db_bench.cc raw
1 // Copyright (c) 2011 The LevelDB Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file. See the AUTHORS file for names of contributors.
4
5 #include <stdio.h>
6 #include <stdlib.h>
7 #include <sys/types.h>
8
9 #include "leveldb/cache.h"
10 #include "leveldb/db.h"
11 #include "leveldb/env.h"
12 #include "leveldb/filter_policy.h"
13 #include "leveldb/write_batch.h"
14 #include "port/port.h"
15 #include "util/crc32c.h"
16 #include "util/histogram.h"
17 #include "util/mutexlock.h"
18 #include "util/random.h"
19 #include "util/testutil.h"
20
21 // Comma-separated list of operations to run in the specified order
22 // Actual benchmarks:
23 // fillseq -- write N values in sequential key order in async mode
24 // fillrandom -- write N values in random key order in async mode
25 // overwrite -- overwrite N values in random key order in async mode
26 // fillsync -- write N/100 values in random key order in sync mode
27 // fill100K -- write N/1000 100K values in random order in async mode
28 // deleteseq -- delete N keys in sequential order
29 // deleterandom -- delete N keys in random order
30 // readseq -- read N times sequentially
31 // readreverse -- read N times in reverse order
32 // readrandom -- read N times in random order
33 // readmissing -- read N missing keys in random order
34 // readhot -- read N times in random order from 1% section of DB
35 // seekrandom -- N random seeks
36 // open -- cost of opening a DB
37 // crc32c -- repeated crc32c of 4K of data
38 // Meta operations:
39 // compact -- Compact the entire DB
40 // stats -- Print DB stats
41 // sstables -- Print sstable info
42 // heapprofile -- Dump a heap profile (if supported by this port)
43 static const char* FLAGS_benchmarks =
44 "fillseq,"
45 "fillsync,"
46 "fillrandom,"
47 "overwrite,"
48 "readrandom,"
49 "readrandom," // Extra run to allow previous compactions to quiesce
50 "readseq,"
51 "readreverse,"
52 "compact,"
53 "readrandom,"
54 "readseq,"
55 "readreverse,"
56 "fill100K,"
57 "crc32c,"
58 "snappycomp,"
59 "snappyuncomp,";
60
61 // Number of key/values to place in database
62 static int FLAGS_num = 1000000;
63
64 // Number of read operations to do. If negative, do FLAGS_num reads.
65 static int FLAGS_reads = -1;
66
67 // Number of concurrent threads to run.
68 static int FLAGS_threads = 1;
69
70 // Size of each value
71 static int FLAGS_value_size = 100;
72
73 // Arrange to generate values that shrink to this fraction of
74 // their original size after compression
75 static double FLAGS_compression_ratio = 0.5;
76
77 // Print histogram of operation timings
78 static bool FLAGS_histogram = false;
79
80 // Number of bytes to buffer in memtable before compacting
81 // (initialized to default value by "main")
82 static int FLAGS_write_buffer_size = 0;
83
84 // Number of bytes written to each file.
85 // (initialized to default value by "main")
86 static int FLAGS_max_file_size = 0;
87
88 // Approximate size of user data packed per block (before compression.
89 // (initialized to default value by "main")
90 static int FLAGS_block_size = 0;
91
92 // Number of bytes to use as a cache of uncompressed data.
93 // Negative means use default settings.
94 static int FLAGS_cache_size = -1;
95
96 // Maximum number of files to keep open at the same time (use default if == 0)
97 static int FLAGS_open_files = 0;
98
99 // Bloom filter bits per key.
100 // Negative means use default settings.
101 static int FLAGS_bloom_bits = -1;
102
103 // If true, do not destroy the existing database. If you set this
104 // flag and also specify a benchmark that wants a fresh database, that
105 // benchmark will fail.
106 static bool FLAGS_use_existing_db = false;
107
108 // If true, reuse existing log/MANIFEST files when re-opening a database.
109 static bool FLAGS_reuse_logs = false;
110
111 // Use the db with the following name.
112 static const char* FLAGS_db = nullptr;
113
114 namespace leveldb {
115
116 namespace {
117 leveldb::Env* g_env = nullptr;
118
119 // Helper for quickly generating random data.
