/root/doris/be/src/util/runtime_profile.h
Line | Count | Source |
1 | | // Licensed to the Apache Software Foundation (ASF) under one |
2 | | // or more contributor license agreements. See the NOTICE file |
3 | | // distributed with this work for additional information |
4 | | // regarding copyright ownership. The ASF licenses this file |
5 | | // to you under the Apache License, Version 2.0 (the |
6 | | // "License"); you may not use this file except in compliance |
7 | | // with the License. You may obtain a copy of the License at |
8 | | // |
9 | | // http://www.apache.org/licenses/LICENSE-2.0 |
10 | | // |
11 | | // Unless required by applicable law or agreed to in writing, |
12 | | // software distributed under the License is distributed on an |
13 | | // "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY |
14 | | // KIND, either express or implied. See the License for the |
15 | | // specific language governing permissions and limitations |
16 | | // under the License. |
17 | | // This file is copied from |
18 | | // https://github.com/apache/impala/blob/branch-2.9.0/be/src/util/runtime-profile.h |
19 | | // and modified by Doris |
20 | | |
21 | | #pragma once |
22 | | |
23 | | #include <gen_cpp/Metrics_types.h> |
24 | | #include <gen_cpp/RuntimeProfile_types.h> |
25 | | #include <gen_cpp/runtime_profile.pb.h> |
26 | | #include <glog/logging.h> |
27 | | #include <stdint.h> |
28 | | |
29 | | #include <algorithm> |
30 | | #include <atomic> |
31 | | #include <cstdint> |
32 | | #include <functional> |
33 | | #include <iostream> |
34 | | #include <map> |
35 | | #include <memory> |
36 | | #include <mutex> |
37 | | #include <set> |
38 | | #include <string> |
39 | | #include <utility> |
40 | | #include <vector> |
41 | | |
42 | | #include "common/compiler_util.h" // IWYU pragma: keep |
43 | | #include "common/logging.h" |
44 | | #include "util/binary_cast.hpp" |
45 | | #include "util/container_util.hpp" |
46 | | #include "util/pretty_printer.h" |
47 | | #include "util/stopwatch.hpp" |
48 | | |
49 | | namespace doris { |
50 | | class TRuntimeProfileNode; |
51 | | class TRuntimeProfileTree; |
52 | | class RuntimeProfileCounterTreeNode; |
53 | | |
54 | | // Some macro magic to generate unique ids using __COUNTER__ |
55 | 2.21M | #define CONCAT_IMPL(x, y) x##y |
56 | 2.21M | #define MACRO_CONCAT(x, y) CONCAT_IMPL(x, y) |
57 | | |
58 | | #define ADD_LABEL_COUNTER(profile, name) (profile)->add_counter(name, TUnit::NONE) |
59 | | #define ADD_LABEL_COUNTER_WITH_LEVEL(profile, name, level) \ |
60 | | (profile)->add_counter_with_level(name, TUnit::NONE, level) |
61 | 5.42M | #define ADD_COUNTER(profile, name, type) (profile)->add_counter(name, type) |
62 | | #define ADD_COUNTER_WITH_LEVEL(profile, name, type, level) \ |
63 | 419k | (profile)->add_counter_with_level(name, type, level) |
64 | 628k | #define ADD_TIMER(profile, name) (profile)->add_counter(name, TUnit::TIME_NS) |
65 | | #define ADD_TIMER_WITH_LEVEL(profile, name, level) \ |
66 | 550k | (profile)->add_counter_with_level(name, TUnit::TIME_NS, level) |
67 | 0 | #define ADD_CHILD_COUNTER(profile, name, type, parent) (profile)->add_counter(name, type, parent) |
68 | | #define ADD_CHILD_COUNTER_WITH_LEVEL(profile, name, type, parent, level) \ |
69 | 357 | (profile)->add_counter(name, type, parent, level) |
70 | 682 | #define ADD_CHILD_TIMER(profile, name, parent) (profile)->add_counter(name, TUnit::TIME_NS, parent) |
71 | | #define ADD_CHILD_TIMER_WITH_LEVEL(profile, name, parent, level) \ |
72 | 437 | (profile)->add_counter(name, TUnit::TIME_NS, parent, level) |
73 | 1.39M | #define SCOPED_TIMER(c) ScopedTimer<MonotonicStopWatch> MACRO_CONCAT(SCOPED_TIMER, __COUNTER__)(c) |
74 | | #define SCOPED_TIMER_ATOMIC(c) \ |
75 | | ScopedTimer<MonotonicStopWatch, std::atomic_bool> MACRO_CONCAT(SCOPED_TIMER, __COUNTER__)(c) |
76 | | #define SCOPED_CPU_TIMER(c) \ |
77 | 87 | ScopedTimer<ThreadCpuStopWatch> MACRO_CONCAT(SCOPED_TIMER, __COUNTER__)(c) |
78 | | #define CANCEL_SAFE_SCOPED_TIMER(c, is_cancelled) \ |
79 | | ScopedTimer<MonotonicStopWatch> MACRO_CONCAT(SCOPED_TIMER, __COUNTER__)(c, is_cancelled) |
80 | | #define SCOPED_RAW_TIMER(c) \ |
81 | 820k | doris::ScopedRawTimer<doris::MonotonicStopWatch, int64_t> MACRO_CONCAT(SCOPED_RAW_TIMER, \ |
82 | 820k | __COUNTER__)(c) |
83 | | #define SCOPED_ATOMIC_TIMER(c) \ |
84 | 0 | ScopedRawTimer<MonotonicStopWatch, std::atomic<int64_t>> MACRO_CONCAT(SCOPED_ATOMIC_TIMER, \ |
85 | 0 | __COUNTER__)(c) |
86 | 133k | #define COUNTER_UPDATE(c, v) (c)->update(v) |
87 | 357k | #define COUNTER_SET(c, v) (c)->set(v) |
88 | | |
89 | | class ObjectPool; |
90 | | |
91 | | // Runtime profile is a group of profiling counters. It supports adding named counters |
92 | | // and being able to serialize and deserialize them. |
93 | | // The profiles support a tree structure to form a hierarchy of counters. |
