NQE: Record correlation metric in UMA

net/nqe/network_quality_estimator.cc

 // Copyright 2015 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 "net/nqe/network_quality_estimator.h"
 
 #include <algorithm>
 #include <cmath>
 #include <limits>
 #include <utility>
 #include <vector>
 
 #include "base/bind_helpers.h"
 #include "base/location.h"
 #include "base/logging.h"
 #include "base/metrics/histogram.h"
 #include "base/metrics/histogram_base.h"
+#include "base/metrics/sparse_histogram.h"
+#include "base/rand_util.h"
 #include "base/single_thread_task_runner.h"
 #include "base/strings/string_number_conversions.h"
 #include "base/threading/thread_task_runner_handle.h"
 #include "base/time/default_tick_clock.h"
 #include "base/trace_event/trace_event.h"
 #include "build/build_config.h"
 #include "net/base/load_flags.h"
 #include "net/base/load_timing_info.h"
 #include "net/base/network_interfaces.h"
 #include "net/base/url_util.h"
+#include "net/http/http_status_code.h"
 #include "net/nqe/socket_watcher_factory.h"
 #include "net/nqe/throughput_analyzer.h"
 #include "net/url_request/url_request.h"
+#include "net/url_request/url_request_status.h"
 #include "url/gurl.h"
 
 #if defined(OS_ANDROID)
 #include "net/android/cellular_signal_strength.h"
 #include "net/android/network_library.h"
 #endif  // OS_ANDROID
 
 namespace {
 
 // Default value of the half life (in seconds) for computing time weighted
@@ skipping 101 matching lines @@
 
 bool GetValueForVariationParam(
     const std::map<std::string, std::string>& variation_params,
     const std::string& parameter_name,
     int32_t* variations_value) {
   const auto it = variation_params.find(parameter_name);
   return it != variation_params.end() &&
          base::StringToInt(it->second, variations_value);
 }
 
+// Returns the variation value for |parameter_name|. If the value is
+// unavailable, |default_value| is returned.
+double GetDoubleValueForVariationParamWithDefaultValue(
+    const std::map<std::string, std::string>& variation_params,
+    const std::string& parameter_name,
+    double default_value) {
+  const auto it = variation_params.find(parameter_name);
+  if (it == variation_params.end())
+    return default_value;
+
+  double variations_value = default_value;
+  if (!base::StringToDouble(it->second, &variations_value))
+    return default_value;
+  return variations_value;
+}
+
 // Returns the algorithm that should be used for computing effective connection
 // type based on field trial params. Returns an empty string if a valid
 // algorithm paramter is not present in the field trial params.
 std::string GetEffectiveConnectionTypeAlgorithm(
     const std::map<std::string, std::string>& variation_params) {
   const auto it = variation_params.find("effective_connection_type_algorithm");
   if (it == variation_params.end())
     return std::string();
   return it->second;
 }
@@ skipping 47 matching lines @@
   static const char* const kSuffixes[] = {
       "0_20",     "20_60",     "60_140",    "140_300",      "300_620",
       "620_1260", "1260_2540", "2540_5100", "5100_Infinity"};
   for (size_t i = 0; i < arraysize(kSuffixes) - 1; ++i) {
     if (observed_throughput_kbps <= static_cast<float>((20 * (2 << i) - 20)))
       return kSuffixes[i];
   }
   return kSuffixes[arraysize(kSuffixes) - 1];
 }
 
+// The least significant kTrimBits of the metric will be discarded. If the
+// trimmed metric value is greater than what can be fit in kBitsPerMetric bits,
+// then the largest value that can be represented in kBitsPerMetric bits is
+// returned.
+const int32_t kTrimBits = 5;
+
+// Maximum number of bits in which one metric should fit. Restricting the amount
+// of space allocated to a single metric makes it possile to fit multiple
+// metrices in a single histogram sample, and ensures that all those metrices

[Ilya Sherman, 2016/07/21 21:38:08]

nit: s/metrices/metrics

[tbansal1, 2016/07/21 22:42:53]

Done.

+// are recorded together as a single tuple.
+const int32_t kBitsPerMetric = 7;
+
+static_assert(31 >= kBitsPerMetric * 4,
+              "Four metrics would not fit in a 32-bit int with first bit "
+              "reserved for sign");

[Ilya Sherman, 2016/07/21 21:38:08]

FWIW, the sign bit is safe to use, as long as you are careful to decode the bit
patterns correctly on the server-side.

[tbansal1, 2016/07/21 22:42:53]

Done.