120 class RandomGenerator {
121 private:
122 std::string data_;
123 int pos_;
124
125 public:
126 RandomGenerator() {
127 // We use a limited amount of data over and over again and ensure
128 // that it is larger than the compression window (32KB), and also
129 // large enough to serve all typical value sizes we want to write.
130 Random rnd(301);
131 std::string piece;
132 while (data_.size() < 1048576) {
133 // Add a short fragment that is as compressible as specified
134 // by FLAGS_compression_ratio.
135 test::CompressibleString(&rnd, FLAGS_compression_ratio, 100, &piece);
136 data_.append(piece);
137 }
138 pos_ = 0;
139 }
140
141 Slice Generate(size_t len) {
142 if (pos_ + len > data_.size()) {
143 pos_ = 0;
144 assert(len < data_.size());
145 }
146 pos_ += len;
147 return Slice(data_.data() + pos_ - len, len);
148 }
149 };
150
151 #if defined(__linux)
152 static Slice TrimSpace(Slice s) {
153 size_t start = 0;
154 while (start < s.size() && isspace(s[start])) {
155 start++;
156 }
157 size_t limit = s.size();
158 while (limit > start && isspace(s[limit - 1])) {
159 limit--;
160 }
161 return Slice(s.data() + start, limit - start);
162 }
163 #endif
164
165 static void AppendWithSpace(std::string* str, Slice msg) {
166 if (msg.empty()) return;
167 if (!str->empty()) {
168 str->push_back(' ');
169 }
170 str->append(msg.data(), msg.size());
171 }
172
173 class Stats {
174 private:
175 double start_;
176 double finish_;
177 double seconds_;
178 int done_;
179 int next_report_;
180 int64_t bytes_;
181 double last_op_finish_;
182 Histogram hist_;
183 std::string message_;
184
185 public:
186 Stats() { Start(); }
187
188 void Start() {
189 next_report_ = 100;
190 hist_.Clear();
191 done_ = 0;
192 bytes_ = 0;
193 seconds_ = 0;
194 message_.clear();
195 start_ = finish_ = last_op_finish_ = g_env->NowMicros();
196 }
197
198 void Merge(const Stats& other) {
199 hist_.Merge(other.hist_);
200 done_ += other.done_;
201 bytes_ += other.bytes_;
202 seconds_ += other.seconds_;
203 if (other.start_ < start_) start_ = other.start_;
204 if (other.finish_ > finish_) finish_ = other.finish_;
205
206 // Just keep the messages from one thread
207 if (message_.empty()) message_ = other.message_;
208 }
209
210 void Stop() {
211 finish_ = g_env->NowMicros();
212 seconds_ = (finish_ - start_) * 1e-6;
213 }
214
215 void AddMessage(Slice msg) { AppendWithSpace(&message_, msg); }
216
217 void FinishedSingleOp() {
218 if (FLAGS_histogram) {
219 double now = g_env->NowMicros();
220 double micros = now - last_op_finish_;
221 hist_.Add(micros);
222 if (micros > 20000) {
223 fprintf(stderr, "long op: %.1f micros%30s\r", micros, "");
224 fflush(stderr);
225 }
226 last_op_finish_ = now;
227 }
228
229 done_++;
230 if (done_ >= next_report_) {
231 if (next_report_ < 1000)
232 next_report_ += 100;
233 else if (next_report_ < 5000)
234 next_report_ += 500;
235 else if (next_report_ < 10000)
236 next_report_ += 1000;
237 else if (next_report_ < 50000)
238 next_report_ += 5000;
239 else if (next_report_ < 100000)
240 next_report_ += 10000;
241 else if (next_report_ < 500000)
242 next_report_ += 50000;
243 else
244 next_report_ += 100000;
245 fprintf(stderr, "... finished %d ops%30s\r", done_, "");
246 fflush(stderr);
247 }
248 }
249
250 void AddBytes(int64_t n) { bytes_ += n; }
251
252 void Report(const Slice& name) {
253 // Pretend at least one op was done in case we are running a benchmark
254 // that does not call FinishedSingleOp().