94 | | // Runtime profiles supports measuring wall clock rate based counters. There is a |
95 | | // single thread per process that will convert an amount (i.e. bytes) counter to a |
96 | | // corresponding rate based counter. This thread wakes up at fixed intervals and updates |
97 | | // all of the rate counters. |
98 | | // Thread-safe. |
99 | | class RuntimeProfile { |
100 | | public: |
101 | | static std::unique_ptr<RuntimeProfile> from_thrift(const TRuntimeProfileTree& node); |
102 | | |
103 | | static std::unique_ptr<RuntimeProfile> from_proto(const PRuntimeProfileTree& tree); |
104 | | |
105 | 14 | static PProfileUnit unit_to_proto(const TUnit::type& type) { |
106 | 14 | switch (type) { |
107 | 11 | case TUnit::UNIT: { |
108 | 11 | return PProfileUnit::UNIT; |
109 | 0 | } |
110 | 0 | case TUnit::UNIT_PER_SECOND: { |
111 | 0 | return PProfileUnit::UNIT_PER_SECOND; |
112 | 0 | } |
113 | 0 | case TUnit::CPU_TICKS: { |
114 | 0 | return PProfileUnit::CPU_TICKS; |
115 | 0 | } |
116 | 3 | case TUnit::BYTES: { |
117 | 3 | return PProfileUnit::BYTES; |
118 | 0 | } |
119 | 0 | case TUnit::BYTES_PER_SECOND: { |
120 | 0 | return PProfileUnit::BYTES_PER_SECOND; |
121 | 0 | } |
122 | 0 | case TUnit::TIME_NS: { |
123 | 0 | return PProfileUnit::TIME_NS; |
124 | 0 | } |
125 | 0 | case TUnit::DOUBLE_VALUE: { |
126 | 0 | return PProfileUnit::DOUBLE_VALUE; |
127 | 0 | } |
128 | 0 | case TUnit::NONE: { |
129 | 0 | return PProfileUnit::NONE; |
130 | 0 | } |
131 | 0 | case TUnit::TIME_MS: { |
132 | 0 | return PProfileUnit::TIME_MS; |
133 | 0 | } |
134 | 0 | case TUnit::TIME_S: { |
135 | 0 | return PProfileUnit::TIME_S; |
136 | 0 | } |
137 | 0 | default: { |
138 | 0 | DCHECK(false); |
139 | 0 | return PProfileUnit::NONE; |
140 | 0 | } |
141 | 14 | } |
142 | 14 | } |
143 | | |
144 | 14 | static TUnit::type unit_to_thrift(const PProfileUnit& unit) { |
145 | 14 | switch (unit) { |
146 | 12 | case PProfileUnit::UNIT: { |
147 | 12 | return TUnit::UNIT; |
148 | 0 | } |
149 | 0 | case PProfileUnit::UNIT_PER_SECOND: { |
150 | 0 | return TUnit::UNIT_PER_SECOND; |
151 | 0 | } |
152 | 0 | case PProfileUnit::CPU_TICKS: { |
153 | 0 | return TUnit::CPU_TICKS; |
154 | 0 | } |
155 | 2 | case PProfileUnit::BYTES: { |
156 | 2 | return TUnit::BYTES; |
157 | 0 | } |
158 | 0 | case PProfileUnit::BYTES_PER_SECOND: { |
159 | 0 | return TUnit::BYTES_PER_SECOND; |
160 | 0 | } |
161 | 0 | case PProfileUnit::TIME_NS: { |
162 | 0 | return TUnit::TIME_NS; |
163 | 0 | } |
164 | 0 | case PProfileUnit::DOUBLE_VALUE: { |
165 | 0 | return TUnit::DOUBLE_VALUE; |
166 | 0 | } |
167 | 0 | case PProfileUnit::NONE: { |
168 | 0 | return TUnit::NONE; |
169 | 0 | } |
170 | 0 | case PProfileUnit::TIME_MS: { |
171 | 0 | return TUnit::TIME_MS; |
172 | 0 | } |
173 | 0 | case PProfileUnit::TIME_S: { |
174 | 0 | return TUnit::TIME_S; |
175 | 0 | } |
176 | 0 | default: { |
177 | 0 | DCHECK(false); |
178 | 0 | return TUnit::NONE; |
179 | 0 | } |
180 | 14 | } |
181 | 14 | } |
182 | | |
183 | | // The root counter name for all top level counters. |
184 | | static const std::string ROOT_COUNTER; |
185 | | class Counter { |
186 | | public: |
187 | | Counter(TUnit::type type, int64_t value = 0, int64_t level = 3) |
188 | 9.88M | : _value(value), _type(type), _level(level) {} |
189 | 9.88M | virtual ~Counter() = default; |
190 | | |
191 | 8 | virtual Counter* clone() const { return new Counter(type(), value(), _level); } |
192 | | |
193 | 1.81M | virtual void update(int64_t delta) { _value.fetch_add(delta, std::memory_order_relaxed); } |
194 | | |
195 | 0 | void bit_or(int64_t delta) { _value.fetch_or(delta, std::memory_order_relaxed); } |
196 | | |
197 | 409k | virtual void set(int64_t value) { _value.store(value, std::memory_order_relaxed); } |
198 | | |
199 | 0 | virtual void set(double value) { |
200 | 0 | DCHECK_EQ(sizeof(value), sizeof(int64_t)); |
201 | 0 | _value.store(binary_cast<double, int64_t>(value), std::memory_order_relaxed); |
202 | 0 | } |
203 | | |
204 | 382k | virtual int64_t value() const { return _value.load(std::memory_order_relaxed); } |
205 | | |
206 | 0 | virtual double double_value() const { |
207 | 0 | return binary_cast<int64_t, double>(_value.load(std::memory_order_relaxed)); |
208 | 0 | } |
209 | | |
210 | 42.1k | virtual TCounter to_thrift(const std::string& name) const { |
211 | 42.1k | TCounter counter; |
212 | 42.1k | counter.name = name; |
213 | 42.1k | counter.value = this->value(); |
214 | 42.1k | counter.type = this->type(); |
215 | 42.1k | counter.__set_level(this->_level); |
216 | 42.1k | return counter; |
217 | 42.1k | } |
218 | | |
219 | 8 | virtual PProfileCounter to_proto(const std::string& name) const { |
220 | 8 | PProfileCounter counter; |
221 | 8 | counter.set_name(name); |
222 | 8 | counter.set_value(this->value()); |
223 | 8 | counter.set_type(unit_to_proto(this->type())); |
