naiveproxy/net/base/percentile_estimator.cc
2018-02-02 05:49:39 -05:00

101 lines
3.7 KiB
C++

// Copyright 2016 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "percentile_estimator.h"
#include "base/bind.h"
#include "base/callback.h"
#include "base/rand_util.h"
namespace {
// Random number wrapper to allow substitutions for testing.
int GenerateRand0To99() {
return base::RandInt(0, 99);
}
} // namespace
namespace net {
// The algorithm used for percentile estimation is "Algorithm 3" from
// https://arxiv.org/pdf/1407.1121v1.pdf. There are several parts to the
// algorithm:
// * The estimate is conditionally moved towards the sample by a step amount.
// This means that if the samples are clustered around a value the estimates
// will converge to that sample.
// * The percentile requested (e.g. 90%l) is handled by the conditional move.
// If the estimate is accurate, there is a chance equal to the percentile
// value that a sample will be lower than it, and a chance equal to
// 1-percentile that it will be higher. So the code balances those
// probabilities by increasing the estimate in the percentile fraction
// of the cases where the sample is over the estimate, and decreases the
// estimate in (1-percentile) fraction of the cases where the sample is under
// the estimate.
// E.g. in the case of the 90%l estimation, the estimate would
// move up in 90% of the cases in which the sample was above the
// estimate (which would be 10% of the total samples, presuming the
// estimate was accurate), and it would move down in 10% of the cases
// in which the sample was below the estimate.
// * Every time the estimate moves in the same direction, the step
// amount is increased by one, and every time the estimate reverses
// direction, the step amount is decreased (to 1, if greater than 1,
// by one, if zero or negative). The effective step amount is
// Max(step, 1).
// * If the estimate
// would be moved beyond the sample causing its move, it is moved to
// be equal to the same (and the step amount set to the distance to
// the sample). See the paper for further details.
PercentileEstimator::PercentileEstimator(int percentile, int initial_estimate)
: percentile_(percentile),
sign_positive_(true),
current_estimate_(initial_estimate),
current_step_(1),
generator_callback_(base::Bind(&GenerateRand0To99)) {}
PercentileEstimator::~PercentileEstimator() = default;
void PercentileEstimator::AddSample(int sample) {
int rand100 = generator_callback_.Run();
if (sample > current_estimate_ && rand100 > 1 - percentile_) {
current_step_ += sign_positive_ ? 1 : -1;
current_estimate_ += (current_step_ > 0) ? current_step_ : 1;
// Clamp movement to distance to sample.
if (current_estimate_ > sample) {
current_step_ -= current_estimate_ - sample;
current_estimate_ = sample;
}
// If we've reversed direction, reset the step down.
if (!sign_positive_ && current_step_ > 1)
current_step_ = 1;
sign_positive_ = true;
} else if (sample < current_estimate_ && rand100 > percentile_) {
current_step_ += !sign_positive_ ? 1 : -1;
current_estimate_ -= (current_step_ > 0) ? current_step_ : 1;
// Clamp movement to distance to sample.
if (current_estimate_ < sample) {
current_step_ -= sample - current_estimate_;
current_estimate_ = sample;
}
// If we've reversed direction, reset the step down.
if (sign_positive_ && current_step_ > 1)
current_step_ = 1;
sign_positive_ = false;
}
}
void PercentileEstimator::SetRandomNumberGeneratorForTesting(
RandomNumberCallback generator_callback) {
generator_callback_ = generator_callback;
}
} // namespace net