be/src/runtime/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 "core/binary_cast.hpp" |
45 | | #include "util/pretty_printer.h" |
46 | | #include "util/stopwatch.hpp" |
47 | | |
48 | | namespace doris { |
49 | | class TRuntimeProfileNode; |
50 | | class TRuntimeProfileTree; |
51 | | class RuntimeProfileCounterTreeNode; |
52 | | inline thread_local bool enable_profile_counter_check = true; |
53 | | // Some macro magic to generate unique ids using __COUNTER__ |
54 | 242M | #define CONCAT_IMPL(x, y) x##y |
55 | 242M | #define MACRO_CONCAT(x, y) CONCAT_IMPL(x, y) |
56 | | |
57 | | #define ADD_LABEL_COUNTER(profile, name) (profile)->add_counter(name, TUnit::NONE) |
58 | | #define ADD_LABEL_COUNTER_WITH_LEVEL(profile, name, level) \ |
59 | | (profile)->add_counter_with_level(name, TUnit::NONE, level) |
60 | 52.9M | #define ADD_COUNTER(profile, name, type) (profile)->add_counter(name, type) |
61 | | #define ADD_COUNTER_WITH_LEVEL(profile, name, type, level) \ |
62 | 18.1M | (profile)->add_counter_with_level(name, type, level) |
63 | 33.9M | #define ADD_TIMER(profile, name) (profile)->add_counter(name, TUnit::TIME_NS) |
64 | | #define ADD_TIMER_WITH_LEVEL(profile, name, level) \ |
65 | 28.2M | (profile)->add_counter_with_level(name, TUnit::TIME_NS, level) |
66 | 2.02M | #define ADD_CHILD_COUNTER(profile, name, type, parent) (profile)->add_counter(name, type, parent) |
67 | | #define ADD_CHILD_COUNTER_WITH_LEVEL(profile, name, type, parent, level) \ |
68 | 13.7M | (profile)->add_counter(name, type, parent, level) |
69 | 20.6M | #define ADD_CHILD_TIMER(profile, name, parent) (profile)->add_counter(name, TUnit::TIME_NS, parent) |
70 | | #define ADD_CHILD_TIMER_WITH_LEVEL(profile, name, parent, level) \ |
71 | 14.3M | (profile)->add_counter(name, TUnit::TIME_NS, parent, level) |
72 | 92.3M | #define SCOPED_TIMER(c) ScopedTimer<MonotonicStopWatch> MACRO_CONCAT(SCOPED_TIMER, __COUNTER__)(c) |
73 | | #define SCOPED_TIMER_ATOMIC(c) \ |
74 | | ScopedTimer<MonotonicStopWatch, std::atomic_bool> MACRO_CONCAT(SCOPED_TIMER, __COUNTER__)(c) |
75 | | #define SCOPED_CPU_TIMER(c) \ |
76 | 13.2M | ScopedTimer<ThreadCpuStopWatch> MACRO_CONCAT(SCOPED_TIMER, __COUNTER__)(c) |
77 | | #define CANCEL_SAFE_SCOPED_TIMER(c, is_cancelled) \ |
78 | | ScopedTimer<MonotonicStopWatch> MACRO_CONCAT(SCOPED_TIMER, __COUNTER__)(c, is_cancelled) |
79 | | #define SCOPED_RAW_TIMER(c) \ |
80 | 136M | doris::ScopedRawTimer<doris::MonotonicStopWatch, int64_t> MACRO_CONCAT(SCOPED_RAW_TIMER, \ |
81 | 136M | __COUNTER__)(c) |
82 | | #define SCOPED_ATOMIC_TIMER(c) \ |
83 | 90.2k | ScopedRawTimer<MonotonicStopWatch, std::atomic<int64_t>> MACRO_CONCAT(SCOPED_ATOMIC_TIMER, \ |
84 | 90.2k | __COUNTER__)(c) |
85 | 193M | #define COUNTER_UPDATE(c, v) (c)->update(v) |
86 | 19.2M | #define COUNTER_SET(c, v) (c)->set(v) |
87 | | |
88 | | class ObjectPool; |
89 | | |
90 | | // Runtime profile is a group of profiling counters. It supports adding named counters |
91 | | // and being able to serialize and deserialize them. |
92 | | // The profiles support a tree structure to form a hierarchy of counters. |
93 | | // Runtime profiles supports measuring wall clock rate based counters. There is a |
94 | | // single thread per process that will convert an amount (i.e. bytes) counter to a |
95 | | // corresponding rate based counter. This thread wakes up at fixed intervals and updates |
96 | | // all of the rate counters. |
97 | | // Thread-safe. |
98 | | class RuntimeProfile { |
99 | | public: |
100 | | static std::unique_ptr<RuntimeProfile> from_thrift(const TRuntimeProfileTree& node); |
101 | | |
102 | | static std::unique_ptr<RuntimeProfile> from_proto(const PRuntimeProfileTree& tree); |
103 | | |
104 | 14 | static PProfileUnit unit_to_proto(const TUnit::type& type) { |
105 | 14 | switch (type) { |
106 | 11 | case TUnit::UNIT: { |
107 | 11 | return PProfileUnit::UNIT; |
108 | 0 | } |
109 | 0 | case TUnit::UNIT_PER_SECOND: { |
110 | 0 | return PProfileUnit::UNIT_PER_SECOND; |
111 | 0 | } |
112 | 0 | case TUnit::CPU_TICKS: { |
113 | 0 | return PProfileUnit::CPU_TICKS; |
114 | 0 | } |
115 | 3 | case TUnit::BYTES: { |
116 | 3 | return PProfileUnit::BYTES; |
117 | 0 | } |
118 | 0 | case TUnit::BYTES_PER_SECOND: { |
119 | 0 | return PProfileUnit::BYTES_PER_SECOND; |
120 | 0 | } |
121 | 0 | case TUnit::TIME_NS: { |
122 | 0 | return PProfileUnit::TIME_NS; |
123 | 0 | } |
124 | 0 | case TUnit::DOUBLE_VALUE: { |
125 | 0 | return PProfileUnit::DOUBLE_VALUE; |
126 | 0 | } |
127 | 0 | case TUnit::NONE: { |
128 | 0 | return PProfileUnit::NONE; |
129 | 0 | } |
130 | 0 | case TUnit::TIME_MS: { |
131 | 0 | return PProfileUnit::TIME_MS; |
132 | 0 | } |
133 | 0 | case TUnit::TIME_S: { |
134 | 0 | return PProfileUnit::TIME_S; |
135 | 0 | } |
136 | 0 | default: { |
137 | 0 | DCHECK(false); |
138 | 0 | return PProfileUnit::NONE; |
139 | 0 | } |
140 | 14 | } |
141 | 14 | } |
142 | | |
143 | 14 | static TUnit::type unit_to_thrift(const PProfileUnit& unit) { |
144 | 14 | switch (unit) { |
145 | 12 | case PProfileUnit::UNIT: { |
146 | 12 | return TUnit::UNIT; |