224 | 8 | counter.set_level(this->value()); |
225 | 8 | return counter; |
226 | 8 | } |
227 | | |
228 | | virtual void pretty_print(std::ostream* s, const std::string& prefix, |
229 | 369 | const std::string& name) const { |
230 | 369 | std::ostream& stream = *s; |
231 | 369 | stream << prefix << " - " << name << ": " |
232 | 369 | << PrettyPrinter::print(_value.load(std::memory_order_relaxed), type()) |
233 | 369 | << std::endl; |
234 | 369 | } |
235 | | |
236 | 1.44M | TUnit::type type() const { return _type; } |
237 | | |
238 | 42.3k | virtual int64_t level() const { return _level; } |
239 | | |
240 | 3 | void set_level(int64_t level) { _level = level; } |
241 | | |
242 | | bool operator==(const Counter& other) const; |
243 | | |
244 | | private: |
245 | | friend class RuntimeProfile; |
246 | | friend class RuntimeProfileCounterTreeNode; |
247 | | |
248 | | std::atomic<int64_t> _value; |
249 | | TUnit::type _type; |
250 | | int64_t _level; |
251 | | }; |
252 | | |
253 | | /// A counter that keeps track of the highest value seen (reporting that |
254 | | /// as value()) and the current value. |
255 | | class HighWaterMarkCounter : public Counter { |
256 | | public: |
257 | | HighWaterMarkCounter(TUnit::type unit, int64_t level, const std::string& parent_name, |
258 | | int64_t value = 0, int64_t current_value = 0) |
259 | 130k | : Counter(unit, value, level), |
260 | 130k | current_value_(current_value), |
261 | 130k | _parent_name(parent_name) {} |
262 | | |
263 | 0 | virtual Counter* clone() const override { |
264 | 0 | return new HighWaterMarkCounter(type(), level(), parent_name(), value(), |
265 | 0 | current_value()); |
266 | 0 | } |
267 | | |
268 | 2.68k | void add(int64_t delta) { |
269 | 2.68k | current_value_.fetch_add(delta, std::memory_order_relaxed); |
270 | 2.68k | if (delta > 0) { |
271 | 1.35k | UpdateMax(current_value_); |
272 | 1.35k | } |
273 | 2.68k | } |
274 | 2.64k | virtual void update(int64_t delta) override { add(delta); } |
275 | | |
276 | 2 | TCounter to_thrift(const std::string& name) const override { |
277 | 2 | TCounter counter; |
278 | 2 | counter.name = std::move(name); |
279 | 2 | counter.value = current_value(); |
280 | 2 | counter.type = type(); |
281 | 2 | counter.__set_level(level()); |
282 | 2 | return counter; |
283 | 2 | } |
284 | | |
285 | 0 | PProfileCounter to_proto(const std::string& name) const override { |
286 | 0 | PProfileCounter counter; |
287 | 0 | counter.set_name(name); |
288 | 0 | counter.set_value(current_value()); |
289 | 0 | counter.set_type(unit_to_proto(this->type())); |
290 | 0 | counter.set_level(this->value()); |
291 | 0 | return counter; |
292 | 0 | } |
293 | | |
294 | 2 | TCounter to_thrift_peak(std::string name) { |
295 | 2 | TCounter counter; |
296 | 2 | counter.name = std::move(name); |
297 | 2 | counter.value = value(); |
298 | 2 | counter.type = type(); |
299 | 2 | counter.__set_level(level()); |
300 | 2 | return counter; |
301 | 2 | } |
302 | | |
303 | 0 | PProfileCounter to_proto_peak(const std::string& name) const { |
304 | 0 | return Counter::to_proto(name); |
305 | 0 | } |
306 | | |
307 | | virtual void pretty_print(std::ostream* s, const std::string& prefix, |
308 | 39 | const std::string& name) const override { |
309 | 39 | std::ostream& stream = *s; |
310 | 39 | stream << prefix << " - " << name |
311 | 39 | << " Current: " << PrettyPrinter::print(current_value(), type()) << " (Peak: " |
312 | 39 | << PrettyPrinter::print(_value.load(std::memory_order_relaxed), type()) << ")" |
313 | 39 | << std::endl; |
314 | 39 | } |
315 | | |
316 | | /// Tries to increase the current value by delta. If current_value() + delta |
317 | | /// exceeds max, return false and current_value is not changed. |
318 | 0 | bool try_add(int64_t delta, int64_t max) { |
319 | 0 | while (true) { |
320 | 0 | int64_t old_val = current_value_.load(std::memory_order_relaxed); |
321 | 0 | int64_t new_val = old_val + delta; |
322 | 0 | if (UNLIKELY(new_val > max)) return false; |
323 | 0 | if (LIKELY(current_value_.compare_exchange_weak(old_val, new_val, |
324 | 0 | std::memory_order_relaxed))) { |
325 | 0 | UpdateMax(new_val); |
326 | 0 | return true; |
327 | 0 | } |
328 | 0 | } |
329 | 0 | } |
330 | | |
331 | 104k | void set(int64_t v) override { |
332 | 104k | current_value_.store(v, std::memory_order_relaxed); |
333 | 104k | UpdateMax(v); |
334 | 104k | } |
335 | | |
336 | 41 | int64_t current_value() const { return current_value_.load(std::memory_order_relaxed); } |
337 | | |
338 | 2 | std::string parent_name() const { return _parent_name; } |
339 | | |
340 | | private: |
341 | | /// Set '_value' to 'v' if 'v' is larger than '_value'. The entire operation is |
342 | | /// atomic. |
343 | 105k | void UpdateMax(int64_t v) { |
344 | 105k | while (true) { |
345 | 105k | int64_t old_max = _value.load(std::memory_order_relaxed); |