+
+// Trims the |metric| by removing the last kTrimBits, and then rounding down
+// the |metric| such that the |metric| fits in kBitsPerMetric.
+int32_t FitInKBitsPerMetricBits(int32_t metric) {
+  // Remove the last kTrimBits. This will allow the metric to fit within
+  // kBitsPerMetric while losing only the least significant bits.
+  metric = metric >> kTrimBits;
+
+  // kLargestValuePossible is the largest value that can be recorded using
+  // kBitsPerMetric.
+  static const int32_t kLargestValuePossible = (1 << kBitsPerMetric) - 1;
+  if (metric > kLargestValuePossible) {
+    // Fit |metric| in kBitsPerMetric by clamping it down.
+    metric = kLargestValuePossible;
+  }
+  DCHECK_EQ(0, metric >> kBitsPerMetric);
+  return metric;
+}
+
 }  // namespace
 
 namespace net {
 
 NetworkQualityEstimator::NetworkQualityEstimator(
     std::unique_ptr<ExternalEstimateProvider> external_estimates_provider,
     const std::map<std::string, std::string>& variation_params)
     : NetworkQualityEstimator(std::move(external_estimates_provider),
                               variation_params,
                               false,
@@ skipping 29 matching lines @@
           NetworkID(NetworkChangeNotifier::ConnectionType::CONNECTION_UNKNOWN,
                     std::string())),
       downstream_throughput_kbps_observations_(weight_multiplier_per_second_),
       rtt_observations_(weight_multiplier_per_second_),
       effective_connection_type_at_last_main_frame_(
           EFFECTIVE_CONNECTION_TYPE_UNKNOWN),
       external_estimate_provider_(std::move(external_estimates_provider)),
       effective_connection_type_(EFFECTIVE_CONNECTION_TYPE_UNKNOWN),
       min_signal_strength_since_connection_change_(INT32_MAX),
       max_signal_strength_since_connection_change_(INT32_MIN),
+      correlation_uma_logging_probability_(
+          GetDoubleValueForVariationParamWithDefaultValue(
+              variation_params,
+              "correlation_logging_probability",
+              0.0)),
       weak_ptr_factory_(this) {
   static_assert(kDefaultHalfLifeSeconds > 0,
                 "Default half life duration must be > 0");
   static_assert(kMaximumNetworkQualityCacheSize > 0,
                 "Size of the network quality cache must be > 0");
   // This limit should not be increased unless the logic for removing the
   // oldest cache entry is rewritten to use a doubly-linked-list LRU queue.
   static_assert(kMaximumNetworkQualityCacheSize <= 10,
                 "Size of the network quality cache must <= 10");
   // None of the algorithms can have an empty name.
   DCHECK(algorithm_name_to_enum_.end() ==
          algorithm_name_to_enum_.find(std::string()));
 
   DCHECK_EQ(algorithm_name_to_enum_.size(),
             static_cast<size_t>(EffectiveConnectionTypeAlgorithm::
                                     EFFECTIVE_CONNECTION_TYPE_ALGORITHM_LAST));
   DCHECK_NE(EffectiveConnectionTypeAlgorithm::
                 EFFECTIVE_CONNECTION_TYPE_ALGORITHM_LAST,
             effective_connection_type_algorithm_);
+  DCHECK_LE(0.0, correlation_uma_logging_probability_);
+  DCHECK_GE(1.0, correlation_uma_logging_probability_);
 
   ObtainOperatingParams(variation_params);
   ObtainEffectiveConnectionTypeModelParams(variation_params);
   NetworkChangeNotifier::AddConnectionTypeObserver(this);
   if (external_estimate_provider_) {
     RecordExternalEstimateProviderMetrics(
         EXTERNAL_ESTIMATE_PROVIDER_STATUS_AVAILABLE);
     external_estimate_provider_->SetUpdatedEstimateDelegate(this);
   } else {
     RecordExternalEstimateProviderMetrics(
@@ skipping 239 matching lines @@
 
   LoadTimingInfo load_timing_info;
   request.GetLoadTimingInfo(&load_timing_info);
 
   // If the load timing info is unavailable, it probably means that the request
   // did not go over the network.
   if (load_timing_info.send_start.is_null() ||
       load_timing_info.receive_headers_end.is_null()) {
     return;
   }
+  DCHECK(!request.response_info().was_cached);
 