147 | 0 | } |
148 | 0 | case PProfileUnit::UNIT_PER_SECOND: { |
149 | 0 | return TUnit::UNIT_PER_SECOND; |
150 | 0 | } |
151 | 0 | case PProfileUnit::CPU_TICKS: { |
152 | 0 | return TUnit::CPU_TICKS; |
153 | 0 | } |
154 | 2 | case PProfileUnit::BYTES: { |
155 | 2 | return TUnit::BYTES; |
156 | 0 | } |
157 | 0 | case PProfileUnit::BYTES_PER_SECOND: { |
158 | 0 | return TUnit::BYTES_PER_SECOND; |
159 | 0 | } |
160 | 0 | case PProfileUnit::TIME_NS: { |
161 | 0 | return TUnit::TIME_NS; |
162 | 0 | } |
163 | 0 | case PProfileUnit::DOUBLE_VALUE: { |
164 | 0 | return TUnit::DOUBLE_VALUE; |
165 | 0 | } |
166 | 0 | case PProfileUnit::NONE: { |
167 | 0 | return TUnit::NONE; |
168 | 0 | } |
169 | 0 | case PProfileUnit::TIME_MS: { |
170 | 0 | return TUnit::TIME_MS; |
171 | 0 | } |
172 | 0 | case PProfileUnit::TIME_S: { |
173 | 0 | return TUnit::TIME_S; |
174 | 0 | } |
175 | 0 | default: { |
176 | 0 | DCHECK(false); |
177 | 0 | return TUnit::NONE; |
178 | 0 | } |
179 | 14 | } |
180 | 14 | } |
181 | | |
182 | | // The root counter name for all top level counters. |
183 | | static const std::string ROOT_COUNTER; |
184 | | class Counter { |
185 | | public: |
186 | | Counter(TUnit::type type, int64_t value = 0, int64_t level = 3) |
187 | 199M | : _value(value), _type(type), _level(level) {} |
188 | 205M | virtual ~Counter() = default; |
189 | | |
190 | 3.47k | virtual Counter* clone() const { return new Counter(type(), value(), _level); } |
191 | | |
192 | 305M | virtual void update(int64_t delta) { |
193 | 305M | #ifndef NDEBUG |
194 | 305M | int64_t prev_value = _value.load(std::memory_order_seq_cst); |
195 | | // Using memory_order_seq_cst to make sure no concurrency issues, it may affect |
196 | | // performance. So that only in debug mode, we check the counter value. |
197 | 305M | _value.fetch_add(delta, std::memory_order_seq_cst); |
198 | | #else |
199 | | _value.fetch_add(delta, std::memory_order_relaxed); |
200 | | #endif |
201 | 305M | #ifndef NDEBUG |
202 | 305M | (void)prev_value; |
203 | 305M | #if !defined(BE_TEST) |
204 | 314M | if (enable_profile_counter_check) { |
205 | 314M | if (delta < 0) { |
206 | 0 | DCHECK_GT(_value.load(std::memory_order_seq_cst), -1L) |
207 | 0 | << " delta: " << delta << " prev_value: " << prev_value; |
208 | 0 | } |
209 | 314M | } |
210 | 305M | #endif |
211 | 305M | #endif |
212 | 305M | } |
213 | | |
214 | 0 | void bit_or(int64_t delta) { _value.fetch_or(delta, std::memory_order_relaxed); } |
215 | | |
216 | 15.2M | virtual void set(int64_t value) { |
217 | 15.2M | #ifndef NDEBUG |
218 | 15.2M | int64_t prev_value = _value.load(std::memory_order_seq_cst); |
219 | 15.2M | _value.store(value, std::memory_order_seq_cst); |
220 | | #else |
221 | | _value.store(value, std::memory_order_relaxed); |
222 | | #endif |
223 | 15.2M | #ifndef NDEBUG |
224 | 15.2M | (void)prev_value; |
225 | 15.2M | #if !defined(BE_TEST) |
226 | 15.3M | if (enable_profile_counter_check) { |
227 | 15.3M | DCHECK_GT(_value.load(std::memory_order_seq_cst), -1L) |
228 | 0 | << " new value: " << value << " prev_value: " << prev_value; |
229 | 15.3M | } |
230 | 15.2M | #endif |
231 | 15.2M | #endif |
232 | 15.2M | } |
233 | | |
234 | 0 | virtual void set(double value) { |
235 | 0 | DCHECK_EQ(sizeof(value), sizeof(int64_t)); |
236 | 0 | #ifndef NDEBUG |
237 | 0 | int64_t prev_value = _value.load(std::memory_order_seq_cst); |
238 | 0 | _value.store(binary_cast<double, int64_t>(value), std::memory_order_seq_cst); |
239 | | #else |
240 | | _value.store(binary_cast<double, int64_t>(value), std::memory_order_relaxed); |
241 | | #endif |
242 | 0 | #ifndef NDEBUG |
243 | 0 | (void)prev_value; |
244 | 0 | #if !defined(BE_TEST) |
245 | 0 | if (enable_profile_counter_check) { |
246 | 0 | DCHECK_GT(_value.load(std::memory_order_seq_cst), -1L) |
247 | 0 | << " new value: " << value << " prev_value: " << prev_value; |
248 | 0 | } |
249 | 0 | #endif |
250 | 0 | #endif |
251 | 0 | } |
252 | | |
253 | 10.2M | virtual int64_t value() const { return _value.load(std::memory_order_relaxed); } |
254 | | |
255 | 0 | virtual double double_value() const { |
256 | 0 | return binary_cast<int64_t, double>(_value.load(std::memory_order_relaxed)); |
257 | 0 | } |
258 | | |
259 | 431k | virtual TCounter to_thrift(const std::string& name) const { |
260 | 431k | TCounter counter; |
261 | 431k | counter.name = name; |
262 | 431k | counter.value = this->value(); |
263 | 431k | counter.type = this->type(); |
264 | 431k | counter.__set_level(this->_level); |
265 | 431k | return counter; |
266 | 431k | } |
267 | | |
268 | 8 | virtual PProfileCounter to_proto(const std::string& name) const { |
269 | 8 | PProfileCounter counter; |
270 | 8 | counter.set_name(name); |
271 | 8 | counter.set_value(this->value()); |
272 | 8 | counter.set_type(unit_to_proto(this->type())); |
273 | 8 | counter.set_level(this->value()); |
274 | 8 | return counter; |
275 | 8 | } |
276 | | |
277 | | virtual void pretty_print(std::ostream* s, const std::string& prefix, |
278 | 350 | const std::string& name) const { |
279 | 350 | std::ostream& stream = *s; |
