queue = priorityQueues[i];
synchronized (queue) {
merged.addAll(queue);
}
}
while (merged.size() > size) {
merged.remove();
}
return merged;
}
}
/**
* Which {@link ProfilerTask}s are profiled.
*/
public enum ProfiledTaskKinds {
/**
* Do not profile anything.
*
* Performance is best with this case, but we lose critical path analysis and slowest
* operation tracking.
*/
NONE {
@Override
boolean isProfiling(ProfilerTask type) {
return false;
}
},
/**
* Profile on a few, known-to-be-slow tasks.
*
*
Performance is somewhat decreased in comparison to {@link #NONE}, but we still track the
* slowest operations (VFS).
*/
SLOWEST {
@Override
boolean isProfiling(ProfilerTask type) {
return type.collectsSlowestInstances();
}
},
/**
* Profile all tasks.
*
*
This is in use when {@code --profile} is specified.
*/
ALL {
@Override
boolean isProfiling(ProfilerTask type) {
return true;
}
};
/** Whether the Profiler collects data for the given task type. */
abstract boolean isProfiling(ProfilerTask type);
}
private Clock clock;
private ProfiledTaskKinds profiledTaskKinds;
private volatile long profileStartTime = 0L;
private volatile boolean recordAllDurations = false;
private AtomicInteger taskId = new AtomicInteger();
private TaskStack taskStack;
private Queue taskQueue;
private DataOutputStream out;
private Timer timer;
private IOException saveException;
private ObjectDescriber describer;
@SuppressWarnings("unchecked")
private final SlowestTaskAggregator[] slowestTasks =
new SlowestTaskAggregator[ProfilerTask.values().length];
private final StatRecorder[] tasksHistograms = new StatRecorder[ProfilerTask.values().length];
private Profiler() {
initHistograms();
for (ProfilerTask task : ProfilerTask.values()) {
if (task.slowestInstancesCount != 0) {
slowestTasks[task.ordinal()] = new SlowestTaskAggregator(task.slowestInstancesCount);
}
}
}
private void initHistograms() {
for (ProfilerTask task : ProfilerTask.values()) {
if (task.isVfs()) {
Map> vfsHeuristics =
VfsHeuristics.vfsTypeHeuristics;
List recorders = new ArrayList<>(vfsHeuristics.size());
for (Entry> e : vfsHeuristics.entrySet()) {
recorders.add(new RecorderAndPredicate(
new SingleStatRecorder(task + " " + e.getKey(), HISTOGRAM_BUCKETS), e.getValue()));
}
tasksHistograms[task.ordinal()] = new PredicateBasedStatRecorder(recorders);
} else {
tasksHistograms[task.ordinal()] = new SingleStatRecorder(task, HISTOGRAM_BUCKETS);
}
}
}
public ImmutableList getTasksHistograms() {
return ImmutableList.copyOf(tasksHistograms);
}
public static Profiler instance() {
return instance;
}
/**
* Returns the nanoTime of the current profiler instance, or an arbitrary
* constant if not active.
*/
public static long nanoTimeMaybe() {
if (instance.isActive()) {
return instance.clock.nanoTime();
}
return -1;
}
/**
* Enable profiling.
*
* Subsequent calls to beginTask/endTask will be recorded
* in the provided output stream. Please note that stream performance is
* extremely important and buffered streams should be utilized.
*
* @param profiledTaskKinds which kinds of {@link ProfilerTask}s to track
* @param stream output stream to store profile data. Note: passing unbuffered stream object
* reference may result in significant performance penalties
* @param comment a comment to insert in the profile data
* @param recordAllDurations iff true, record all tasks regardless of their duration; otherwise
* some tasks may get aggregated if they finished quick enough
* @param clock a {@code BlazeClock.instance()}
* @param execStartTimeNanos execution start time in nanos obtained from {@code clock.nanoTime()}
*/
public synchronized void start(ProfiledTaskKinds profiledTaskKinds, OutputStream stream,
String comment, boolean recordAllDurations, Clock clock, long execStartTimeNanos)
throws IOException {
Preconditions.checkState(!isActive(), "Profiler already active");
taskStack = new TaskStack();
taskQueue = new ConcurrentLinkedQueue<>();
describer = new ObjectDescriber();
this.profiledTaskKinds = profiledTaskKinds;
this.clock = clock;
// sanity check for current limitation on the number of supported types due
// to using enum.ordinal() to store them instead of EnumSet for performance reasons.
