Reporting top cpu and memory consumers via rappor on chromeos

chrome/browser/chromeos/resource_reporter/resource_reporter.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 "chrome/browser/chromeos/resource_reporter/resource_reporter.h"
+
+#include <stdint.h>

[Daniel Erat, 2015/11/23 23:57:19]

nit: cstdint

[afakhry, 2015/11/24 01:52:33]

Done.

+
+#include <queue>
+
+#include "base/bind.h"
+#include "base/memory/singleton.h"
+#include "base/rand_util.h"
+#include "base/strings/utf_string_conversions.h"
+#include "base/sys_info.h"
+#include "chrome/browser/browser_process.h"
+#include "chrome/browser/task_management/task_manager_interface.h"
+#include "components/rappor/rappor_service.h"
+#include "content/public/browser/browser_thread.h"
+
+namespace chromeos {
+
+namespace {
+
+// The task manager refresh interval, currently at 1 minute.
+const int64_t kRefreshIntervalSeconds = 60;
+
+// 1 GB in bytes.
+const int64_t kMemory1GB = 1024 * 1024 * 1024;
+
+// 800 MB in bytes.

[Daniel Erat, 2015/11/23 23:57:19]

nit: these comments aren't very useful; i'd just put one at the top like
"Various sizes in bytes." and then delete the lines between all of these (or
better, just get rid of these entirely and inline the values below; these will
never change and look like they're only used once)

[afakhry, 2015/11/24 01:52:33]

Done. Went with the first suggestion.

+const int64_t kMemory800MB = 800 * 1024 * 1024;
+
+// 600 MB in bytes.
+const int64_t kMemory600MB = 600 * 1024 * 1024;
+
+// 400 MB in bytes.
+const int64_t kMemory400MB = 400 * 1024 * 1024;
+
+// 200 MB in bytes.
+const int64_t kMemory200MB = 200 * 1024 * 1024;
+
+// The name of the Rappor metric to report the CPU usage.
+const char kCpuRapporMetric[] = "ResourceReporter.Cpu";
+
+// The name of the Rappor metric to report the memory usage.
+const char kMemoryRapporMetric[] = "ResourceReporter.Memory";
+
+// The name of the string field of the Rappor metrics in which we'll record the
+// task's Rappor sample name.
+const char kRapporTaskStringField[] = "task";
+
+// The name of the flags field of the Rappor metrics in which we'll store the
+// priority of the process on which the task is running.
+const char kRapporPriorityFlagsField[] = "priority";
+
+// The name of the flags field of the CPU usage Rappor metrics in which we'll
+// record the number of cores in the current system.
+const char kRapporNumCoresRangeFlagsField[] = "num_cores_range";
+
+// The name of the flags field of the Rappor metrics in which we'll store the
+// CPU / memory usage ranges.
+const char kRapporUsageRangeFlagsField[] = "usage_range";
+
+// Currently set to be one day.
+const int kMinimumTimeBetweenReportsInMS = 1 * 24 * 60 * 60 * 1000;

[Daniel Erat, 2015/11/23 23:57:19]

nit: "Ms" looks like it's more common than "MS" in the chrome source. you're
also using e.g. "Cpu" rather than "CPU" elsewhere

[afakhry, 2015/11/24 01:52:33]

Done.

+
+// A functor to sort the TaskRecords by their |cpu|.
+struct TaskRecordByCpuSorter {

[Daniel Erat, 2015/11/23 23:57:19]

would a name like "TaskRecordCpuLessThan" be clearer?

[afakhry, 2015/11/24 01:52:33]

Done.

+  bool operator()(ResourceReporter::TaskRecord* const& lhs,
+                  ResourceReporter::TaskRecord* const& rhs) const {
+    if (lhs->cpu == rhs->cpu)
+      return lhs->id < rhs->id;
+    return lhs->cpu < rhs->cpu;
+  }
+};
+
+// A functor to sort the TaskRecords by their |memory|.
+struct TaskRecordByMemorySorter {

[Daniel Erat, 2015/11/23 23:57:19]

ditto here with "TaskRecordMemoryLessThan"

[afakhry, 2015/11/24 01:52:33]

Done.