280 | 350 | stream << prefix << " - " << name << ": " |
281 | 350 | << PrettyPrinter::print(_value.load(std::memory_order_relaxed), type()) |
282 | 350 | << std::endl; |
283 | 350 | } |
284 | | |
285 | 104M | TUnit::type type() const { return _type; } |
286 | | |
287 | 1.37M | virtual int64_t level() const { return _level; } |
288 | | |
289 | | void set_level(int64_t level) { _level = level; } |
290 | | |
291 | | bool operator==(const Counter& other) const; |
292 | | |
293 | | private: |
294 | | friend class RuntimeProfile; |
295 | | friend class RuntimeProfileCounterTreeNode; |
296 | | |
297 | | std::atomic<int64_t> _value; |
298 | | TUnit::type _type; |
299 | | int64_t _level; |
300 | | }; |
301 | | |
302 | | /// A counter that keeps track of the highest value seen (reporting that |
303 | | /// as value()) and the current value. |
304 | | class HighWaterMarkCounter : public Counter { |
305 | | public: |
306 | | HighWaterMarkCounter(TUnit::type unit, int64_t level, const std::string& parent_name, |
307 | | int64_t value = 0, int64_t current_value = 0) |
308 | 7.37M | : Counter(unit, value, level), |
309 | 7.37M | current_value_(current_value), |
310 | 7.37M | _parent_name(parent_name) {} |
311 | | |
312 | 8.26k | virtual Counter* clone() const override { |
313 | 8.26k | return new HighWaterMarkCounter(type(), level(), parent_name(), value(), |
314 | 8.26k | current_value()); |
315 | 8.26k | } |
316 | | |
317 | 6.15M | void add(int64_t delta) { |
318 | 6.15M | #ifndef NDEBUG |
319 | 6.15M | current_value_.fetch_add(delta, std::memory_order_seq_cst); |
320 | 6.15M | if (delta > 0) { |
321 | 1.55M | UpdateMax(current_value_); |
322 | 1.55M | } |
323 | | //if (enable_profile_counter_check) { |
324 | | // DCHECK_GT(current_value_.load(std::memory_order_seq_cst), -1L); |
325 | | //} |
326 | | #else |
327 | | current_value_.fetch_add(delta, std::memory_order_relaxed); |
328 | | if (delta > 0) { |
329 | | UpdateMax(current_value_); |
330 | | } |
331 | | #endif |
332 | 6.15M | } |
333 | 4.28M | virtual void update(int64_t delta) override { add(delta); } |
334 | | |
335 | 38.6k | TCounter to_thrift(const std::string& name) const override { |
336 | 38.6k | TCounter counter; |
337 | 38.6k | counter.name = std::move(name); |
338 | 38.6k | counter.value = current_value(); |
339 | 38.6k | counter.type = type(); |
340 | 38.6k | counter.__set_level(level()); |
341 | 38.6k | return counter; |
342 | 38.6k | } |
343 | | |
344 | 0 | PProfileCounter to_proto(const std::string& name) const override { |
345 | 0 | PProfileCounter counter; |
346 | 0 | counter.set_name(name); |
347 | 0 | counter.set_value(current_value()); |
348 | 0 | counter.set_type(unit_to_proto(this->type())); |
349 | 0 | counter.set_level(this->value()); |
350 | 0 | return counter; |
351 | 0 | } |
352 | | |
353 | 38.6k | TCounter to_thrift_peak(std::string name) { |
354 | 38.6k | TCounter counter; |
355 | 38.6k | counter.name = std::move(name); |
356 | 38.6k | counter.value = value(); |
357 | 38.6k | counter.type = type(); |
358 | 38.6k | counter.__set_level(level()); |
359 | 38.6k | return counter; |
360 | 38.6k | } |
361 | | |
362 | 0 | PProfileCounter to_proto_peak(const std::string& name) const { |
363 | 0 | return Counter::to_proto(name); |
364 | 0 | } |
365 | | |
366 | | virtual void pretty_print(std::ostream* s, const std::string& prefix, |
367 | 82 | const std::string& name) const override { |
368 | 82 | std::ostream& stream = *s; |
369 | 82 | stream << prefix << " - " << name |
370 | 82 | << " Current: " << PrettyPrinter::print(current_value(), type()) << " (Peak: " |
371 | 82 | << PrettyPrinter::print(_value.load(std::memory_order_relaxed), type()) << ")" |
372 | 82 | << std::endl; |
373 | 82 | } |
374 | | |
375 | | /// Tries to increase the current value by delta. If current_value() + delta |
376 | | /// exceeds max, return false and current_value is not changed. |
377 | 0 | bool try_add(int64_t delta, int64_t max) { |
378 | 0 | while (true) { |
379 | 0 | int64_t old_val = current_value_.load(std::memory_order_relaxed); |
380 | 0 | int64_t new_val = old_val + delta; |
381 | 0 | if (UNLIKELY(new_val > max)) return false; |
382 | 0 | if (LIKELY(current_value_.compare_exchange_weak(old_val, new_val, |
383 | 0 | std::memory_order_relaxed))) { |
384 | 0 | UpdateMax(new_val); |
385 | 0 | return true; |
386 | 0 | } |
387 | 0 | } |
388 | 0 | } |
389 | | |
390 | 7.97M | void set(int64_t v) override { |
391 | 7.97M | #ifndef NDEBUG |
392 | 7.97M | int64_t prev_value = current_value_.load(std::memory_order_seq_cst); |
393 | 7.97M | int64_t prev_max_value = _value.load(std::memory_order_seq_cst); |
394 | 7.97M | current_value_.store(v, std::memory_order_seq_cst); |
395 | | #else |
396 | | current_value_.store(v, std::memory_order_relaxed); |
397 | | #endif |
398 | 7.97M | UpdateMax(v); |
399 | 7.97M | #ifndef NDEBUG |
400 | 7.97M | (void)prev_value; |
401 | 7.97M | (void)prev_max_value; |
402 | 7.97M | #if !defined(BE_TEST) |
403 | | |
404 | 7.97M | if (enable_profile_counter_check) { |
