Add V8.GC.ParallelTaskLatencyMicroSeconds metric.

It will record the time-to-schedule-after-job-start for different
task types to try to highlight use cases where contention might
be a problem (and show improvements to it later).

Also introducing AsyncTimedHistogram to support this use case whose
reported timings go beyond a single scope (i.e. the async version of
ScopedTimedHistogram).

Bug: chromium:807606
Change-Id: Ib4d581fa8b001723dfe8c91102280e9608b4fabb
Reviewed-on: https://chromium-review.googlesource.com/899365Reviewed-by: Hannes Payer <hpayer@chromium.org>
Reviewed-by: Michael Lippautz <mlippautz@chromium.org>
Commit-Queue: Gabriel Charette <gab@chromium.org>
Cr-Commit-Position: refs/heads/master@{#51131}

src/counters.cc

@@ -84,7 +84,6 @@ void TimedHistogram::Start(base::ElapsedTimer* timer, Isolate* isolate) {
 
 void TimedHistogram::Stop(base::ElapsedTimer* timer, Isolate* isolate) {
   if (Enabled()) {
-    // Compute the delta between start and stop, in microseconds.
     int64_t sample = resolution_ == HistogramTimerResolution::MICROSECOND
                          ? timer->Elapsed().InMicroseconds()
                          : timer->Elapsed().InMilliseconds();
@@ -96,6 +95,21 @@ void TimedHistogram::Stop(base::ElapsedTimer* timer, Isolate* isolate) {
   }
 }
 
+void TimedHistogram::RecordAbandon(base::ElapsedTimer* timer,
+                                   Isolate* isolate) {
+  if (Enabled()) {
+    DCHECK(timer->IsStarted());
+    timer->Stop();
+    int64_t sample = resolution_ == HistogramTimerResolution::MICROSECOND
+                         ? base::TimeDelta::Max().InMicroseconds()
+                         : base::TimeDelta::Max().InMilliseconds();
+    AddSample(static_cast<int>(sample));
+  }
+  if (isolate != nullptr) {
+    Logger::CallEventLogger(isolate, name(), Logger::END, true);
+  }
+}
+
 Counters::Counters(Isolate* isolate)
     : isolate_(isolate),
       stats_table_(this),

src/counters.h

@@ -220,6 +220,12 @@ class Histogram {
   int max() const { return max_; }
   int num_buckets() const { return num_buckets_; }
 
+  // Asserts that |expected_counters| are the same as the Counters this
+  // Histogram reports to.
+  void AssertReportsToCounters(Counters* expected_counters) {
+    DCHECK_EQ(counters_, expected_counters);
+  }
+
  protected:
   Histogram() {}
   Histogram(const char* name, int min, int max, int num_buckets,
@@ -229,7 +235,9 @@ class Histogram {
         max_(max),
         num_buckets_(num_buckets),
         histogram_(nullptr),
-        counters_(counters) {}
+        counters_(counters) {
+    DCHECK(counters_);
+  }
 
   Counters* counters() const { return counters_; }
 
@@ -253,7 +261,7 @@ enum class HistogramTimerResolution { MILLISECOND, MICROSECOND };
 
 // A thread safe histogram timer. It also allows distributions of
 // nested timed results.
-class TimedHistogram : public Histogram {
+class V8_EXPORT_PRIVATE TimedHistogram : public Histogram {
  public:
   // Start the timer. Log if isolate non-null.
   void Start(base::ElapsedTimer* timer, Isolate* isolate);
@@ -261,6 +269,10 @@ class TimedHistogram : public Histogram {
   // Stop the timer and record the results. Log if isolate non-null.
   void Stop(base::ElapsedTimer* timer, Isolate* isolate);
 
+  // Records a TimeDelta::Max() result. Useful to record percentage of tasks
+  // that never got to run in a given scenario. Log if isolate non-null.
+  void RecordAbandon(base::ElapsedTimer* timer, Isolate* isolate);
+
  protected:
   friend class Counters;
   HistogramTimerResolution resolution_;
@@ -282,6 +294,7 @@ class TimedHistogramScope {
       : histogram_(histogram), isolate_(isolate) {
     histogram_->Start(&timer_, isolate);
   }
+
   ~TimedHistogramScope() { histogram_->Stop(&timer_, isolate_); }
 
  private:
@@ -292,6 +305,42 @@ class TimedHistogramScope {
   DISALLOW_IMPLICIT_CONSTRUCTORS(TimedHistogramScope);
 };
 