+  bool operator()(ResourceReporter::TaskRecord* const& lhs,
+                  ResourceReporter::TaskRecord* const& rhs) const {
+    if (lhs->memory == rhs->memory)
+      return lhs->id < rhs->id;
+    return lhs->memory < rhs->memory;
+  }
+};
+
+}  // namespace
+
+ResourceReporter::TaskRecord::TaskRecord(task_management::TaskId the_id)
+    : id(the_id), cpu(0.0), memory(0), is_background(false) {
+}
+
+ResourceReporter::TaskRecord::TaskRecord(task_management::TaskId the_id,
+                                         const std::string& task_name,
+                                         double the_cpu,
+                                         int64_t the_memory,
+                                         bool background)
+    : id(the_id),
+      task_name_for_rappor(task_name),
+      cpu(the_cpu),
+      memory(the_memory),
+      is_background(background) {
+}
+
+ResourceReporter::~ResourceReporter() {
+}
+
+// static
+ResourceReporter* ResourceReporter::GetInstance() {
+  return base::Singleton<ResourceReporter>::get();
+}
+
+void ResourceReporter::StartMonitoring() {
+  DCHECK_CURRENTLY_ON(content::BrowserThread::UI);
+
+  if (is_monitoring_)
+    return;
+
+  is_monitoring_ = true;
+  task_management::TaskManagerInterface::GetTaskManager()->AddObserver(this);
+  memory_pressure_listener_.reset(new base::MemoryPressureListener(
+      base::Bind(&ResourceReporter::OnMemoryPressure, base::Unretained(this))));
+}
+
+void ResourceReporter::StopMonitoring() {
+  DCHECK_CURRENTLY_ON(content::BrowserThread::UI);
+
+  if (!is_monitoring_)
+    return;
+
+  is_monitoring_ = false;
+  memory_pressure_listener_.reset();
+  task_management::TaskManagerInterface::GetTaskManager()->RemoveObserver(this);
+}
+
+void ResourceReporter::OnTaskAdded(task_management::TaskId id) {
+  // Ignore this event.
+}
+
+void ResourceReporter::OnTaskToBeRemoved(task_management::TaskId id) {
+  auto itr = task_records_.find(id);

[Daniel Erat, 2015/11/23 23:57:19]

nit: i think that "it" is much more common for iterators in the chromium source
(also below)

[afakhry, 2015/11/24 01:52:33]

Done.

+  if (itr == task_records_.end())
+    return;
+
+  // Must be erased from the sorted set first.
+  // Note: this could mean that the sorted records are now less than
+  // |kTopConsumerCount| with other records in |task_records_| that can be
+  // added now. That's ok, we ignore this case.
+  auto cpu_itr = std::find(task_records_by_cpu_.begin(),
+                           task_records_by_cpu_.end(),
+                           itr->second.get());
+  if (cpu_itr != task_records_by_cpu_.end())
+    task_records_by_cpu_.erase(cpu_itr);
+
+  auto memory_itr = std::find(task_records_by_memory_.begin(),
+                              task_records_by_memory_.end(),
+                              itr->second.get());
+  if (memory_itr != task_records_by_memory_.end())
+    task_records_by_memory_.erase(memory_itr);
+
+  task_records_.erase(itr);
+}
+
+void ResourceReporter::OnTasksRefreshed(
+    const task_management::TaskIdList& task_ids) {
+  // A priority queue to sort the task records by their |cpu|. Greatest |cpu|
+  // first.
+  std::priority_queue<TaskRecord*,
+                      std::vector<TaskRecord*>,
+                      TaskRecordByCpuSorter> records_by_cpu_queue;
+  // A priority queue to sort the task records by their |memory|. Greatest
+  // |memory| first.
+  std::priority_queue<TaskRecord*,
+                      std::vector<TaskRecord*>,
+                      TaskRecordByMemorySorter> records_by_memory_queue;
+  task_records_by_cpu_.clear();
+  task_records_by_cpu_.reserve(kTopConsumersCount);
+  task_records_by_memory_.clear();
+  task_records_by_memory_.reserve(kTopConsumersCount);
+
+  for (const auto& id : task_ids) {
+    const double cpu_usage = observed_task_manager()->GetCpuUsage(id);
+    const int64_t memory_usage =
+        observed_task_manager()->GetPhysicalMemoryUsage(id);
+
+    // Browser and GPU processes are reported later using UMA histograms as they
+    // don't have any privacy issues.
+    const auto task_type = observed_task_manager()->GetType(id);
+    switch (task_type) {
+      case task_management::Task::UNKNOWN:
+      case task_management::Task::ZYGOTE:
+        break;
+
+      case task_management::Task::BROWSER:
+        last_browser_process_cpu_ = cpu_usage;
+        last_browser_process_memory_ = memory_usage != -1 ? memory_usage : 0;