405 | 1.90M | DCHECK_GT(current_value_.load(std::memory_order_seq_cst), -1L) |
406 | 0 | << " prev_value: " << prev_value; |
407 | 1.90M | DCHECK_GT(_value.load(std::memory_order_seq_cst), -1L) |
408 | 0 | << " prev_max_value: " << prev_max_value << " prev_value: " << prev_value; |
409 | 1.90M | } |
410 | 7.97M | #endif |
411 | 7.97M | #endif |
412 | 7.97M | } |
413 | | |
414 | 48.0k | int64_t current_value() const { return current_value_.load(std::memory_order_relaxed); } |
415 | | |
416 | 46.9k | std::string parent_name() const { return _parent_name; } |
417 | | |
418 | | private: |
419 | | /// Set '_value' to 'v' if 'v' is larger than '_value'. The entire operation is |
420 | | /// atomic. |
421 | 9.62M | void UpdateMax(int64_t v) { |
422 | 9.62M | while (true) { |
423 | 9.62M | int64_t old_max = _value.load(std::memory_order_relaxed); |
424 | 9.62M | int64_t new_max = std::max(old_max, v); |
425 | 9.62M | if (new_max == old_max) { |
426 | 8.35M | break; // Avoid atomic update. |
427 | 8.35M | } |
428 | 1.26M | if (LIKELY(_value.compare_exchange_weak(old_max, new_max, |
429 | 1.26M | std::memory_order_relaxed))) { |
430 | 1.26M | break; |
431 | 1.26M | } |
432 | 1.26M | } |
433 | 9.62M | } |
434 | | |
435 | | /// The current value of the counter. _value in the super class represents |
436 | | /// the high water mark. |
437 | | std::atomic<int64_t> current_value_; |
438 | | |
439 | | const std::string _parent_name; |
440 | | }; |
441 | | |
442 | | using DerivedCounterFunction = std::function<int64_t()>; |
443 | | |
444 | | // A DerivedCounter also has a name and type, but the value is computed. |
445 | | // Do not call Set() and Update(). |
446 | | class DerivedCounter : public Counter { |
447 | | public: |
448 | | DerivedCounter(TUnit::type type, const DerivedCounterFunction& counter_fn, |
449 | | int64_t value = 0, int64_t level = 1) |
450 | 563k | : Counter(type, value, level), _counter_fn(counter_fn) {} |
451 | | |
452 | 0 | virtual Counter* clone() const override { |
453 | 0 | return new DerivedCounter(type(), _counter_fn, value(), level()); |
454 | 0 | } |
455 | | |
456 | 8.03k | int64_t value() const override { return _counter_fn(); } |
457 | | |
458 | | private: |
459 | | DerivedCounterFunction _counter_fn; |
460 | | }; |
461 | | |
462 | | using ConditionCounterFunction = std::function<bool(int64_t, int64_t)>; |
463 | | |
464 | | // ConditionCounter is a specialized counter that only updates its value when a specific condition is met. |
465 | | // It uses a condition function (condition_func) to determine when the counter's value should be updated. |
466 | | // This type of counter is particularly useful for tracking maximum values, minimum values, or other metrics |
467 | | // that should only be updated when they meet certain criteria. |
468 | | // For example, it can be used to record the maximum value of a specific metric during query execution, |
469 | | // or to update the counter only when a new value exceeds some threshold. |
470 | | class ConditionCounter : public Counter { |
471 | | public: |
472 | | ConditionCounter(TUnit::type type, const ConditionCounterFunction& condition_func, |
473 | | int64_t level = 2, int64_t condition = 0, int64_t value = 0) |
474 | 4 | : Counter(type, value, level), |
475 | 4 | _condition(condition), |
476 | 4 | _stored_value(value), |
477 | 4 | _condition_func(condition_func) {} |
478 | | |
479 | 0 | Counter* clone() const override { |
480 | 0 | std::lock_guard<std::mutex> l(_mutex); |
481 | 0 | return new ConditionCounter(type(), _condition_func, _condition, value(), level()); |
482 | 0 | } |
483 | | |
484 | 7 | int64_t value() const override { |
485 | 7 | std::lock_guard<std::mutex> l(_mutex); |
486 | 7 | return _stored_value; |
487 | 7 | } |
488 | | |
489 | 9 | void conditional_update(int64_t c, int64_t v) { |
490 | 9 | std::lock_guard<std::mutex> l(_mutex); |
491 | 9 | if (_condition_func(_condition, c)) { |
492 | 6 | _stored_value = v; |
493 | 6 | _condition = c; |
494 | 6 | } |
495 | 9 | } |
496 | | |
497 | | private: |
498 | | mutable std::mutex _mutex; |
499 | | int64_t _condition; |
500 | | int64_t _stored_value; |
501 | | ConditionCounterFunction _condition_func; |
502 | | }; |
503 | | |
504 | | // NonZeroCounter will not be converted to Thrift if the value is 0. |
505 | | class NonZeroCounter : public Counter { |
506 | | public: |
507 | | NonZeroCounter(TUnit::type type, int64_t level, const std::string& parent_name, |
508 | | int64_t value = 0) |
509 | 2.57M | : Counter(type, value, level), _parent_name(parent_name) {} |
510 | | |
511 | 0 | virtual Counter* clone() const override { |
512 | 0 | return new NonZeroCounter(type(), level(), parent_name(), value()); |
513 | 0 | } |
514 | | |
515 | 12.9k | std::string parent_name() const { return _parent_name; } |
516 | | |
517 | | private: |
518 | | const std::string _parent_name; |
519 | | }; |
520 | | |
521 | | class DescriptionEntry : public Counter { |
522 | | public: |
523 | | DescriptionEntry(const std::string& name, const std::string& description) |