Preconditions.checkState(TASK_COUNT < 256,
"The profiler implementation supports only up to 255 different ProfilerTask values.");
// reset state for the new profiling session
taskId.set(0);
this.recordAllDurations = recordAllDurations;
this.saveException = null;
if (stream != null) {
this.timer = new Timer("ProfilerTimer", true);
// Wrapping deflater stream in the buffered stream proved to reduce CPU consumption caused by
// the save() method. Values for buffer sizes were chosen by running small amount of tests
// and identifying point of diminishing returns - but I have not really tried to optimize
// them.
this.out = new DataOutputStream(new BufferedOutputStream(new DeflaterOutputStream(
stream, new Deflater(Deflater.BEST_SPEED, false), 65536), 262144));
this.out.writeInt(MAGIC); // magic
this.out.writeInt(VERSION); // protocol_version
this.out.writeUTF(comment);
// ProfileTask.values() method sorts enums using their ordinal() value, so
// there there is no need to store ordinal() value for each entry.
this.out.writeInt(TASK_COUNT);
for (ProfilerTask type : ProfilerTask.values()) {
this.out.writeUTF(type.toString());
}
// Start save thread
timer.schedule(new TimerTask() {
@Override public void run() { save(); }
}, SAVE_DELAY, SAVE_DELAY);
} else {
this.out = null;
}
// activate profiler
profileStartTime = execStartTimeNanos;
}
public synchronized Iterable getSlowestTasks() {
List> slowestTasksByType = new ArrayList<>();
for (SlowestTaskAggregator aggregator : slowestTasks) {
if (aggregator != null) {
slowestTasksByType.add(aggregator.getSlowestTasks());
}
}
return Iterables.concat(slowestTasksByType);
}
/**
* Disable profiling and complete profile file creation.
* Subsequent calls to beginTask/endTask will no longer
* be recorded in the profile.
*/
public synchronized void stop() throws IOException {
if (saveException != null) {
throw saveException;
}
if (!isActive()) {
return;
}
// Log a final event to update the duration of ProfilePhase.FINISH.
logEvent(ProfilerTask.INFO, "Finishing");
save();
clear();
for (SlowestTaskAggregator aggregator : slowestTasks) {
if (aggregator != null) {
aggregator.clear();
}
}
if (saveException != null) {
throw saveException;
}
if (out != null) {
out.writeInt(EOF_MARKER);
out.close();
out = null;
}
}
/**
* Returns true iff profiling is currently enabled.
*/
public boolean isActive() {
return profileStartTime != 0L;
}
public boolean isProfiling(ProfilerTask type) {
return profiledTaskKinds.isProfiling(type);
}
/**
* Saves all gathered information from taskQueue queue to the file.
* Method is invoked internally by the Timer-based thread and at the end of
* profiling session.
*/
private synchronized void save() {
if (out == null) {
return;
}
try {
// Allocate the sink once to avoid GC
ByteBuffer sink = ByteBuffer.allocate(1024);
while (!taskQueue.isEmpty()) {
sink.clear();
TaskData data = taskQueue.poll();
VarInt.putVarLong(data.threadId, sink);
VarInt.putVarInt(data.id, sink);
VarInt.putVarInt(data.parentId, sink);
VarInt.putVarLong(data.startTime - profileStartTime, sink);
VarInt.putVarLong(data.duration, sink);
// To save space (and improve performance), convert all description
// strings to the canonical object and use IdentityHashMap to assign
// unique numbers for each string.
int descIndex = describer.getDescriptionIndex(data.object);
VarInt.putVarInt(descIndex + 1, sink); // Add 1 to avoid encoding negative values.