+// Helper class for recording a TimedHistogram asynchronously with manual
+// controls (it will not generate a report if destroyed without explicitly
+// triggering a report). |async_counters| should be a shared_ptr to
+// |histogram->counters()|, making it is safe to report to an
+// AsyncTimedHistogram after the associated isolate has been destroyed.
+// AsyncTimedHistogram can be moved/copied to avoid computing Now() multiple
+// times when the times of multiple tasks are identical; each copy will generate
+// its own report.
+class V8_EXPORT_PRIVATE AsyncTimedHistogram {
+ public:
+  explicit AsyncTimedHistogram(TimedHistogram* histogram,
+                               std::shared_ptr<Counters> async_counters)
+      : histogram_(histogram), async_counters_(std::move(async_counters)) {
+    histogram_->AssertReportsToCounters(async_counters_.get());
+    histogram_->Start(&timer_, nullptr);
+  }
+
+  ~AsyncTimedHistogram() = default;
+
+  AsyncTimedHistogram(const AsyncTimedHistogram& other) = default;
+  AsyncTimedHistogram& operator=(const AsyncTimedHistogram& other) = default;
+  AsyncTimedHistogram(AsyncTimedHistogram&& other) = default;
+  AsyncTimedHistogram& operator=(AsyncTimedHistogram&& other) = default;
+
+  // Records the time elapsed to |histogram_| and stops |timer_|.
+  void RecordDone() { histogram_->Stop(&timer_, nullptr); }
+
+  // Records TimeDelta::Max() to |histogram_| and stops |timer_|.
+  void RecordAbandon() { histogram_->RecordAbandon(&timer_, nullptr); }
+
+ private:
+  base::ElapsedTimer timer_;
+  TimedHistogram* histogram_;
+  std::shared_ptr<Counters> async_counters_;
+};
+
 // Helper class for scoping a TimedHistogram, where the histogram is selected at
 // stop time rather than start time.
 // TODO(leszeks): This is heavily reliant on TimedHistogram::Start() doing
@@ -1168,7 +1217,9 @@ class RuntimeCallTimerScope {
   HT(compile_script_no_cache_because_cache_too_cold,                           \
      V8.CompileScriptMicroSeconds.NoCache.CacheTooCold, 1000000, MICROSECOND)  \
   HT(compile_script_on_background,                                             \
-     V8.CompileScriptMicroSeconds.BackgroundThread, 1000000, MICROSECOND)
+     V8.CompileScriptMicroSeconds.BackgroundThread, 1000000, MICROSECOND)      \
+  HT(gc_parallel_task_latency, V8.GC.ParallelTaskLatencyMicroSeconds, 1000000, \
+     MICROSECOND)
 
 #define AGGREGATABLE_HISTOGRAM_TIMER_LIST(AHT) \
   AHT(compile_lazy, V8.CompileLazyMicroSeconds)

src/heap/heap.cc

@@ -2100,7 +2100,7 @@ void Heap::Scavenge() {
     {
       // Parallel phase scavenging all copied and promoted objects.
       TRACE_GC(tracer(), GCTracer::Scope::SCAVENGER_SCAVENGE_PARALLEL);
-      job.Run();
+      job.Run(isolate()->async_counters());
       DCHECK(copied_list.IsGlobalEmpty());
       DCHECK(promotion_list.IsGlobalEmpty());
     }

src/heap/item-parallel-job.h

@@ -8,8 +8,10 @@
 #include <vector>
 
 #include "src/base/macros.h"
+#include "src/base/optional.h"
 #include "src/base/platform/semaphore.h"
 #include "src/cancelable-task.h"
+#include "src/counters.h"
 #include "src/utils.h"
 #include "src/v8.h"
 