346 | 105k | int64_t new_max = std::max(old_max, v); |
347 | 105k | if (new_max == old_max) { |
348 | 1.27k | break; // Avoid atomic update. |
349 | 1.27k | } |
350 | 104k | if (LIKELY(_value.compare_exchange_weak(old_max, new_max, |
351 | 104k | std::memory_order_relaxed))) { |
352 | 104k | break; |
353 | 104k | } |
354 | 104k | } |
355 | 105k | } |
356 | | |
357 | | /// The current value of the counter. _value in the super class represents |
358 | | /// the high water mark. |
359 | | std::atomic<int64_t> current_value_; |
360 | | |
361 | | const std::string _parent_name; |
362 | | }; |
363 | | |
364 | | using DerivedCounterFunction = std::function<int64_t()>; |
365 | | |
366 | | // A DerivedCounter also has a name and type, but the value is computed. |
367 | | // Do not call Set() and Update(). |
368 | | class DerivedCounter : public Counter { |
369 | | public: |
370 | | DerivedCounter(TUnit::type type, const DerivedCounterFunction& counter_fn, |
371 | | int64_t value = 0, int64_t level = 1) |
372 | 8 | : Counter(type, value, level), _counter_fn(counter_fn) {} |
373 | | |
374 | 0 | virtual Counter* clone() const override { |
375 | 0 | return new DerivedCounter(type(), _counter_fn, value(), level()); |
376 | 0 | } |
377 | | |
378 | 3 | int64_t value() const override { return _counter_fn(); } |
379 | | |
380 | | private: |
381 | | DerivedCounterFunction _counter_fn; |
382 | | }; |
383 | | |
384 | | using ConditionCounterFunction = std::function<bool(int64_t, int64_t)>; |
385 | | |
386 | | // ConditionCounter is a specialized counter that only updates its value when a specific condition is met. |
387 | | // It uses a condition function (condition_func) to determine when the counter's value should be updated. |
388 | | // This type of counter is particularly useful for tracking maximum values, minimum values, or other metrics |
389 | | // that should only be updated when they meet certain criteria. |
390 | | // For example, it can be used to record the maximum value of a specific metric during query execution, |
391 | | // or to update the counter only when a new value exceeds some threshold. |
392 | | class ConditionCounter : public Counter { |
393 | | public: |
394 | | ConditionCounter(TUnit::type type, const ConditionCounterFunction& condition_func, |
395 | | int64_t level = 2, int64_t condition = 0, int64_t value = 0) |
396 | 2 | : Counter(type, value, level), |
397 | 2 | _condition(condition), |
398 | 2 | _value(value), |
399 | 2 | _condition_func(condition_func) {} |
400 | | |
401 | 0 | Counter* clone() const override { |
402 | 0 | std::lock_guard<std::mutex> l(_mutex); |
403 | 0 | return new ConditionCounter(type(), _condition_func, _condition, value(), level()); |
404 | 0 | } |
405 | | |
406 | 7 | int64_t value() const override { |
407 | 7 | std::lock_guard<std::mutex> l(_mutex); |
408 | 7 | return _value; |
409 | 7 | } |
410 | | |
411 | 7 | void conditional_update(int64_t c, int64_t v) { |
412 | 7 | std::lock_guard<std::mutex> l(_mutex); |
413 | 7 | if (_condition_func(_condition, c)) { |
414 | 4 | _value = v; |
415 | 4 | _condition = c; |
416 | 4 | } |
417 | 7 | } |
418 | | |
419 | | private: |
420 | | mutable std::mutex _mutex; |
421 | | int64_t _condition; |
422 | | int64_t _value; |
423 | | ConditionCounterFunction _condition_func; |
424 | | }; |
425 | | |
426 | | // NonZeroCounter will not be converted to Thrift if the value is 0. |
427 | | class NonZeroCounter : public Counter { |
428 | | public: |
429 | | NonZeroCounter(TUnit::type type, int64_t level, const std::string& parent_name, |
430 | | int64_t value = 0) |
431 | 34 | : Counter(type, value, level), _parent_name(parent_name) {} |
432 | | |
433 | 0 | virtual Counter* clone() const override { |
434 | 0 | return new NonZeroCounter(type(), level(), parent_name(), value()); |
435 | 0 | } |
436 | | |
437 | 1 | std::string parent_name() const { return _parent_name; } |
438 | | |
439 | | private: |
440 | | const std::string _parent_name; |
441 | | }; |
442 | | |
443 | | class DescriptionEntry : public Counter { |
444 | | public: |
445 | | DescriptionEntry(const std::string& name, const std::string& description) |
446 | 78.0k | : Counter(TUnit::NONE, 0, 2), _description(description), _name(name) {} |
447 | | |
448 | 0 | virtual Counter* clone() const override { |
449 | 0 | return new DescriptionEntry(_name, _description); |
450 | 0 | } |
451 | | |
452 | 0 | void set(int64_t value) override { |
453 | | // Do nothing |
454 | 0 | } |
455 | 0 | void set(double value) override { |
456 | | // Do nothing |
457 | 0 | } |
458 | 0 | void update(int64_t delta) override { |
459 | | // Do nothing |
460 | 0 | } |
461 | | |
462 | 1 | TCounter to_thrift(const std::string& name) const override { |
463 | 1 | TCounter counter; |
464 | 1 | counter.name = name; |
465 | 1 | counter.__set_level(2); |