[Daniel Erat, 2015/11/23 23:57:19]

memory_usage >= 0? (also below)

[afakhry, 2015/11/24 01:52:32]

Done.

+        break;
+
+      case task_management::Task::GPU:
+        last_gpu_process_cpu_ = cpu_usage;
+        last_gpu_process_memory_ = memory_usage != -1 ? memory_usage : 0;
+        break;
+
+      default:
+        // Other tasks types will be reported using Rappor.
+        TaskRecord* task_data = nullptr;
+        auto itr = task_records_.find(id);
+        if (itr == task_records_.end()) {
+          task_data = new TaskRecord(id);
+          task_records_[id] = make_scoped_ptr(task_data);
+        } else {
+          task_data = itr->second.get();
+        }
+
+        DCHECK_EQ(task_data->id, id);
+        task_data->task_name_for_rappor =
+            observed_task_manager()->GetTaskNameForRappor(id);
+        task_data->cpu = cpu_usage;
+        task_data->memory = memory_usage;
+        task_data->is_background =
+            observed_task_manager()->IsTaskOnBackgroundedProcess(id);
+
+        // Push only valid or useful data to both priority queues. They might
+        // end up having more records than |kTopConsumerCount|, that's fine.
+        // We'll take care of that next.
+        if (task_data->cpu > 0)
+          records_by_cpu_queue.push(task_data);
+        if (task_data->memory > 0)
+          records_by_memory_queue.push(task_data);
+    }
+  }
+
+  // Sort the |kTopConsumersCount| task records by their CPU and memory usage.
+  while (!records_by_cpu_queue.empty() &&
+         task_records_by_cpu_.size() < kTopConsumersCount) {
+    task_records_by_cpu_.push_back(records_by_cpu_queue.top());
+    records_by_cpu_queue.pop();
+  }
+
+  while (!records_by_memory_queue.empty() &&
+         task_records_by_memory_.size() < kTopConsumersCount) {
+    task_records_by_memory_.push_back(records_by_memory_queue.top());
+    records_by_memory_queue.pop();
+  }
+}
+
+// static
+const size_t ResourceReporter::kTopConsumersCount = 10U;
+
+ResourceReporter::ResourceReporter()
+    : TaskManagerObserver(base::TimeDelta::FromSeconds(kRefreshIntervalSeconds),
+                          task_management::REFRESH_TYPE_CPU |
+                              task_management::REFRESH_TYPE_MEMORY |
+                                  task_management::REFRESH_TYPE_PRIORITY),
+      system_cpu_cores_range_(GetCurrentSystemCpuCoresRange()) {
+}
+
+// static
+scoped_ptr<rappor::Sample> ResourceReporter::CreateRapporSample(
+    rappor::RapporService* rappor_service,
+    const ResourceReporter::TaskRecord& task_record) {
+  scoped_ptr<rappor::Sample> sample(rappor_service->CreateSample(
+      rappor::UMA_RAPPOR_TYPE));
+  sample->SetStringField(kRapporTaskStringField,
+                         task_record.task_name_for_rappor);
+  sample->SetFlagsField(kRapporPriorityFlagsField,