@@ -27,6 +29,9 @@ class Isolate;
 //
 // Items need to be marked as finished after processing them. Task and Item
 // ownership is transferred to the job.
+//
+// Each parallel (non-main thread) task will report the time between the job
+// being created and it being scheduled to |gc_parallel_task_latency_histogram|.
 class ItemParallelJob {
  public:
   class Task;
@@ -63,7 +68,12 @@ class ItemParallelJob {
           cur_index_(0),
           items_considered_(0),
           on_finish_(nullptr) {}
-    virtual ~Task() {}
+    virtual ~Task() {
+      // The histogram is reset in RunInternal(). If it's still around it means
+      // this task was cancelled before being scheduled.
+      if (gc_parallel_task_latency_histogram_)
+        gc_parallel_task_latency_histogram_->RecordAbandon();
+    }
 
     virtual void RunInParallel() = 0;
 
@@ -90,19 +100,33 @@ class ItemParallelJob {
     // Sets up state required before invoking Run(). If
     // |start_index is >= items_.size()|, this task will not process work items
     // (some jobs have more tasks than work items in order to parallelize post-
-    // processing, e.g. scavenging).
+    // processing, e.g. scavenging). If |gc_parallel_task_latency_histogram| is
+    // provided, it will be used to report histograms on the latency between
+    // posting the task and it being scheduled.
     void SetupInternal(base::Semaphore* on_finish, std::vector<Item*>* items,
-                       size_t start_index) {
+                       size_t start_index,
+                       base::Optional<AsyncTimedHistogram>
+                           gc_parallel_task_latency_histogram) {
       on_finish_ = on_finish;
       items_ = items;
-      if (start_index < items->size())
+
+      if (start_index < items->size()) {
         cur_index_ = start_index;
-      else
+      } else {
         items_considered_ = items_->size();
       }
 
+      gc_parallel_task_latency_histogram_ =
+          std::move(gc_parallel_task_latency_histogram);
+    }
+
     // We don't allow overriding this method any further.
     void RunInternal() final {
+      if (gc_parallel_task_latency_histogram_) {
+        gc_parallel_task_latency_histogram_->RecordDone();
+        gc_parallel_task_latency_histogram_.reset();
+      }
+
       RunInParallel();
       on_finish_->Signal();
     }
@@ -111,6 +135,7 @@ class ItemParallelJob {
     size_t cur_index_;
     size_t items_considered_;
     base::Semaphore* on_finish_;
+    base::Optional<AsyncTimedHistogram> gc_parallel_task_latency_histogram_;
 
     friend class ItemParallelJob;
     friend class Item;
@@ -140,11 +165,16 @@ class ItemParallelJob {
   int NumberOfItems() const { return static_cast<int>(items_.size()); }
   int NumberOfTasks() const { return static_cast<int>(tasks_.size()); }
 
-  void Run() {
+  // Runs this job. Reporting metrics in a thread-safe manner to
+  // |async_counters|.
+  void Run(std::shared_ptr<Counters> async_counters) {
     DCHECK_GT(tasks_.size(), 0);
     const size_t num_items = items_.size();
     const size_t num_tasks = tasks_.size();
 
+    AsyncTimedHistogram gc_parallel_task_latency_histogram(
+        async_counters->gc_parallel_task_latency(), async_counters);
+
     // Some jobs have more tasks than items (when the items are mere coarse
     // grain tasks that generate work dynamically for a second phase which all
     // tasks participate in). Some jobs even have 0 items to preprocess but
@@ -174,7 +204,9 @@ class ItemParallelJob {
       // assigning work items.
       DCHECK_IMPLIES(start_index >= num_items, i >= num_tasks_processing_items);
 
-      task->SetupInternal(pending_tasks_, &items_, start_index);
+      task->SetupInternal(pending_tasks_, &items_, start_index,
+                          i > 0 ? gc_parallel_task_latency_histogram
+                                : base::Optional<AsyncTimedHistogram>());
       task_ids[i] = task->id();
       if (i > 0) {
         V8::GetCurrentPlatform()->CallOnBackgroundThread(