   // Duration between when the resource was requested and when the response
   // headers were received.
   base::TimeDelta observed_http_rtt =
       load_timing_info.receive_headers_end - load_timing_info.send_start;
   DCHECK_GE(observed_http_rtt, base::TimeDelta());
   if (observed_http_rtt < peak_network_quality_.http_rtt()) {
     peak_network_quality_ = nqe::internal::NetworkQuality(
         observed_http_rtt, peak_network_quality_.transport_rtt(),
         peak_network_quality_.downstream_throughput_kbps());
@@ skipping 121 matching lines @@
 void NetworkQualityEstimator::NotifyRequestCompleted(
     const URLRequest& request) {
   TRACE_EVENT0(TRACE_DISABLED_BY_DEFAULT("net"),
                "NetworkQualityEstimator::NotifyRequestCompleted");
   DCHECK(thread_checker_.CalledOnValidThread());
 
   if (!RequestSchemeIsHTTPOrHTTPS(request))
     return;
 
   throughput_analyzer_->NotifyRequestCompleted(request);
+  RecordCorrelationMetric(request);
+}
+
+void NetworkQualityEstimator::RecordCorrelationMetric(
+    const URLRequest& request) const {
+  DCHECK(thread_checker_.CalledOnValidThread());
+
+  // The histogram is recorded with probability
+  // |correlation_uma_logging_probability_| to reduce overhead involved with
+  // sparse histograms. Also, recording the correlation on each request is
+  // unnecessary.
+  if (RandDouble() >= correlation_uma_logging_probability_)
+    return;
+
+  if (request.response_info().was_cached ||
+      !request.response_info().network_accessed) {
+    return;
+  }
+
+  LoadTimingInfo load_timing_info;
+  request.GetLoadTimingInfo(&load_timing_info);
+  DCHECK(!load_timing_info.send_start.is_null() &&
+         !load_timing_info.receive_headers_end.is_null());
+
+  // Record UMA only for successful requests that have completed.
+  if (!request.status().is_success() || request.status().is_io_pending())
+    return;
+  if (request.GetResponseCode() != HTTP_OK)
+    return;
+  if (load_timing_info.receive_headers_end < last_main_frame_request_)
+    return;
+
+  const base::TimeTicks now = tick_clock_->NowTicks();
+  // Record UMA only for requests that started recently.
+  if (now - last_main_frame_request_ > base::TimeDelta::FromSeconds(15))
+    return;
+
+  DCHECK_GE(now, load_timing_info.send_start);
+
+  int32_t rtt = 0;
+
+  if (UseTransportRTT()) {
+    rtt = estimated_quality_at_last_main_frame_.transport_rtt() !=
+                  nqe::internal::InvalidRTT()
+              ? FitInKBitsPerMetricBits(
+                    estimated_quality_at_last_main_frame_.transport_rtt()
+                        .InMilliseconds())
+              : 0;
+  } else {
+    rtt = estimated_quality_at_last_main_frame_.http_rtt() !=
+                  nqe::internal::InvalidRTT()
+              ? FitInKBitsPerMetricBits(
+                    estimated_quality_at_last_main_frame_.http_rtt()
+                        .InMilliseconds())
+              : 0;
+  }
+
+  const int32_t downstream_throughput =
+      estimated_quality_at_last_main_frame_.downstream_throughput_kbps() !=
+              nqe::internal::kInvalidThroughput
+          ? FitInKBitsPerMetricBits(estimated_quality_at_last_main_frame_
+                                        .downstream_throughput_kbps())
+          : 0;
+
+  const int32_t resource_load_time = FitInKBitsPerMetricBits(
+      (now - load_timing_info.send_start).InMilliseconds());
+
+  int64_t resource_size = (request.GetTotalReceivedBytes() * 8) / 1024;
+  std::string histogram_name = "NQE.Correlation.ResourceLoadTime.0Kb_128Kb";

[Ilya Sherman, 2016/07/21 21:38:08]

nit: I'd just inline this constant below, where it is used.

[tbansal1, 2016/07/21 22:42:53]

Done.

+  if (resource_size >= (1 << kBitsPerMetric)) {
+    // Too large resource size (at least 128 Kb).
+    return;
+  }
+
+  DCHECK_EQ(
+      0, (rtt | downstream_throughput | resource_load_time | resource_size) >>
+             kBitsPerMetric);
+
+  // First 32 - (4* kBitsPerMetric) of the sample are unset. Next
+  // kBitsPerMetric of the sample contain |rtt|. Next
+  // kBitsPerMetric contain |downstream_throughput|. Next kBitsPerMetric
+  // contain |resource_load_time|. And, the last kBitsPerMetric
+  // contain |resource_size|.
+  int32_t sample = rtt;
+  sample = (sample << kBitsPerMetric) | downstream_throughput;
+  sample = (sample << kBitsPerMetric) | resource_load_time;
+  sample = (sample << kBitsPerMetric) | resource_size;

[Ilya Sherman, 2016/07/21 21:38:08]

How many distinct values for this metric is each client likely to record?  It
seems like the answer is "a lot", in which case there could be a significant
memory cost to recording this metric.