+                        task_record.is_background ? BACKGROUND : FOREGROUND,
+                        PRIORITIES_NUM);
+  return sample.Pass();
+}
+
+// static
+ResourceReporter::CpuUsageRange
+ResourceReporter::GetCpuUsageRange(double cpu) {
+  if (cpu > 60.0)
+    return RANGE_ABOVE_60_PERCENT;
+  if (cpu > 30.0)
+    return RANGE_30_TO_60_PERCENT;
+  if (cpu > 10.0)
+    return RANGE_10_TO_30_PERCENT;
+
+  return RANGE_0_TO_10_PERCENT;
+}
+
+// static
+ResourceReporter::MemoryUsageRange
+ResourceReporter::GetMemoryUsageRange(int64_t memory_in_bytes) {
+  if (memory_in_bytes > kMemory1GB)
+    return RANGE_ABOVE_1_GB;
+  if (memory_in_bytes > kMemory800MB)
+    return RANGE_800_TO_1_GB;
+  if (memory_in_bytes > kMemory600MB)
+    return RANGE_600_TO_800_MB;
+  if (memory_in_bytes > kMemory400MB)
+      return RANGE_400_TO_600_MB;
+  if (memory_in_bytes > kMemory200MB)
+      return RANGE_200_TO_400_MB;
+
+  return RANGE_0_TO_200_MB;
+}
+
+// static
+ResourceReporter::CpuCoresNumberRange
+ResourceReporter::GetCurrentSystemCpuCoresRange() {
+  const int cpus = base::SysInfo::NumberOfProcessors();
+
+  if (cpus > 16)
+    return RANGE_CORES_ABOVE_16_CORES;
+  if (cpus > 8)
+    return RANGE_CORES_9_TO_16_CORES;
+  if (cpus > 4)
+    return RANGE_CORES_5_TO_8_CORES;
+  if (cpus > 2)
+    return RANGE_CORES_3_TO_4_CORES;
+  if (cpus == 2)
+    return RANGE_CORES_2_CORES;
+  if (cpus == 1)
+    return RANGE_CORES_1_CORE;
+
+  NOTREACHED();
+  return RANGE_CORES_NA;
+}
+
+bool ResourceReporter::ShouldRecordSamples() {
+  if (is_first_memory_pressure_event_) {
+    is_first_memory_pressure_event_ = false;

[Daniel Erat, 2015/11/23 23:57:19]

i think it's unexpected to have a Should/Is method mutating internal state like
this. i'd rather that you had some have_seen_first_memory_pressure_event_ or
similar member that got updated by OnMemoryPressure().

[afakhry, 2015/11/24 01:52:32]

Thanks! |have_seen_first_memory_pressure_event_| is indeed a much better name.
Done.

+    return true;
+  }
+
+  return (base::TimeTicks::Now() - last_memory_pressure_event_time_) >=
+      base::TimeDelta::FromMilliseconds(kMinimumTimeBetweenReportsInMS);
+}
+
+const ResourceReporter::TaskRecord* ResourceReporter::SampleTaskByCpu() const {
+  TaskRecord* sampled_task = nullptr;
+  double cpu_weights_sum = 0;
+  for (const auto& task_data : task_records_by_cpu_) {

[Daniel Erat, 2015/11/23 23:57:19]

it's non-obvious to me what this is doing without staring at the code for a
while. please add a comment summarizing it.

[afakhry, 2015/11/24 01:52:32]

This is an algorithm for doing weighted random sampling, taking the CPU usage of
the tasks as their weights. I added a comment, but I doubt it helps much. Maybe
add a link to this page: https://en.wikipedia.org/wiki/Reservoir_sampling ?