466 | 1 | counter.__set_description(_description); |
467 | 1 | return counter; |
468 | 1 | } |
469 | | |
470 | 0 | PProfileCounter to_proto(const std::string& name) const override { |
471 | 0 | PProfileCounter counter; |
472 | 0 | counter.set_name(name); |
473 | 0 | counter.set_level(2); |
474 | 0 | counter.set_description(_description); |
475 | 0 | return counter; |
476 | 0 | } |
477 | | |
478 | | private: |
479 | | const std::string _description; |
480 | | const std::string _name; |
481 | | }; |
482 | | |
483 | | // Create a runtime profile object with 'name'. |
484 | | RuntimeProfile(const std::string& name, bool is_averaged_profile = false); |
485 | | |
486 | | ~RuntimeProfile(); |
487 | | |
488 | | // Adds a child profile. This is thread safe. |
489 | | // 'indent' indicates whether the child will be printed w/ extra indentation |
490 | | // relative to the parent. |
491 | | // If location is non-null, child will be inserted after location. Location must |
492 | | // already be added to the profile. |
493 | | void add_child(RuntimeProfile* child, bool indent, RuntimeProfile* location = nullptr); |
494 | | |
495 | | void add_child_unlock(RuntimeProfile* child, bool indent, RuntimeProfile* loc); |
496 | | |
497 | | /// Creates a new child profile with the given 'name'. A child profile with that name |
498 | | /// must not already exist. If 'prepend' is true, prepended before other child profiles, |
499 | | /// otherwise appended after other child profiles. |
500 | | RuntimeProfile* create_child(const std::string& name, bool indent = true, bool prepend = false); |
501 | | |
502 | | // Merges the src profile into this one, combining counters that have an identical |
503 | | // path. Info strings from profiles are not merged. 'src' would be a const if it |
504 | | // weren't for locking. |
505 | | // Calling this concurrently on two RuntimeProfiles in reverse order results in |
506 | | // undefined behavior. |
507 | | void merge(RuntimeProfile* src); |
508 | | |
509 | | // Updates this profile w/ the thrift profile: behaves like Merge(), except |
510 | | // that existing counters are updated rather than added up. |
511 | | // Info strings matched up by key and are updated or added, depending on whether |
512 | | // the key has already been registered. |
513 | | void update(const TRuntimeProfileTree& thrift_profile); |
514 | | |
515 | | //Similar to `void update(const TRuntimeProfileTree& thrift_profile)` |
516 | | void update(const PRuntimeProfileTree& proto_profile); |
517 | | |
518 | | // Add a counter with 'name'/'type'. Returns a counter object that the caller can |
519 | | // update. The counter is owned by the RuntimeProfile object. |
520 | | // If parent_counter_name is a non-empty string, the counter is added as a child of |
521 | | // parent_counter_name. |
522 | | // If the counter already exists, the existing counter object is returned. |
523 | | Counter* add_counter(const std::string& name, TUnit::type type, |
524 | | const std::string& parent_counter_name, int64_t level = 2); |
525 | | |
526 | 6.05M | Counter* add_counter(const std::string& name, TUnit::type type) { |
527 | 6.05M | return add_counter(name, type, RuntimeProfile::ROOT_COUNTER); |
528 | 6.05M | } |
529 | | |
530 | 1.04M | Counter* add_counter_with_level(const std::string& name, TUnit::type type, int64_t level) { |
531 | 1.04M | return add_counter(name, type, RuntimeProfile::ROOT_COUNTER, level); |
532 | 1.04M | } |
533 | | |
534 | | NonZeroCounter* add_nonzero_counter( |
535 | | const std::string& name, TUnit::type type, |
536 | | const std::string& parent_counter_name = RuntimeProfile::ROOT_COUNTER, |
537 | | int64_t level = 2); |
538 | | |
539 | | // Add a description entry under target counter. |
540 | | void add_description(const std::string& name, const std::string& description, |
541 | | std::string parent_counter_name); |
542 | | // Add a derived counter with 'name'/'type'. The counter is owned by the |
543 | | // RuntimeProfile object. |
544 | | // If parent_counter_name is a non-empty string, the counter is added as a child of |
545 | | // parent_counter_name. |
546 | | // Returns nullptr if the counter already exists. |
547 | | DerivedCounter* add_derived_counter(const std::string& name, TUnit::type type, |
548 | | const DerivedCounterFunction& counter_fn, |
549 | | const std::string& parent_counter_name); |
550 | | |
551 | | ConditionCounter* add_conditition_counter(const std::string& name, TUnit::type type, |
552 | | const ConditionCounterFunction& counter_fn, |
553 | | const std::string& parent_counter_name, |
554 | | int64_t level = 2); |
555 | | |
556 | | // Gets the counter object with 'name'. Returns nullptr if there is no counter with |
557 | | // that name. |
558 | | Counter* get_counter(const std::string& name); |
559 | | |