524 | 197k | : Counter(TUnit::NONE, 0, 2), _description(description), _name(name) {} |
525 | | |
526 | 0 | virtual Counter* clone() const override { |
527 | 0 | return new DescriptionEntry(_name, _description); |
528 | 0 | } |
529 | | |
530 | 0 | void set(int64_t value) override { |
531 | | // Do nothing |
532 | 0 | } |
533 | 0 | void set(double value) override { |
534 | | // Do nothing |
535 | 0 | } |
536 | 0 | void update(int64_t delta) override { |
537 | | // Do nothing |
538 | 0 | } |
539 | | |
540 | 1 | TCounter to_thrift(const std::string& name) const override { |
541 | 1 | TCounter counter; |
542 | 1 | counter.name = name; |
543 | 1 | counter.__set_level(2); |
544 | 1 | counter.__set_description(_description); |
545 | 1 | return counter; |
546 | 1 | } |
547 | | |
548 | 0 | PProfileCounter to_proto(const std::string& name) const override { |
549 | 0 | PProfileCounter counter; |
550 | 0 | counter.set_name(name); |
551 | 0 | counter.set_level(2); |
552 | 0 | counter.set_description(_description); |
553 | 0 | return counter; |
554 | 0 | } |
555 | | |
556 | | private: |
557 | | const std::string _description; |
558 | | const std::string _name; |
559 | | }; |
560 | | |
561 | | // Create a runtime profile object with 'name'. |
562 | | RuntimeProfile(const std::string& name, bool is_averaged_profile = false); |
563 | | |
564 | | ~RuntimeProfile(); |
565 | | |
566 | | // Adds a child profile. This is thread safe. |
567 | | // 'indent' indicates whether the child will be printed w/ extra indentation |
568 | | // relative to the parent. |
569 | | // If location is non-null, child will be inserted after location. Location must |
570 | | // already be added to the profile. |
571 | | void add_child(RuntimeProfile* child, bool indent, RuntimeProfile* location = nullptr); |
572 | | |
573 | | void add_child_unlock(RuntimeProfile* child, bool indent, RuntimeProfile* loc); |
574 | | |
575 | | /// Creates a new child profile with the given 'name'. A child profile with that name |
576 | | /// must not already exist. If 'prepend' is true, prepended before other child profiles, |
577 | | /// otherwise appended after other child profiles. |
578 | | RuntimeProfile* create_child(const std::string& name, bool indent = true, bool prepend = false); |
579 | | |
580 | | // Merges the src profile into this one, combining counters that have an identical |
581 | | // path. Info strings from profiles are not merged. 'src' would be a const if it |
582 | | // weren't for locking. |
583 | | // Calling this concurrently on two RuntimeProfiles in reverse order results in |
584 | | // undefined behavior. |
585 | | void merge(const RuntimeProfile* src); |
586 | | |
587 | | // Updates this profile w/ the thrift profile: behaves like Merge(), except |
588 | | // that existing counters are updated rather than added up. |
589 | | // Info strings matched up by key and are updated or added, depending on whether |
590 | | // the key has already been registered. |
591 | | void update(const TRuntimeProfileTree& thrift_profile); |
592 | | |
593 | | //Similar to `void update(const TRuntimeProfileTree& thrift_profile)` |
594 | | void update(const PRuntimeProfileTree& proto_profile); |
595 | | |
596 | | // Add a counter with 'name'/'type'. Returns a counter object that the caller can |
597 | | // update. The counter is owned by the RuntimeProfile object. |
598 | | // If parent_counter_name is a non-empty string, the counter is added as a child of |
599 | | // parent_counter_name. |
600 | | // If the counter already exists, the existing counter object is returned. |
601 | | Counter* add_counter(const std::string& name, TUnit::type type, |
602 | | const std::string& parent_counter_name, int64_t level = 2); |
603 | | |
604 | 85.4M | Counter* add_counter(const std::string& name, TUnit::type type) { |
605 | 85.4M | return add_counter(name, type, RuntimeProfile::ROOT_COUNTER); |
606 | 85.4M | } |
607 | | |
608 | 45.6M | Counter* add_counter_with_level(const std::string& name, TUnit::type type, int64_t level) { |
609 | 45.6M | return add_counter(name, type, RuntimeProfile::ROOT_COUNTER, level); |
610 | 45.6M | } |
611 | | |
612 | | NonZeroCounter* add_nonzero_counter( |
613 | | const std::string& name, TUnit::type type, |
614 | | const std::string& parent_counter_name = RuntimeProfile::ROOT_COUNTER, |
615 | | int64_t level = 2); |
616 | | |
617 | | // Add a description entry under target counter. |
618 | | void add_description(const std::string& name, const std::string& description, |
619 | | std::string parent_counter_name); |
620 | | // Add a derived counter with 'name'/'type'. The counter is owned by the |
621 | | // RuntimeProfile object. |
622 | | // If parent_counter_name is a non-empty string, the counter is added as a child of |
623 | | // parent_counter_name. |
624 | | // Returns nullptr if the counter already exists. |
625 | | DerivedCounter* add_derived_counter(const std::string& name, TUnit::type type, |
626 | | const DerivedCounterFunction& counter_fn, |
627 | | const std::string& parent_counter_name); |
628 | | |