src/heap/mark-compact.cc

@@ -2072,7 +2072,7 @@ void MinorMarkCompactCollector::MarkRootSetInParallel() {
         job.AddTask(
             new YoungGenerationMarkingTask(isolate(), this, worklist(), i));
       }
-      job.Run();
+      job.Run(isolate()->async_counters());
       DCHECK(worklist()->IsGlobalEmpty());
     }
   }
@@ -3253,7 +3253,7 @@ void MarkCompactCollectorBase::CreateAndExecuteEvacuationTasks(
       evacuators[i]->AddObserver(migration_observer);
     job->AddTask(new PageEvacuationTask(heap()->isolate(), evacuators[i]));
   }
-  job->Run();
+  job->Run(isolate()->async_counters());
   for (int i = 0; i < wanted_num_tasks; i++) {
     evacuators[i]->Finalize();
     delete evacuators[i];
@@ -3936,7 +3936,7 @@ void MarkCompactCollector::UpdatePointersAfterEvacuation() {
           isolate(),
           GCTracer::BackgroundScope::MC_BACKGROUND_EVACUATE_UPDATE_POINTERS));
     }
-    updating_job.Run();
+    updating_job.Run(isolate()->async_counters());
   }
 
   {
@@ -3968,7 +3968,7 @@ void MarkCompactCollector::UpdatePointersAfterEvacuation() {
             isolate(),
             GCTracer::BackgroundScope::MC_BACKGROUND_EVACUATE_UPDATE_POINTERS));
       }
-      updating_job.Run();
+      updating_job.Run(isolate()->async_counters());
       heap()->array_buffer_collector()->FreeAllocationsOnBackgroundThread();
     }
   }
@@ -4030,7 +4030,7 @@ void MinorMarkCompactCollector::UpdatePointersAfterEvacuation() {
   {
     TRACE_GC(heap()->tracer(),
              GCTracer::Scope::MINOR_MC_EVACUATE_UPDATE_POINTERS_SLOTS);
-    updating_job.Run();
+    updating_job.Run(isolate()->async_counters());
     heap()->array_buffer_collector()->FreeAllocationsOnBackgroundThread();
   }
 

test/unittests/heap/item-parallel-job-unittest.cc

@@ -202,7 +202,7 @@ TEST_F(ItemParallelJobTest, SimpleTaskWithNoItemsRuns) {
                       parallel_job_semaphore());
   job.AddTask(new SimpleTask(i_isolate(), &did_run));
 
-  job.Run();
+  job.Run(i_isolate()->async_counters());
   EXPECT_TRUE(did_run);
 }
 
@@ -214,7 +214,7 @@ TEST_F(ItemParallelJobTest, SimpleTaskWithSimpleItemRuns) {
 
   job.AddItem(new ItemParallelJob::Item);
 
-  job.Run();
+  job.Run(i_isolate()->async_counters());
   EXPECT_TRUE(did_run);
 }
 
@@ -244,7 +244,7 @@ TEST_F(ItemParallelJobTest, MoreTasksThanItems) {
     job.AddItem(new SimpleItem);
   }
 
-  job.Run();
+  job.Run(i_isolate()->async_counters());
 
   for (int i = 0; i < kNumTasks; i++) {
     // Only the first kNumItems tasks should have been assigned a work item.
@@ -261,7 +261,7 @@ TEST_F(ItemParallelJobTest, SingleThreadProcessing) {
   for (int i = 0; i < kItems; i++) {
     job.AddItem(new SimpleItem(&was_processed[i]));
   }
-  job.Run();
+  job.Run(i_isolate()->async_counters());
   for (int i = 0; i < kItems; i++) {
     EXPECT_TRUE(was_processed[i]);
   }
@@ -282,7 +282,7 @@ TEST_F(ItemParallelJobTest, DistributeItemsMultipleTasks) {
     job.AddTask(
         new TaskProcessingOneItem(i_isolate(), &barrier, wait_when_done));
   }
-  job.Run();
+  job.Run(i_isolate()->async_counters());
   for (int i = 0; i < kItemsAndTasks; i++) {
     EXPECT_TRUE(was_processed[i]);
   }
@@ -296,7 +296,7 @@ TEST_F(ItemParallelJobTest, DifferentItems) {
   job.AddItem(new ItemA());
   job.AddItem(new ItemB());
   job.AddTask(new TaskForDifferentItems(i_isolate(), &item_a, &item_b));
-  job.Run();
+  job.Run(i_isolate()->async_counters());
   EXPECT_TRUE(item_a);
   EXPECT_TRUE(item_b);
 }