560 | | // Adds all counters with 'name' that are registered either in this or |
561 | | // in any of the child profiles to 'counters'. |
562 | | void get_counters(const std::string& name, std::vector<Counter*>* counters); |
563 | | |
564 | | // Helper to append to the "ExecOption" info string. |
565 | 0 | void append_exec_option(const std::string& option) { add_info_string("ExecOption", option); } |
566 | | |
567 | | // Adds a string to the runtime profile. If a value already exists for 'key', |
568 | | // the value will be updated. |
569 | | void add_info_string(const std::string& key, const std::string& value); |
570 | | |
571 | | // Returns a pointer to the info string value for 'key'. Returns nullptr if |
572 | | // the key does not exist. |
573 | | const std::string* get_info_string(const std::string& key); |
574 | | |
575 | | // Returns the counter for the total elapsed time. |
576 | 133 | Counter* total_time_counter() { return &_counter_total_time; } |
577 | | |
578 | | // Prints the counters in a name: value format. |
579 | | // Does not hold locks when it makes any function calls. |
580 | | void pretty_print(std::ostream* s, const std::string& prefix = "", |
581 | | int64_t profile_level = 2) const; |
582 | 17 | std::string pretty_print() const { |
583 | 17 | std::stringstream ss; |
584 | 17 | pretty_print(&ss); |
585 | 17 | return ss.str(); |
586 | 17 | }; |
587 | | |
588 | | // Serializes profile to thrift. |
589 | | // Does not hold locks when it makes any function calls. |
590 | | void to_thrift(TRuntimeProfileTree* tree, int64_t profile_level = 2); |
591 | | void to_thrift(std::vector<TRuntimeProfileNode>* nodes, int64_t profile_level = 2); |
592 | | |
593 | | // Similar to `to_thrift`. |
594 | | void to_proto(PRuntimeProfileTree* tree, int64_t profile_level = 2); |
595 | | void to_proto(google::protobuf::RepeatedPtrField<PRuntimeProfileNode>* nodes, |
596 | | int64_t profile_level = 2); |
597 | | |
598 | | // Divides all counters by n |
599 | | void divide(int n); |
600 | | |
601 | | RuntimeProfile* get_child(std::string name); |
602 | | |
603 | | void get_children(std::vector<RuntimeProfile*>* children); |
604 | | |
605 | | // Gets all profiles in tree, including this one. |
606 | | void get_all_children(std::vector<RuntimeProfile*>* children); |
607 | | |
608 | | // Returns the number of counters in this profile |
609 | 16 | int num_counters() const { return _counter_map.size(); } |
610 | | |
611 | | // Returns name of this profile |
612 | 24 | const std::string& name() const { return _name; } |
613 | | |
614 | | // *only call this on top-level profiles* |
615 | | // (because it doesn't re-file child profiles) |
616 | 0 | void set_name(const std::string& name) { _name = name; } |
617 | | |
618 | 0 | int64_t metadata() const { return _metadata; } |
619 | 104k | void set_metadata(int64_t md) { |
620 | 104k | _is_set_metadata = true; |
621 | 104k | _metadata = md; |
622 | 104k | } |
623 | | |
624 | 118 | bool is_set_metadata() const { return _is_set_metadata; } |
625 | | |
626 | 0 | time_t timestamp() const { return _timestamp; } |
627 | 0 | void set_timestamp(time_t ss) { _timestamp = ss; } |
628 | | |
629 | | // Derived counter function: return measured throughput as input_value/second. |
630 | | static int64_t units_per_second(const Counter* total_counter, const Counter* timer); |
631 | | |
632 | | // Derived counter function: return aggregated value |
633 | | static int64_t counter_sum(const std::vector<Counter*>* counters); |
634 | | |
635 | | /// Adds a high water mark counter to the runtime profile. Otherwise, same behavior |
636 | | /// as AddCounter(). |
637 | | HighWaterMarkCounter* AddHighWaterMarkCounter( |
638 | | const std::string& name, TUnit::type unit, |
639 | | const std::string& parent_counter_name = RuntimeProfile::ROOT_COUNTER, |
640 | | int64_t profile_level = 2); |
641 | | |
642 | | // Recursively compute the fraction of the 'total_time' spent in this profile and |
643 | | // its children. |
644 | | // This function updates _local_time_percent for each profile. |
645 | | void compute_time_in_profile(); |
646 | | |
647 | | void clear_children(); |
648 | | |
649 | | private: |
650 | | // RuntimeProfileCounterTreeNode needs to access the counter map and child counter map |
651 | | friend class RuntimeProfileCounterTreeNode; |
652 | | // Pool for allocated counters. Usually owned by the creator of this |
653 | | // object, but occasionally allocated in the constructor. |
654 | | std::unique_ptr<ObjectPool> _pool; |
655 | | |
656 | | // Pool for allocated counters. These counters are shared with some other objects. |
657 | | std::map<std::string, std::shared_ptr<HighWaterMarkCounter>> _shared_counter_pool; |
658 | | |
659 | | // Name for this runtime profile. |
660 | | std::string _name; |
661 | | |
662 | | // user-supplied, uninterpreted metadata. |
