rippled/src/libxrpl/core/detail/LoadMonitor.cpp

#include <xrpl/core/LoadMonitor.h>

#include <xrpl/basics/Log.h>
#include <xrpl/basics/UptimeClock.h>
#include <xrpl/beast/utility/Journal.h>
#include <xrpl/core/LoadEvent.h>

#include <chrono>
#include <mutex>

namespace xrpl {

/*

TODO
----

- Use Journal for logging

*/

//------------------------------------------------------------------------------

LoadMonitor::Stats::Stats() : latencyAvg(0), latencyPeak(0)
{
}

//------------------------------------------------------------------------------

LoadMonitor::LoadMonitor(beast::Journal j)
    : mLatencyMSAvg(0)
    , mLatencyMSPeak(0)
    , mTargetLatencyAvg(0)
    , mTargetLatencyPk(0)
    , mLastUpdate(UptimeClock::now())
    , j_(j)
{
}

// VFALCO NOTE WHY do we need "the mutex?" This dependence on
//         a hidden global, especially a synchronization primitive,
//         is a flawed design.
//         It's not clear exactly which data needs to be protected.
//
// call with the mutex
void
LoadMonitor::update()
{
    using namespace std::chrono_literals;
    auto now = UptimeClock::now();
    if (now == mLastUpdate)  // current
        return;

    // VFALCO TODO Why 8?
    if ((now < mLastUpdate) || (now > (mLastUpdate + 8s)))
    {
        // way out of date
        mCounts = 0;
        mLatencyEvents = 0;
        mLatencyMSAvg = 0ms;
        mLatencyMSPeak = 0ms;
        mLastUpdate = now;
        return;
    }

    // do exponential decay
    /*
        David:

        "Imagine if you add 10 to something every second. And you
         also reduce it by 1/4 every second. It will "idle" at 40,
         corresponding to 10 counts per second."
    */
    do
    {
        mLastUpdate += 1s;
        mCounts -= ((mCounts + 3) / 4);
        mLatencyEvents -= ((mLatencyEvents + 3) / 4);
        mLatencyMSAvg -= (mLatencyMSAvg / 4);
        mLatencyMSPeak -= (mLatencyMSPeak / 4);
    } while (mLastUpdate < now);
}

void
LoadMonitor::addLoadSample(LoadEvent const& s)
{
    using namespace std::chrono;

    auto const total = s.runTime() + s.waitTime();
    // Don't include "jitter" as part of the latency
    auto const latency = total < 2ms ? 0ms : round<milliseconds>(total);

    if (latency > 500ms)
    {
        auto mj = (latency > 1s) ? j_.warn() : j_.info();
        JLOG(mj) << "Job: " << s.name() << " run: " << round<milliseconds>(s.runTime()).count()
                 << "ms"
                 << " wait: " << round<milliseconds>(s.waitTime()).count() << "ms";
    }

    addSamples(1, latency);
}

/* Add multiple samples
   @param count The number of samples to add
   @param latencyMS The total number of milliseconds
*/
void
LoadMonitor::addSamples(int count, std::chrono::milliseconds latency)
{
    std::scoped_lock const sl(mutex_);

    update();
    mCounts += count;
    mLatencyEvents += count;
    mLatencyMSAvg += latency;
    mLatencyMSPeak += latency;

    auto const latencyPeak = mLatencyEvents * latency * 4 / count;

    if (mLatencyMSPeak < latencyPeak)
        mLatencyMSPeak = latencyPeak;
}

void
LoadMonitor::setTargetLatency(std::chrono::milliseconds avg, std::chrono::milliseconds pk)
{
    mTargetLatencyAvg = avg;
    mTargetLatencyPk = pk;
}

bool
LoadMonitor::isOverTarget(std::chrono::milliseconds avg, std::chrono::milliseconds peak)
{
    using namespace std::chrono_literals;
    return (mTargetLatencyPk > 0ms && (peak > mTargetLatencyPk)) ||
        (mTargetLatencyAvg > 0ms && (avg > mTargetLatencyAvg));
}

bool
LoadMonitor::isOver()
{
    std::scoped_lock const sl(mutex_);

    update();

    if (mLatencyEvents == 0)
        return false;

    return isOverTarget(
        mLatencyMSAvg / (mLatencyEvents * 4), mLatencyMSPeak / (mLatencyEvents * 4));
}

LoadMonitor::Stats
LoadMonitor::getStats()
{
    using namespace std::chrono_literals;
    Stats stats;

    std::scoped_lock const sl(mutex_);

    update();

    stats.count = mCounts / 4;

    if (mLatencyEvents == 0)
    {
        stats.latencyAvg = 0ms;
        stats.latencyPeak = 0ms;
    }
    else
    {
        stats.latencyAvg = mLatencyMSAvg / (mLatencyEvents * 4);
        stats.latencyPeak = mLatencyMSPeak / (mLatencyEvents * 4);
    }

    stats.isOverloaded = isOverTarget(stats.latencyAvg, stats.latencyPeak);

    return stats;
}

}  // namespace xrpl