629 | | ConditionCounter* add_conditition_counter(const std::string& name, TUnit::type type, |
630 | | const ConditionCounterFunction& counter_fn, |
631 | | const std::string& parent_counter_name, |
632 | | int64_t level = 2); |
633 | | |
634 | | // Gets the counter object with 'name'. Returns nullptr if there is no counter with |
635 | | // that name. |
636 | | Counter* get_counter(const std::string& name); |
637 | | |
638 | | // Adds all counters with 'name' that are registered either in this or |
639 | | // in any of the child profiles to 'counters'. |
640 | | void get_counters(const std::string& name, std::vector<Counter*>* counters); |
641 | | |
642 | | // Helper to append to the "ExecOption" info string. |
643 | 0 | void append_exec_option(const std::string& option) { add_info_string("ExecOption", option); } |
644 | | |
645 | | // Adds a string to the runtime profile. If a value already exists for 'key', |
646 | | // the value will be updated. |
647 | | void add_info_string(const std::string& key, const std::string& value); |
648 | | |
649 | | // Returns a pointer to the info string value for 'key'. Returns nullptr if |
650 | | // the key does not exist. |
651 | | const std::string* get_info_string(const std::string& key); |
652 | | |
653 | | // Returns the counter for the total elapsed time. |
654 | 22.8M | Counter* total_time_counter() { return &_counter_total_time; } |
655 | | |
656 | | // Prints the counters in a name: value format. |
657 | | // Does not hold locks when it makes any function calls. |
658 | | void pretty_print(std::ostream* s, const std::string& prefix = "", |
659 | | int64_t profile_level = 2) const; |
660 | 26 | std::string pretty_print() const { |
661 | 26 | std::stringstream ss; |
662 | 26 | pretty_print(&ss); |
663 | 26 | return ss.str(); |
664 | 26 | }; |
665 | | |
666 | | // Serializes profile to thrift. |
667 | | // Does not hold locks when it makes any function calls. |
668 | | void to_thrift(TRuntimeProfileTree* tree, int64_t profile_level = 2); |
669 | | void to_thrift(std::vector<TRuntimeProfileNode>* nodes, int64_t profile_level = 2); |
670 | | |
671 | | // Similar to `to_thrift`. |
672 | | void to_proto(PRuntimeProfileTree* tree, int64_t profile_level = 2); |
673 | | void to_proto(google::protobuf::RepeatedPtrField<PRuntimeProfileNode>* nodes, |
674 | | int64_t profile_level = 2); |
675 | | |
676 | | // Divides all counters by n |
677 | | void divide(int n); |
678 | | |
679 | | RuntimeProfile* get_child(std::string name); |
680 | | |
681 | | void get_children(std::vector<RuntimeProfile*>* children) const; |
682 | | |
683 | | // Gets all profiles in tree, including this one. |
684 | | void get_all_children(std::vector<RuntimeProfile*>* children); |
685 | | |
686 | | // Returns the number of counters in this profile |
687 | | int num_counters() const { return cast_set<int>(_counter_map.size()); } |
688 | | |
689 | | // Returns name of this profile |
690 | 5.20k | const std::string& name() const { return _name; } |
691 | | |
692 | | // *only call this on top-level profiles* |
693 | | // (because it doesn't re-file child profiles) |
694 | 570 | void set_name(const std::string& name) { _name = name; } |
695 | | |
696 | 608 | int64_t metadata() const { return _metadata; } |
697 | 4.03M | void set_metadata(int64_t md) { |
698 | 4.03M | _is_set_metadata = true; |
699 | 4.03M | _metadata = md; |
700 | 4.03M | } |
701 | | |
702 | 3.01M | bool is_set_metadata() const { return _is_set_metadata; } |
703 | | |
704 | 0 | time_t timestamp() const { return _timestamp; } |
705 | 11 | void set_timestamp(time_t ss) { _timestamp = ss; } |
706 | | |
707 | | // Derived counter function: return measured throughput as input_value/second. |
708 | | static int64_t units_per_second(const Counter* total_counter, const Counter* timer); |
709 | | |
710 | | // Derived counter function: return aggregated value |
711 | | static int64_t counter_sum(const std::vector<Counter*>* counters); |
712 | | |
713 | | /// Adds a high water mark counter to the runtime profile. Otherwise, same behavior |
714 | | /// as AddCounter(). |
715 | | HighWaterMarkCounter* AddHighWaterMarkCounter( |
716 | | const std::string& name, TUnit::type unit, |
717 | | const std::string& parent_counter_name = RuntimeProfile::ROOT_COUNTER, |
718 | | int64_t profile_level = 2); |
719 | | |
720 | | // Recursively compute the fraction of the 'total_time' spent in this profile and |
721 | | // its children. |
722 | | // This function updates _local_time_percent for each profile. |
723 | | void compute_time_in_profile(); |
724 | | |
725 | | void clear_children(); |
726 | | |
727 | | private: |
728 | | // RuntimeProfileCounterTreeNode needs to access the counter map and child counter map |
729 | | friend class RuntimeProfileCounterTreeNode; |
730 | | // Pool for allocated counters. Usually owned by the creator of this |
731 | | // object, but occasionally allocated in the constructor. |
732 | | std::unique_ptr<ObjectPool> _pool; |
733 | | |
734 | | // Pool for allocated counters. These counters are shared with some other objects. |
