jfr/support/jfrAdaptiveSampler.cpp

/*
* Copyright (c) 2020, Oracle and/or its affiliates. All rights reserved.
* Copyright (c) 2020, Datadog, Inc. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/

#include "precompiled.hpp"
#include "jfr/support/jfrAdaptiveSampler.hpp"
#include "jfr/utilities/jfrRandom.inline.hpp"
#include "jfr/utilities/jfrSpinlockHelper.hpp"
#include "jfr/utilities/jfrTime.hpp"
#include "jfr/utilities/jfrTimeConverter.hpp"
#include "jfr/utilities/jfrTryLock.hpp"
#include "logging/log.hpp"
#include "runtime/atomic.hpp"
#include "utilities/globalDefinitions.hpp"
#include <cmath>

JfrSamplerWindow::JfrSamplerWindow() :
  _params(),
  _end_ticks(0),
  _sampling_interval(1),
  _projected_population_size(0),
  _measured_population_size(0) {}

JfrAdaptiveSampler::JfrAdaptiveSampler() :
  _prng(this),
  _window_0(NULL),
  _window_1(NULL),
  _active_window(NULL),
  _avg_population_size(0),
  _ewma_population_size_alpha(0),
  _acc_debt_carry_limit(0),
  _acc_debt_carry_count(0),
  _lock(0) {}

JfrAdaptiveSampler::~JfrAdaptiveSampler() {
  delete _window_0;
  delete _window_1;
}

bool JfrAdaptiveSampler::initialize() {
  assert(_window_0 == NULL, "invariant");
  _window_0 = new JfrSamplerWindow();
  if (_window_0 == NULL) {
    return false;
  }
  assert(_window_1 == NULL, "invariant");
  _window_1 = new JfrSamplerWindow();
  if (_window_1 == NULL) {
    return false;
  }
  _active_window = _window_0;
  return true;
}

/*
 * The entry point to the sampler.
 */
bool JfrAdaptiveSampler::sample(int64_t timestamp) {
  bool expired_window;
  const bool result = active_window()->sample(timestamp, &expired_window);
  if (expired_window) {
    JfrTryLock mutex(&_lock);
    if (mutex.acquired()) {
      rotate_window(timestamp);
    }
  }
  return result;
}

inline const JfrSamplerWindow* JfrAdaptiveSampler::active_window() const {
  return Atomic::load_acquire(&_active_window);
}

inline int64_t now() {
  return JfrTicks::now().value();
}

inline bool JfrSamplerWindow::is_expired(int64_t timestamp) const {
  const int64_t end_ticks = Atomic::load(&_end_ticks);
  return timestamp == 0 ? now() >= end_ticks : timestamp >= end_ticks;
}

bool JfrSamplerWindow::sample(int64_t timestamp, bool* expired_window) const {
  assert(expired_window != NULL, "invariant");
  *expired_window = is_expired(timestamp);
  return *expired_window ? false : sample();
}

inline bool JfrSamplerWindow::sample() const {
  const size_t ordinal = Atomic::add(&_measured_population_size, static_cast<size_t>(1));
  return ordinal <= _projected_population_size && ordinal % _sampling_interval == 0;
}

// Called exclusively by the holder of the lock when a window is determined to have expired.
void JfrAdaptiveSampler::rotate_window(int64_t timestamp) {
  assert(_lock, "invariant");
  const JfrSamplerWindow* const current = active_window();
  assert(current != NULL, "invariant");
  if (!current->is_expired(timestamp)) {
    // Someone took care of it.
    return;
  }
  rotate(current);
}

// Subclasses can call this to immediately trigger a reconfiguration of the sampler.
// There is no need to await the expiration of the current active window.
void JfrAdaptiveSampler::reconfigure() {
  assert(_lock, "invariant");
  rotate(active_window());
}

// Call next_window_param() to report the expired window and to retreive params for the next window.
void JfrAdaptiveSampler::rotate(const JfrSamplerWindow* expired) {
  assert(expired == active_window(), "invariant");
  install(configure(next_window_params(expired), expired));
}

inline void JfrAdaptiveSampler::install(const JfrSamplerWindow* next) {
  assert(next != active_window(), "invariant");
  Atomic::release_store(&_active_window, next);
}

const JfrSamplerWindow* JfrAdaptiveSampler::configure(const JfrSamplerParams& params, const JfrSamplerWindow* expired) {
  assert(_lock, "invariant");
  if (params.reconfigure) {
    // Store updated params once to both windows.
    const_cast<JfrSamplerWindow*>(expired)->_params = params;
    next_window(expired)->_params = params;
    configure(params);
  }
  JfrSamplerWindow* const next = set_rate(params, expired);
  next->initialize(params);
  return next;
}