663 | | int64_t _metadata; |
664 | | bool _is_set_metadata = false; |
665 | | |
666 | | bool _is_sink = false; |
667 | | bool _is_set_sink = false; |
668 | | |
669 | | // The timestamp when the profile was modified, make sure the update is up to date. |
670 | | time_t _timestamp; |
671 | | |
672 | | /// True if this profile is an average derived from other profiles. |
673 | | /// All counters in this profile must be of unit AveragedCounter. |
674 | | bool _is_averaged_profile; |
675 | | |
676 | | // Map from counter names to counters. The profile owns the memory for the |
677 | | // counters. |
678 | | using CounterMap = std::map<std::string, Counter*>; |
679 | | CounterMap _counter_map; |
680 | | |
681 | | // Map from parent counter name to a set of child counter name. |
682 | | // All top level counters are the child of RuntimeProfile::ROOT_COUNTER (root). |
683 | | using ChildCounterMap = std::map<std::string, std::set<std::string>>; |
684 | | ChildCounterMap _child_counter_map; |
685 | | |
686 | | // protects _counter_map, _counter_child_map and _bucketing_counters |
687 | | mutable std::mutex _counter_map_lock; |
688 | | |
689 | | // Child profiles. Does not own memory. |
690 | | // We record children in both a map (to facilitate updates) and a vector |
691 | | // (to print things in the order they were registered) |
692 | | using ChildMap = std::map<std::string, RuntimeProfile*>; |
693 | | ChildMap _child_map; |
694 | | // vector of (profile, indentation flag) |
695 | | using ChildVector = std::vector<std::pair<RuntimeProfile*, bool>>; |
696 | | ChildVector _children; |
697 | | mutable std::mutex _children_lock; // protects _child_map and _children |
698 | | |
699 | | using InfoStrings = std::map<std::string, std::string>; |
700 | | InfoStrings _info_strings; |
701 | | |
702 | | // Keeps track of the order in which InfoStrings are displayed when printed |
703 | | using InfoStringsDisplayOrder = std::vector<std::string>; |
704 | | InfoStringsDisplayOrder _info_strings_display_order; |
705 | | |
706 | | // Protects _info_strings and _info_strings_display_order |
707 | | mutable std::mutex _info_strings_lock; |
708 | | |
709 | | Counter _counter_total_time; |
710 | | // Time spent in just in this profile (i.e. not the children) as a fraction |
711 | | // of the total time in the entire profile tree. |
712 | | double _local_time_percent; |
713 | | |
714 | | // update a subtree of profiles from nodes, rooted at *idx. |
715 | | // On return, *idx points to the node immediately following this subtree. |
716 | | void update(const std::vector<TRuntimeProfileNode>& nodes, int* idx); |
717 | | |
718 | | // Similar to `void update(const std::vector<TRuntimeProfileNode>& nodes, int* idx)` |
719 | | void update(const google::protobuf::RepeatedPtrField<PRuntimeProfileNode>& nodes, int* idx); |
720 | | |
721 | | // Helper function to compute compute the fraction of the total time spent in |
722 | | // this profile and its children. |
723 | | // Called recursively. |
724 | | void compute_time_in_profile(int64_t total_time); |
725 | | |
726 | | // Print the child counters of the given counter name |
727 | | static void print_child_counters(const std::string& prefix, const std::string& counter_name, |
728 | | const CounterMap& counter_map, |
729 | | const ChildCounterMap& child_counter_map, std::ostream* s); |
730 | | }; |
731 | | |
732 | | // Utility class to update the counter at object construction and destruction. |
733 | | // When the object is constructed, decrement the counter by val. |
734 | | // When the object goes out of scope, increment the counter by val. |
735 | | class ScopedCounter { |
736 | | public: |
737 | 0 | ScopedCounter(RuntimeProfile::Counter* counter, int64_t val) : _val(val), _counter(counter) { |
738 | 0 | if (counter == nullptr) { |
739 | 0 | return; |
740 | 0 | } |
741 | 0 |
|
742 | 0 | _counter->update(-1L * _val); |
743 | 0 | } |
744 | | |
745 | | // Increment the counter when object is destroyed |
746 | 0 | ~ScopedCounter() { |
747 | 0 | if (_counter != nullptr) { |
748 | 0 | _counter->update(_val); |
749 | 0 | } |
750 | 0 | } |
751 | | |
752 | | // Disable copy constructor and assignment |
753 | | ScopedCounter(const ScopedCounter& counter) = delete; |
754 | | ScopedCounter& operator=(const ScopedCounter& counter) = delete; |
755 | | |
756 | | private: |
757 | | int64_t _val; |
758 | | RuntimeProfile::Counter* _counter = nullptr; |
759 | | }; |
760 | | |
761 | | // Utility class to update time elapsed when the object goes out of scope. |
762 | | // 'T' must implement the stopWatch "interface" (start,stop,elapsed_time) but |
763 | | // we use templates not to pay for virtual function overhead. |
764 | | template <class T, typename Bool = bool> |
765 | | class ScopedTimer { |
766 | | public: |
767 | | ScopedTimer(RuntimeProfile::Counter* counter, const Bool* is_cancelled = nullptr) |