735 | | std::map<std::string, std::shared_ptr<HighWaterMarkCounter>> _shared_counter_pool; |
736 | | |
737 | | // Name for this runtime profile. |
738 | | std::string _name; |
739 | | |
740 | | // user-supplied, uninterpreted metadata. |
741 | | int64_t _metadata; |
742 | | bool _is_set_metadata = false; |
743 | | |
744 | | bool _is_sink = false; |
745 | | bool _is_set_sink = false; |
746 | | |
747 | | // The timestamp when the profile was modified, make sure the update is up to date. |
748 | | time_t _timestamp; |
749 | | |
750 | | /// True if this profile is an average derived from other profiles. |
751 | | /// All counters in this profile must be of unit AveragedCounter. |
752 | | bool _is_averaged_profile; |
753 | | |
754 | | // Map from counter names to counters. The profile owns the memory for the |
755 | | // counters. |
756 | | using CounterMap = std::map<std::string, Counter*>; |
757 | | CounterMap _counter_map; |
758 | | |
759 | | // Map from parent counter name to a set of child counter name. |
760 | | // All top level counters are the child of RuntimeProfile::ROOT_COUNTER (root). |
761 | | using ChildCounterMap = std::map<std::string, std::set<std::string>>; |
762 | | ChildCounterMap _child_counter_map; |
763 | | |
764 | | // protects _counter_map, _counter_child_map and _bucketing_counters |
765 | | mutable std::mutex _counter_map_lock; |
766 | | |
767 | | // Child profiles. Does not own memory. |
768 | | // We record children in both a map (to facilitate updates) and a vector |
769 | | // (to print things in the order they were registered) |
770 | | using ChildMap = std::map<std::string, RuntimeProfile*>; |
771 | | ChildMap _child_map; |
772 | | // vector of (profile, indentation flag) |
773 | | using ChildVector = std::vector<std::pair<RuntimeProfile*, bool>>; |
774 | | ChildVector _children; |
775 | | mutable std::mutex _children_lock; // protects _child_map and _children |
776 | | |
777 | | using InfoStrings = std::map<std::string, std::string>; |
778 | | InfoStrings _info_strings; |
779 | | |
780 | | // Keeps track of the order in which InfoStrings are displayed when printed |
781 | | using InfoStringsDisplayOrder = std::vector<std::string>; |
782 | | InfoStringsDisplayOrder _info_strings_display_order; |
783 | | |
784 | | // Protects _info_strings and _info_strings_display_order |
785 | | mutable std::mutex _info_strings_lock; |
786 | | |
787 | | Counter _counter_total_time; |
788 | | // Time spent in just in this profile (i.e. not the children) as a fraction |
789 | | // of the total time in the entire profile tree. |
790 | | double _local_time_percent; |
791 | | |
792 | | // update a subtree of profiles from nodes, rooted at *idx. |
793 | | // On return, *idx points to the node immediately following this subtree. |
794 | | void update(const std::vector<TRuntimeProfileNode>& nodes, int* idx); |
795 | | |
796 | | // Similar to `void update(const std::vector<TRuntimeProfileNode>& nodes, int* idx)` |
797 | | void update(const google::protobuf::RepeatedPtrField<PRuntimeProfileNode>& nodes, int* idx); |
798 | | |
799 | | // Helper function to compute compute the fraction of the total time spent in |
800 | | // this profile and its children. |
801 | | // Called recursively. |
802 | | void compute_time_in_profile(int64_t total_time); |
803 | | |
804 | | // Print the child counters of the given counter name |
805 | | static void print_child_counters(const std::string& prefix, const std::string& counter_name, |
806 | | const CounterMap& counter_map, |
807 | | const ChildCounterMap& child_counter_map, std::ostream* s); |
808 | | }; |
809 | | |
810 | | // Utility class to update the counter at object construction and destruction. |
811 | | // When the object is constructed, decrement the counter by val. |
812 | | // When the object goes out of scope, increment the counter by val. |
813 | | class ScopedCounter { |
814 | | public: |
815 | 0 | ScopedCounter(RuntimeProfile::Counter* counter, int64_t val) : _val(val), _counter(counter) { |
816 | 0 | if (counter == nullptr) { |
817 | 0 | return; |
818 | 0 | } |
819 | 0 |
|
820 | 0 | _counter->update(-1L * _val); |
821 | 0 | } |
822 | | |
823 | | // Increment the counter when object is destroyed |
824 | 0 | ~ScopedCounter() { |
825 | 0 | if (_counter != nullptr) { |
826 | 0 | _counter->update(_val); |
827 | 0 | } |
828 | 0 | } |
829 | | |
830 | | // Disable copy constructor and assignment |
831 | | ScopedCounter(const ScopedCounter& counter) = delete; |
832 | | ScopedCounter& operator=(const ScopedCounter& counter) = delete; |
833 | | |
834 | | private: |
835 | | int64_t _val; |
836 | | RuntimeProfile::Counter* _counter = nullptr; |
837 | | }; |
838 | | |
839 | | // Utility class to update time elapsed when the object goes out of scope. |
840 | | // 'T' must implement the stopWatch "interface" (start,stop,elapsed_time) but |
841 | | // we use templates not to pay for virtual function overhead. |
842 | | template <class T, typename Bool = bool> |
843 | | class ScopedTimer { |
844 | | public: |