/*
 * Exponentially Weighted Moving Average (EWMA):
 *
 * Y is a datapoint (at time t)
 * S is the current EMWA (at time t-1)
 * alpha represents the degree of weighting decrease, a constant smoothing factor between 0 and 1.
 *
 * A higher alpha discounts older observations faster.
 * Returns the new EWMA for S
*/

inline double exponentially_weighted_moving_average(double Y, double alpha, double S) {
  return alpha * Y + (1 - alpha) * S;
}

inline double compute_ewma_alpha_coefficient(size_t lookback_count) {
  return lookback_count <= 1 ? 1 : static_cast<double>(1) / static_cast<double>(lookback_count);
}

inline size_t compute_accumulated_debt_carry_limit(const JfrSamplerParams& params) {
  if (params.window_duration_ms == 0 || params.window_duration_ms >= MILLIUNITS) {
    return 1;
  }
  return MILLIUNITS / params.window_duration_ms;
}

void JfrAdaptiveSampler::configure(const JfrSamplerParams& params) {
  assert(params.reconfigure, "invariant");
  _avg_population_size = 0;
  _ewma_population_size_alpha = compute_ewma_alpha_coefficient(params.window_lookback_count);
  _acc_debt_carry_limit = compute_accumulated_debt_carry_limit(params);
  _acc_debt_carry_count = _acc_debt_carry_limit;
  params.reconfigure = false;
}

inline int64_t millis_to_countertime(int64_t millis) {
  return JfrTimeConverter::nanos_to_countertime(millis * NANOSECS_PER_MILLISEC);
}

void JfrSamplerWindow::initialize(const JfrSamplerParams& params) {
  assert(_sampling_interval >= 1, "invariant");
  if (params.window_duration_ms == 0) {
    Atomic::store(&_end_ticks, static_cast<int64_t>(0));
    return;
  }
  Atomic::store(&_measured_population_size, static_cast<size_t>(0));
  const int64_t end_ticks = now() + millis_to_countertime(params.window_duration_ms);
  Atomic::store(&_end_ticks, end_ticks);
}

/*
 * Based on what it has learned from the past, the sampler creates a future 'projection',
 * a speculation, or model, of what the situation will be like during the next window.
 * This projection / model is used to derive values for the parameters, which are estimates for
 * collecting a sample set that, should the model hold, is as close as possible to the target,
 * i.e. the set point, which is a function of the number of sample_points_per_window + amortization.
 * The model is a geometric distribution over the number of trials / selections required until success.
 * For each window, the sampling interval is a random variable from this geometric distribution.
 */
JfrSamplerWindow* JfrAdaptiveSampler::set_rate(const JfrSamplerParams& params, const JfrSamplerWindow* expired) {
  JfrSamplerWindow* const next = next_window(expired);
  assert(next != expired, "invariant");
  const size_t sample_size = project_sample_size(params, expired);
  if (sample_size == 0) {
    next->_projected_population_size = 0;
    return next;
  }
  next->_sampling_interval = derive_sampling_interval(sample_size, expired);
  assert(next->_sampling_interval >= 1, "invariant");
  next->_projected_population_size = sample_size * next->_sampling_interval;
  return next;
}

inline JfrSamplerWindow* JfrAdaptiveSampler::next_window(const JfrSamplerWindow* expired) const {
  assert(expired != NULL, "invariant");
  return expired == _window_0 ? _window_1 : _window_0;
}

size_t JfrAdaptiveSampler::project_sample_size(const JfrSamplerParams& params, const JfrSamplerWindow* expired) {
  return params.sample_points_per_window + amortize_debt(expired);
}

/*
 * When the sampler is configured to maintain a rate, is employs the concepts
 * of 'debt' and 'accumulated debt'. 'Accumulated debt' can be thought of as
 * a cumulative error term, and is indicative for how much the sampler is
 * deviating from a set point, i.e. the ideal target rate. Debt accumulates naturally
 * as a function of undersampled windows, caused by system fluctuations,
 * i.e. too small populations.
 *
 * A specified rate is implicitly a _maximal_ rate, so the sampler must ensure
 * to respect this 'limit'. Rates are normalized as per-second ratios, hence the
 * limit to respect is on a per second basis. During this second, the sampler
 * has freedom to dynamically re-adjust, and it does so by 'amortizing'
 * accumulated debt over a certain number of windows that fall within the second.
 *
 * Intuitively, accumulated debt 'carry over' from the predecessor to the successor
 * window if within the allowable time frame (determined in # of 'windows' given by
 * _acc_debt_carry_limit). The successor window will sample more points to make amends,
 * or 'amortize' debt accumulated by its predecessor(s).
 */
size_t JfrAdaptiveSampler::amortize_debt(const JfrSamplerWindow* expired) {
  assert(expired != NULL, "invariant");
  const intptr_t accumulated_debt = expired->accumulated_debt();
  assert(accumulated_debt <= 0, "invariant");
  if (_acc_debt_carry_count == _acc_debt_carry_limit) {
    _acc_debt_carry_count = 1;
    return 0;
  }
  ++_acc_debt_carry_count;
  return -accumulated_debt; // negation
}

inline size_t JfrSamplerWindow::max_sample_size() const {
  return _projected_population_size / _sampling_interval;
}