// Save types using their ordinal() value
sink.put((byte) data.type.ordinal());
// Save aggregated data stats.
if (data.counts != null) {
for (int i = 0; i < TASK_COUNT; i++) {
if (data.counts[i] > 0) {
sink.put((byte) i); // aggregated type ordinal value
VarInt.putVarInt(data.counts[i], sink);
VarInt.putVarLong(data.durations[i], sink);
}
}
}
this.out.writeInt(sink.position());
this.out.write(sink.array(), 0, sink.position());
if (describer.isUnassigned(descIndex)) {
this.out.writeUTF(describer.getDescription(data.object));
}
}
this.out.flush();
} catch (IOException e) {
saveException = e;
clear();
try {
out.close();
} catch (IOException e2) {
// ignore it
}
}
}
private synchronized void clear() {
initHistograms();
profileStartTime = 0L;
if (timer != null) {
timer.cancel();
timer = null;
}
taskStack = null;
taskQueue = null;
describer = null;
// Note that slowest task aggregator are not cleared here because clearing happens
// periodically over the course of a command invocation.
}
/**
* Unless --record_full_profiler_data is given we drop small tasks and add their time to the
* parents duration.
*/
private boolean wasTaskSlowEnoughToRecord(ProfilerTask type, long duration) {
return (recordAllDurations || duration >= type.minDuration);
}
/**
* Adds task directly to the main queue bypassing task stack. Used for simple
* tasks that are known to not have any subtasks.
*
* @param startTime task start time (obtained through {@link Profiler#nanoTimeMaybe()})
* @param duration task duration
* @param type task type
* @param object object associated with that task. Can be String object that
* describes it.
*/
private void logTask(long startTime, long duration, ProfilerTask type, Object object) {
Preconditions.checkNotNull(object);
Preconditions.checkState(startTime > 0, "startTime was %s", startTime);
if (duration < 0) {
// See note in Clock#nanoTime, which is used by Profiler#nanoTimeMaybe.
duration = 0;
}
tasksHistograms[type.ordinal()].addStat((int) TimeUnit.NANOSECONDS.toMillis(duration), object);
// Store instance fields as local variables so they are not nulled out from under us by #clear.
TaskStack localStack = taskStack;
Queue localQueue = taskQueue;
if (localStack == null || localQueue == null) {
// Variables have been nulled out by #clear in between the check the caller made and this
// point in the code. Probably due to an asynchronous crash.
LOG.severe("Variables null in profiler for " + type + ", probably due to async crash");
return;
}
TaskData parent = localStack.peek();
if (parent != null) {
parent.aggregateChild(type, duration);
}
if (wasTaskSlowEnoughToRecord(type, duration)) {
TaskData data = localStack.create(startTime, type, object);
data.duration = duration;
if (out != null) {
localQueue.add(data);
}
SlowestTaskAggregator aggregator = slowestTasks[type.ordinal()];
if (aggregator != null) {
aggregator.add(data);
}
}
}
/**
* Used externally to submit simple task (one that does not have any subtasks).
* Depending on the minDuration attribute of the task type, task may be
* just aggregated into the parent task and not stored directly.
*
* @param startTime task start time (obtained through {@link
* Profiler#nanoTimeMaybe()})
* @param type task type
* @param object object associated with that task. Can be String object that
* describes it.
*/
public void logSimpleTask(long startTime, ProfilerTask type, Object object) {
if (isActive() && isProfiling(type)) {
logTask(startTime, clock.nanoTime() - startTime, type, object);
}
}
/**
* Used externally to submit simple task (one that does not have any
* subtasks). Depending on the minDuration attribute of the task type, task
* may be just aggregated into the parent task and not stored directly.
*
* Note that start and stop time must both be acquired from the same clock
* instance.
*
* @param startTime task start time
* @param stopTime task stop time
* @param type task type
* @param object object associated with that task. Can be String object that
* describes it.