768 | 1.39M | : _counter(counter), _is_cancelled(is_cancelled) { |
769 | 1.39M | if (counter == nullptr) { |
770 | 218 | return; |
771 | 218 | } |
772 | 1.39M | DCHECK_EQ(counter->type(), TUnit::TIME_NS); |
773 | 1.39M | _sw.start(); |
774 | 1.39M | } _ZN5doris11ScopedTimerINS_15CustomStopWatchILi1EEEbEC2EPNS_14RuntimeProfile7CounterEPKb Line | Count | Source | 768 | 1.39M | : _counter(counter), _is_cancelled(is_cancelled) { | 769 | 1.39M | if (counter == nullptr) { | 770 | 218 | return; | 771 | 218 | } | 772 | 1.39M | DCHECK_EQ(counter->type(), TUnit::TIME_NS); | 773 | 1.39M | _sw.start(); | 774 | 1.39M | } |
_ZN5doris11ScopedTimerINS_15CustomStopWatchILi3EEEbEC2EPNS_14RuntimeProfile7CounterEPKb Line | Count | Source | 768 | 87 | : _counter(counter), _is_cancelled(is_cancelled) { | 769 | 87 | if (counter == nullptr) { | 770 | 0 | return; | 771 | 0 | } | 772 | 87 | DCHECK_EQ(counter->type(), TUnit::TIME_NS); | 773 | 87 | _sw.start(); | 774 | 87 | } |
|
775 | | |
776 | | void stop() { _sw.stop(); } |
777 | | |
778 | | void start() { _sw.start(); } |
779 | | |
780 | 1.39M | bool is_cancelled() { return _is_cancelled != nullptr && *_is_cancelled; } _ZN5doris11ScopedTimerINS_15CustomStopWatchILi1EEEbE12is_cancelledEv Line | Count | Source | 780 | 1.39M | bool is_cancelled() { return _is_cancelled != nullptr && *_is_cancelled; } |
_ZN5doris11ScopedTimerINS_15CustomStopWatchILi3EEEbE12is_cancelledEv Line | Count | Source | 780 | 87 | bool is_cancelled() { return _is_cancelled != nullptr && *_is_cancelled; } |
|
781 | | |
782 | 1.39M | void UpdateCounter() { |
783 | 1.39M | if (_counter != nullptr && !is_cancelled()) { |
784 | 1.39M | _counter->update(_sw.elapsed_time()); |
785 | 1.39M | } |
786 | 1.39M | } _ZN5doris11ScopedTimerINS_15CustomStopWatchILi1EEEbE13UpdateCounterEv Line | Count | Source | 782 | 1.39M | void UpdateCounter() { | 783 | 1.39M | if (_counter != nullptr && !is_cancelled()) { | 784 | 1.39M | _counter->update(_sw.elapsed_time()); | 785 | 1.39M | } | 786 | 1.39M | } |
_ZN5doris11ScopedTimerINS_15CustomStopWatchILi3EEEbE13UpdateCounterEv Line | Count | Source | 782 | 87 | void UpdateCounter() { | 783 | 87 | if (_counter != nullptr && !is_cancelled()) { | 784 | 87 | _counter->update(_sw.elapsed_time()); | 785 | 87 | } | 786 | 87 | } |
|
787 | | |
788 | | // Update counter when object is destroyed |
789 | 1.39M | ~ScopedTimer() { |
790 | 1.39M | if (_counter == nullptr) { |
791 | 218 | return; |
792 | 218 | } |
793 | 1.39M | _sw.stop(); |
794 | 1.39M | UpdateCounter(); |
795 | 1.39M | } _ZN5doris11ScopedTimerINS_15CustomStopWatchILi1EEEbED2Ev Line | Count | Source | 789 | 1.39M | ~ScopedTimer() { | 790 | 1.39M | if (_counter == nullptr) { | 791 | 218 | return; | 792 | 218 | } | 793 | 1.39M | _sw.stop(); | 794 | 1.39M | UpdateCounter(); | 795 | 1.39M | } |
_ZN5doris11ScopedTimerINS_15CustomStopWatchILi3EEEbED2Ev Line | Count | Source | 789 | 87 | ~ScopedTimer() { | 790 | 87 | if (_counter == nullptr) { | 791 | 0 | return; | 792 | 0 | } | 793 | 87 | _sw.stop(); | 794 | 87 | UpdateCounter(); | 795 | 87 | } |
|
796 | | |
797 | | // Disable copy constructor and assignment |
798 | | ScopedTimer(const ScopedTimer& timer) = delete; |
799 | | ScopedTimer& operator=(const ScopedTimer& timer) = delete; |
800 | | |
801 | | private: |
802 | | T _sw; |
803 | | RuntimeProfile::Counter* _counter = nullptr; |
804 | | const Bool* _is_cancelled = nullptr; |
805 | | }; |
806 | | |
807 | | // Utility class to update time elapsed when the object goes out of scope. |
808 | | // 'T' must implement the stopWatch "interface" (start,stop,elapsed_time) but |
809 | | // we use templates not to pay for virtual function overhead. |
810 | | template <class T, class C> |
811 | | class ScopedRawTimer { |
812 | | public: |
813 | 820k | ScopedRawTimer(C* counter) : _counter(counter) { _sw.start(); } _ZN5doris14ScopedRawTimerINS_15CustomStopWatchILi1EEElEC2EPl Line | Count | Source | 813 | 820k | ScopedRawTimer(C* counter) : _counter(counter) { _sw.start(); } |
Unexecuted instantiation: _ZN5doris14ScopedRawTimerINS_15CustomStopWatchILi1EEESt6atomicIlEEC2EPS4_ |
814 | | // Update counter when object is destroyed |
815 | 820k | ~ScopedRawTimer() { *_counter += _sw.elapsed_time(); } _ZN5doris14ScopedRawTimerINS_15CustomStopWatchILi1EEElED2Ev Line | Count | Source | 815 | 820k | ~ScopedRawTimer() { *_counter += _sw.elapsed_time(); } |
Unexecuted instantiation: _ZN5doris14ScopedRawTimerINS_15CustomStopWatchILi1EEESt6atomicIlEED2Ev |
816 | | |
817 | | // Disable copy constructor and assignment |
818 | | ScopedRawTimer(const ScopedRawTimer& timer) = delete; |
819 | | ScopedRawTimer& operator=(const ScopedRawTimer& timer) = delete; |
820 | | |
821 | | private: |
822 | | T _sw; |
823 | | C* _counter = nullptr; |
824 | | }; |
825 | | |
826 | | } // namespace doris |