        const auto f1 = std::async(std::launch::async, do_something,
                                   results->data1());
        const auto f2 = std::async(std::launch::async, do_something,
                                   results->data2());
        const auto f3 = std::async(std::launch::async, do_something,
                                   results->data3());
        const auto f4 = std::async(std::launch::async, do_something,
                                   results->data4());
    }

    std::cout << "random sums: " << *results->data1()
              << ", " << *results->data2()
              << ", " << *results->data3()
              << ", " << *results->data4() << '\n';
    return 0;
}

[Milian_W_]

Ah, and I forgot to add: I have seen this on the following hardware, all with VTune 2017 Update 4, running Ubuntu 17.04.

Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz
Intel(R) Core(TM) i7-7700 CPU @ 3.60GHz
Intel(R) Core(TM) i5-2400 CPU @ 3.10GHz
 

[Milian_W_]

Ping? Can someone please have a look at this and explain me what's going on here?

Thanks

[Dmitry_R_Intel1]

Hello,

When you are saying that you used to see it as store bound - what do you mean? Did you see it with previous VTune versions? Or on other platforms? What has changed?

[Milian_W_]

Yes, it was with a previous VTune version, but also potentially an older Linux kernel. I don't know the exact version anymore, sorry. Can you or anyone else reproduce this issue?