845 | | ScopedTimer(RuntimeProfile::Counter* counter, const Bool* is_cancelled = nullptr) |
846 | 104M | : _counter(counter), _is_cancelled(is_cancelled) { |
847 | 104M | if (counter == nullptr) { |
848 | 848k | return; |
849 | 848k | } |
850 | 104M | DCHECK_EQ(counter->type(), TUnit::TIME_NS); |
851 | 103M | _sw.start(); |
852 | 103M | } _ZN5doris11ScopedTimerINS_15CustomStopWatchILi1EEEbEC2EPNS_14RuntimeProfile7CounterEPKb Line | Count | Source | 846 | 91.3M | : _counter(counter), _is_cancelled(is_cancelled) { | 847 | 91.3M | if (counter == nullptr) { | 848 | 848k | return; | 849 | 848k | } | 850 | 91.3M | DCHECK_EQ(counter->type(), TUnit::TIME_NS); | 851 | 90.4M | _sw.start(); | 852 | 90.4M | } |
_ZN5doris11ScopedTimerINS_15CustomStopWatchILi3EEEbEC2EPNS_14RuntimeProfile7CounterEPKb Line | Count | Source | 846 | 13.1M | : _counter(counter), _is_cancelled(is_cancelled) { | 847 | 13.1M | if (counter == nullptr) { | 848 | 0 | return; | 849 | 0 | } | 850 | 13.1M | DCHECK_EQ(counter->type(), TUnit::TIME_NS); | 851 | 13.1M | _sw.start(); | 852 | 13.1M | } |
|
853 | | |
854 | | void stop() { _sw.stop(); } |
855 | | |
856 | | void start() { _sw.start(); } |
857 | | |
858 | 104M | bool is_cancelled() { return _is_cancelled != nullptr && *_is_cancelled; }_ZN5doris11ScopedTimerINS_15CustomStopWatchILi1EEEbE12is_cancelledEv Line | Count | Source | 858 | 91.7M | bool is_cancelled() { return _is_cancelled != nullptr && *_is_cancelled; } |
_ZN5doris11ScopedTimerINS_15CustomStopWatchILi3EEEbE12is_cancelledEv Line | Count | Source | 858 | 13.2M | bool is_cancelled() { return _is_cancelled != nullptr && *_is_cancelled; } |
|
859 | | |
860 | 104M | void UpdateCounter() { |
861 | 105M | if (_counter != nullptr && !is_cancelled()) { |
862 | 105M | _counter->update(_sw.elapsed_time()); |
863 | 105M | } |
864 | 104M | } _ZN5doris11ScopedTimerINS_15CustomStopWatchILi1EEEbE13UpdateCounterEv Line | Count | Source | 860 | 91.6M | void UpdateCounter() { | 861 | 91.7M | if (_counter != nullptr && !is_cancelled()) { | 862 | 91.7M | _counter->update(_sw.elapsed_time()); | 863 | 91.7M | } | 864 | 91.6M | } |
_ZN5doris11ScopedTimerINS_15CustomStopWatchILi3EEEbE13UpdateCounterEv Line | Count | Source | 860 | 13.2M | void UpdateCounter() { | 861 | 13.2M | if (_counter != nullptr && !is_cancelled()) { | 862 | 13.2M | _counter->update(_sw.elapsed_time()); | 863 | 13.2M | } | 864 | 13.2M | } |
|
865 | | |
866 | | // Update counter when object is destroyed |
867 | 105M | ~ScopedTimer() { |
868 | 105M | if (_counter == nullptr) { |
869 | 848k | return; |
870 | 848k | } |
871 | 104M | _sw.stop(); |
872 | 104M | UpdateCounter(); |
873 | 104M | } _ZN5doris11ScopedTimerINS_15CustomStopWatchILi1EEEbED2Ev Line | Count | Source | 867 | 91.9M | ~ScopedTimer() { | 868 | 91.9M | if (_counter == nullptr) { | 869 | 848k | return; | 870 | 848k | } | 871 | 91.1M | _sw.stop(); | 872 | 91.1M | UpdateCounter(); | 873 | 91.1M | } |
_ZN5doris11ScopedTimerINS_15CustomStopWatchILi3EEEbED2Ev Line | Count | Source | 867 | 13.2M | ~ScopedTimer() { | 868 | 13.2M | if (_counter == nullptr) { | 869 | 0 | return; | 870 | 0 | } | 871 | 13.2M | _sw.stop(); | 872 | 13.2M | UpdateCounter(); | 873 | 13.2M | } |
|
874 | | |
875 | | // Disable copy constructor and assignment |
876 | | ScopedTimer(const ScopedTimer& timer) = delete; |
877 | | ScopedTimer& operator=(const ScopedTimer& timer) = delete; |
878 | | |
879 | | private: |
880 | | T _sw; |
881 | | RuntimeProfile::Counter* _counter = nullptr; |
882 | | const Bool* _is_cancelled = nullptr; |
883 | | }; |
884 | | |
885 | | // Utility class to update time elapsed when the object goes out of scope. |
886 | | // 'T' must implement the stopWatch "interface" (start,stop,elapsed_time) but |
887 | | // we use templates not to pay for virtual function overhead. |
888 | | template <class T, class C> |
889 | | class ScopedRawTimer { |
890 | | public: |
891 | 134M | ScopedRawTimer(C* counter) : _counter(counter) { _sw.start(); }_ZN5doris14ScopedRawTimerINS_15CustomStopWatchILi1EEElEC2EPl Line | Count | Source | 891 | 134M | ScopedRawTimer(C* counter) : _counter(counter) { _sw.start(); } |
_ZN5doris14ScopedRawTimerINS_15CustomStopWatchILi1EEESt6atomicIlEEC2EPS4_ Line | Count | Source | 891 | 90.2k | ScopedRawTimer(C* counter) : _counter(counter) { _sw.start(); } |
|
892 | | // Update counter when object is destroyed |
893 | 135M | ~ScopedRawTimer() { *_counter += _sw.elapsed_time(); }_ZN5doris14ScopedRawTimerINS_15CustomStopWatchILi1EEElED2Ev Line | Count | Source | 893 | 135M | ~ScopedRawTimer() { *_counter += _sw.elapsed_time(); } |
_ZN5doris14ScopedRawTimerINS_15CustomStopWatchILi1EEESt6atomicIlEED2Ev Line | Count | Source | 893 | 90.1k | ~ScopedRawTimer() { *_counter += _sw.elapsed_time(); } |
|
894 | | |
895 | | // Disable copy constructor and assignment |
896 | | ScopedRawTimer(const ScopedRawTimer& timer) = delete; |
897 | | ScopedRawTimer& operator=(const ScopedRawTimer& timer) = delete; |
898 | | |
899 | | private: |
900 | | T _sw; |
901 | | C* _counter = nullptr; |
902 | | }; |
903 | | } // namespace doris |