// The sample size is derived from the measured population size.
size_t JfrSamplerWindow::sample_size() const {
  const size_t size = population_size();
  return size > _projected_population_size ? max_sample_size() : size / _sampling_interval;
}

size_t JfrSamplerWindow::population_size() const {
  return Atomic::load(&_measured_population_size);
}

intptr_t JfrSamplerWindow::accumulated_debt() const {
  return _projected_population_size == 0 ? 0 : static_cast<intptr_t>(_params.sample_points_per_window - max_sample_size()) + debt();
}

intptr_t JfrSamplerWindow::debt() const {
  return _projected_population_size == 0 ? 0 : static_cast<intptr_t>(sample_size() - _params.sample_points_per_window);
}

/*
 * Inverse transform sampling from a uniform to a geometric distribution.
 *
 * PMF: f(x)  = P(X=x) = ((1-p)^x-1)p
 *
 * CDF: F(x)  = P(X<=x) = 1 - (1-p)^x
 *
 * Inv
 * CDF: F'(u) = ceil( ln(1-u) / ln(1-p) ) // u = random uniform, 0.0 < u < 1.0
 *
 */
inline size_t next_geometric(double p, double u) {
  assert(u >= 0.0, "invariant");
  assert(u <= 1.0, "invariant");
  if (u == 0.0) {
    u = 0.01;
  } else if (u == 1.0) {
    u = 0.99;
  }
  // Inverse CDF for the geometric distribution.
  return ceil(log(1.0 - u) / log(1.0 - p));
}

size_t JfrAdaptiveSampler::derive_sampling_interval(double sample_size, const JfrSamplerWindow* expired) {
  assert(sample_size > 0, "invariant");
  const size_t population_size = project_population_size(expired);
  if (population_size <= sample_size) {
    return 1;
  }
  assert(population_size > 0, "invariant");
  const double projected_probability = sample_size / population_size;
  return next_geometric(projected_probability, _prng.next_uniform());
}

// The projected population size is an exponentially weighted moving average, a function of the window_lookback_count.
inline size_t JfrAdaptiveSampler::project_population_size(const JfrSamplerWindow* expired) {
  assert(expired != NULL, "invariant");
  _avg_population_size = exponentially_weighted_moving_average(expired->population_size(), _ewma_population_size_alpha, _avg_population_size);
  return _avg_population_size;
}

/* GTEST support */
JfrGTestFixedRateSampler::JfrGTestFixedRateSampler(size_t sample_points_per_window, size_t window_duration_ms, size_t lookback_count) : JfrAdaptiveSampler(), _params() {
  _sample_size_ewma = 0.0;
  _params.sample_points_per_window = sample_points_per_window;
  _params.window_duration_ms = window_duration_ms;
  _params.window_lookback_count = lookback_count;
  _params.reconfigure = true;
}

bool JfrGTestFixedRateSampler::initialize() {
  const bool result = JfrAdaptiveSampler::initialize();
  JfrSpinlockHelper mutex(&_lock);
  reconfigure();
  return result;
}

/*
 * To start debugging the sampler: -Xlog:jfr+system+throttle=debug
 * It will log details of each expired window together with an average sample size.
 *
 * Excerpt:
 *
 * "JfrGTestFixedRateSampler: avg.sample size: 19.8377, window set point: 20 ..."
 *
 * Monitoring the relation of average sample size to the window set point, i.e the target,
 * is a good indicator of how the sampler is performing over time.
 *
 */
static void log(const JfrSamplerWindow* expired, double* sample_size_ewma) {
  assert(sample_size_ewma != NULL, "invariant");
  if (log_is_enabled(Debug, jfr, system, throttle)) {
    *sample_size_ewma = exponentially_weighted_moving_average(expired->sample_size(), compute_ewma_alpha_coefficient(expired->params().window_lookback_count), *sample_size_ewma);
    log_debug(jfr, system, throttle)("JfrGTestFixedRateSampler: avg.sample size: %0.4f, window set point: %zu, sample size: %zu, population size: %zu, ratio: %.4f, window duration: %zu ms\n",
      *sample_size_ewma, expired->params().sample_points_per_window, expired->sample_size(), expired->population_size(),
      expired->population_size() == 0 ? 0 : (double)expired->sample_size() / (double)expired->population_size(),
      expired->params().window_duration_ms);
  }
}

/*
 * This is the feedback control loop.
 *
 * The JfrAdaptiveSampler engine calls this when a sampler window has expired, providing
 * us with an opportunity to perform some analysis.To reciprocate, we returns a set of
 * parameters, possibly updated, for the engine to apply to the next window.
 */
const JfrSamplerParams& JfrGTestFixedRateSampler::next_window_params(const JfrSamplerWindow* expired) {
  assert(expired != NULL, "invariant");
  assert(_lock, "invariant");
  log(expired, &_sample_size_ewma);
  return _params;
}