[tbansal1, 2016/07/21 22:42:54]

This could potentially be recorded once per HTTP resource (so ~100 per page
load). However, this metric is recorded randomly with a probability that is
configured by field trial variation param. See
|correlation_uma_logging_probability_| above.

By default, correlation_uma_logging_probability_ is 0.0. So, until I land the
server CL to enable logging, this CL would have no effect.

I plan to set it to not more than 0.01 (so approximately once per page loads).

+
+  UMA_HISTOGRAM_SPARSE_SLOWLY(histogram_name, sample);
 }
 
 void NetworkQualityEstimator::NotifyURLRequestDestroyed(
     const URLRequest& request) {
   DCHECK(thread_checker_.CalledOnValidThread());
 
-  NotifyRequestCompleted(request);
+  if (!RequestSchemeIsHTTPOrHTTPS(request))
+    return;
+
+  throughput_analyzer_->NotifyRequestCompleted(request);
 }
 
 void NetworkQualityEstimator::AddRTTObserver(RTTObserver* rtt_observer) {
   DCHECK(thread_checker_.CalledOnValidThread());
   rtt_observer_list_.AddObserver(rtt_observer);
 }
 
 void NetworkQualityEstimator::RemoveRTTObserver(RTTObserver* rtt_observer) {
   DCHECK(thread_checker_.CalledOnValidThread());
   rtt_observer_list_.RemoveObserver(rtt_observer);
@@ skipping 258 matching lines @@
         NetworkQualityEstimator::MetricUsage::
             USE_IF_AVAILABLE /* transport_rtt_metric */,
         NetworkQualityEstimator::MetricUsage::
             USE_IF_AVAILABLE /* downstream_throughput_kbps_metric */);
   }
   // Add additional algorithms here.
   NOTREACHED();
   return EFFECTIVE_CONNECTION_TYPE_UNKNOWN;
 }
 
+bool NetworkQualityEstimator::UseTransportRTT() const {
+  DCHECK(thread_checker_.CalledOnValidThread());
+
+  if (effective_connection_type_algorithm_ ==
+      EffectiveConnectionTypeAlgorithm::HTTP_RTT_AND_DOWNSTREAM_THROUGHOUT) {
+    return false;
+  }
+  if (effective_connection_type_algorithm_ ==
+      EffectiveConnectionTypeAlgorithm::
+          TRANSPORT_RTT_OR_DOWNSTREAM_THROUGHOUT) {
+    return true;
+  }
+  // Add additional algorithms here.
+  NOTREACHED();
+  return false;
+}
+
 NetworkQualityEstimator::EffectiveConnectionType
 NetworkQualityEstimator::GetRecentEffectiveConnectionTypeUsingMetrics(
     const base::TimeTicks& start_time,
     NetworkQualityEstimator::MetricUsage http_rtt_metric,
     NetworkQualityEstimator::MetricUsage transport_rtt_metric,
     NetworkQualityEstimator::MetricUsage downstream_throughput_kbps_metric)
     const {
   DCHECK(thread_checker_.CalledOnValidThread());
 
   // If the device is currently offline, then return
@@ skipping 348 matching lines @@
   NOTREACHED();
   return EFFECTIVE_CONNECTION_TYPE_UNKNOWN;
 }
 
 void NetworkQualityEstimator::SetTickClockForTesting(
     std::unique_ptr<base::TickClock> tick_clock) {
   DCHECK(thread_checker_.CalledOnValidThread());
   tick_clock_ = std::move(tick_clock);
 }
 
+double NetworkQualityEstimator::RandDouble() const {
+  return base::RandDouble();
+}
+
 void NetworkQualityEstimator::CacheNetworkQualityEstimate() {
   DCHECK(thread_checker_.CalledOnValidThread());
   DCHECK_LE(cached_network_qualities_.size(),
             static_cast<size_t>(kMaximumNetworkQualityCacheSize));
 
   // If the network name is unavailable, caching should not be performed.
   if (current_network_id_.id.empty())
     return;
 
   base::TimeDelta http_rtt = nqe::internal::InvalidRTT();
@@ skipping 122 matching lines @@
     NotifyObserversOfEffectiveConnectionTypeChanged() {
   DCHECK(thread_checker_.CalledOnValidThread());
 
   // TODO(tbansal): Add hysteresis in the notification.
   FOR_EACH_OBSERVER(
       EffectiveConnectionTypeObserver, effective_connection_type_observer_list_,
       OnEffectiveConnectionTypeChanged(effective_connection_type_));
 }
 
 }  // namespace net