*/
public void logSimpleTask(long startTime, long stopTime, ProfilerTask type, Object object) {
if (isActive() && isProfiling(type)) {
logTask(startTime, stopTime - startTime, type, object);
}
}
/**
* Used externally to submit simple task (one that does not have any
* subtasks). Depending on the minDuration attribute of the task type, task
* may be just aggregated into the parent task and not stored directly.
*
* @param startTime task start time (obtained through {@link
* Profiler#nanoTimeMaybe()})
* @param duration the duration of the task
* @param type task type
* @param object object associated with that task. Can be String object that
* describes it.
*/
public void logSimpleTaskDuration(long startTime, long duration, ProfilerTask type,
Object object) {
if (isActive() && isProfiling(type)) {
logTask(startTime, duration, type, object);
}
}
/**
* Used to log "events" - tasks with zero duration.
*/
public void logEvent(ProfilerTask type, Object object) {
if (isActive() && isProfiling(type)) {
logTask(clock.nanoTime(), 0, type, object);
}
}
/**
* Records the beginning of the task specified by the parameters. This method
* should always be followed by completeTask() invocation to mark the end of
* task execution (usually ensured by try {} finally {} block). Failure to do
* so will result in task stack corruption.
*
* Use of this method allows to support nested task monitoring. For tasks that
* are known to not have any subtasks, logSimpleTask() should be used instead.
*
* @param type predefined task type - see ProfilerTask for available types.
* @param object object associated with that task. Can be String object that
* describes it.
*/
public void startTask(ProfilerTask type, Object object) {
// ProfilerInfo.allTasksById is supposed to be an id -> Task map, but it is in fact a List,
// which means that we cannot drop tasks to which we had already assigned ids. Therefore,
// non-leaf tasks must not have a minimum duration. However, we don't quite consistently
// enforce this, and Blaze only works because we happen not to add child tasks to those parent
// tasks that have a minimum duration.
Preconditions.checkNotNull(object);
if (isActive() && isProfiling(type)) {
taskStack.push(type, object);
}
}
/**
* Records the end of the task and moves tasks from the thread-local stack to
* the main queue. Will validate that given task type matches task at the top
* of the stack.
*
* @param type task type.
*/
public void completeTask(ProfilerTask type) {
if (isActive() && isProfiling(type)) {
long endTime = clock.nanoTime();
TaskData data = taskStack.pop();
Preconditions.checkState(
data.type == type,
"Inconsistent Profiler.completeTask() call: should have been %s but got %s (%s, %s)",
data.type,
type,
data,
taskStack);
data.duration = endTime - data.startTime;
if (data.parentId > 0) {
taskStack.peek().aggregateChild(data.type, data.duration);
}
boolean shouldRecordTask = wasTaskSlowEnoughToRecord(type, data.duration);
if (out != null && (shouldRecordTask || data.counts != null)) {
taskQueue.add(data);
}
if (shouldRecordTask) {
SlowestTaskAggregator aggregator = slowestTasks[type.ordinal()];
if (aggregator != null) {
aggregator.add(data);
}
}
}
}
/**
* Convenience method to log phase marker tasks.
*/
public void markPhase(ProfilePhase phase) {
MemoryProfiler.instance().markPhase(phase);
if (isActive() && isProfiling(ProfilerTask.PHASE)) {
Preconditions.checkState(taskStack.isEmpty(), "Phase tasks must not be nested");
logEvent(ProfilerTask.PHASE, phase.description);
}
}
/**
* Convenience method to log spawn tasks.
*
* TODO(bazel-team): Right now method expects single string of the spawn action
* as task description (usually either argv[0] or a name of the main executable
* in case of complex shell commands). Maybe it should accept Command object
* and create more user friendly description.
*/
public void logSpawn(long startTime, String arg0) {
if (isActive() && isProfiling(ProfilerTask.SPAWN)) {
logTask(startTime, clock.nanoTime() - startTime, ProfilerTask.SPAWN, arg0);
}
}
}