+    if ((base::RandDouble() * (cpu_weights_sum + task_data->cpu)) >=
+        cpu_weights_sum) {
+      sampled_task = task_data;
+    }
+    cpu_weights_sum += task_data->cpu;
+  }
+
+  return sampled_task;
+}
+
+const ResourceReporter::TaskRecord*
+ResourceReporter::SampleTaskByMemory() const {
+  TaskRecord* sampled_task = nullptr;
+  int64_t memory_weights_sum = 0;
+  for (const auto& task_data : task_records_by_memory_) {
+    if ((base::RandDouble() *
+         (memory_weights_sum + task_data->memory)) >=
+          memory_weights_sum) {
+      sampled_task = task_data;
+    }
+    memory_weights_sum += task_data->memory;
+  }
+
+  return sampled_task;
+}
+
+void ResourceReporter::OnMemoryPressure(
+    MemoryPressureLevel memory_pressure_level) {
+  if (memory_pressure_level >=
+      MemoryPressureLevel::MEMORY_PRESSURE_LEVEL_MODERATE) {
+    // Report browser and GPU processes usage using UMA histograms.
+    UMA_HISTOGRAM_ENUMERATION("ResourceReporter.BrowserProcess.CpuUsage",
+                              GetCpuUsageRange(last_browser_process_cpu_),
+                              CPU_RANGES_NUM);
+    UMA_HISTOGRAM_ENUMERATION("ResourceReporter.BrowserProcess.MemoryUsage",
+                              GetMemoryUsageRange(last_browser_process_memory_),
+                              MEMORY_RANGES_NUM);
+    UMA_HISTOGRAM_ENUMERATION("ResourceReporter.GpuProcess.CpuUsage",
+                              GetCpuUsageRange(last_gpu_process_cpu_),
+                              CPU_RANGES_NUM);
+    UMA_HISTOGRAM_ENUMERATION("ResourceReporter.GpuProcess.MemoryUsage",
+                              GetMemoryUsageRange(last_gpu_process_memory_),
+                              MEMORY_RANGES_NUM);
+
+    // For the rest of tasks, report them using Rappor.
+    auto rappor_service = g_browser_process->rappor_service();
+    if (!rappor_service)
+      return;
+
+    if (!ShouldRecordSamples())
+      return;
+    last_memory_pressure_event_time_ = base::TimeTicks::Now();
+
+    // Use weighted random sampling to select a task to report in the CPU
+    // metric.
+    const TaskRecord* sampled_cpu_task = SampleTaskByCpu();
+    if (sampled_cpu_task) {
+      scoped_ptr<rappor::Sample> cpu_sample(
+          CreateRapporSample(rappor_service, *sampled_cpu_task));
+      cpu_sample->SetFlagsField(kRapporNumCoresRangeFlagsField,
+                                system_cpu_cores_range_,
+                                CORES_RANGES_NUM);
+      cpu_sample->SetFlagsField(kRapporUsageRangeFlagsField,
+                                GetCpuUsageRange(sampled_cpu_task->cpu),
+                                CPU_RANGES_NUM);
+      rappor_service->RecordSampleObj(kCpuRapporMetric, cpu_sample.Pass());
+    }
+
+    // Use weighted random sampling to select a task to report in the memory
+    // metric.
+    const TaskRecord* sampled_memory_task = SampleTaskByMemory();
+    if (sampled_memory_task) {
+      scoped_ptr<rappor::Sample> memory_sample(
+          CreateRapporSample(rappor_service, *sampled_memory_task));
+      memory_sample->SetFlagsField(
+          kRapporUsageRangeFlagsField,
+          GetMemoryUsageRange(sampled_memory_task->memory),
+          MEMORY_RANGES_NUM);
+      rappor_service->RecordSampleObj(kMemoryRapporMetric,
+                                      memory_sample.Pass());
+    }
+  }
+}
+
+}  // namespace chromeos