255 if (done_ < 1) done_ = 1;
256
257 std::string extra;
258 if (bytes_ > 0) {
259 // Rate is computed on actual elapsed time, not the sum of per-thread
260 // elapsed times.
261 double elapsed = (finish_ - start_) * 1e-6;
262 char rate[100];
263 snprintf(rate, sizeof(rate), "%6.1f MB/s",
264 (bytes_ / 1048576.0) / elapsed);
265 extra = rate;
266 }
267 AppendWithSpace(&extra, message_);
268
269 fprintf(stdout, "%-12s : %11.3f micros/op;%s%s\n", name.ToString().c_str(),
270 seconds_ * 1e6 / done_, (extra.empty() ? "" : " "), extra.c_str());
271 if (FLAGS_histogram) {
272 fprintf(stdout, "Microseconds per op:\n%s\n", hist_.ToString().c_str());
273 }
274 fflush(stdout);
275 }
276 };
277
278 // State shared by all concurrent executions of the same benchmark.
279 struct SharedState {
280 port::Mutex mu;
281 port::CondVar cv GUARDED_BY(mu);
282 int total GUARDED_BY(mu);
283
284 // Each thread goes through the following states:
285 // (1) initializing
286 // (2) waiting for others to be initialized
287 // (3) running
288 // (4) done
289
290 int num_initialized GUARDED_BY(mu);
291 int num_done GUARDED_BY(mu);
292 bool start GUARDED_BY(mu);
293
294 SharedState(int total)
295 : cv(&mu), total(total), num_initialized(0), num_done(0), start(false) {}
296 };
297
298 // Per-thread state for concurrent executions of the same benchmark.
299 struct ThreadState {
300 int tid; // 0..n-1 when running in n threads
301 Random rand; // Has different seeds for different threads
302 Stats stats;
303 SharedState* shared;
304
305 ThreadState(int index) : tid(index), rand(1000 + index), shared(nullptr) {}
306 };
307
308 } // namespace
309
310 class Benchmark {
311 private:
312 Cache* cache_;
313 const FilterPolicy* filter_policy_;
314 DB* db_;
315 int num_;
316 int value_size_;
317 int entries_per_batch_;
318 WriteOptions write_options_;
319 int reads_;
320 int heap_counter_;
321
322 void PrintHeader() {
323 const int kKeySize = 16;
324 PrintEnvironment();
325 fprintf(stdout, "Keys: %d bytes each\n", kKeySize);
326 fprintf(stdout, "Values: %d bytes each (%d bytes after compression)\n",
327 FLAGS_value_size,
328 static_cast<int>(FLAGS_value_size * FLAGS_compression_ratio + 0.5));
329 fprintf(stdout, "Entries: %d\n", num_);
330 fprintf(stdout, "RawSize: %.1f MB (estimated)\n",
331 ((static_cast<int64_t>(kKeySize + FLAGS_value_size) * num_) /
332 1048576.0));
333 fprintf(stdout, "FileSize: %.1f MB (estimated)\n",
334 (((kKeySize + FLAGS_value_size * FLAGS_compression_ratio) * num_) /
335 1048576.0));
336 PrintWarnings();
337 fprintf(stdout, "------------------------------------------------\n");
338 }
339
340 void PrintWarnings() {
341 #if defined(__GNUC__) && !defined(__OPTIMIZE__)
342 fprintf(
343 stdout,
344 "WARNING: Optimization is disabled: benchmarks unnecessarily slow\n");
345 #endif
346 #ifndef NDEBUG
347 fprintf(stdout,
348 "WARNING: Assertions are enabled; benchmarks unnecessarily slow\n");
349 #endif
350
351 // See if snappy is working by attempting to compress a compressible string
352 const char text[] = "yyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyy";
353 std::string compressed;
354 if (!port::Snappy_Compress(text, sizeof(text), &compressed)) {
355 fprintf(stdout, "WARNING: Snappy compression is not enabled\n");
356 } else if (compressed.size() >= sizeof(text)) {
357 fprintf(stdout, "WARNING: Snappy compression is not effective\n");
358 }
359 }
360
361 void PrintEnvironment() {
362 fprintf(stderr, "LevelDB: version %d.%d\n", kMajorVersion,
363 kMinorVersion);
364
365 #if defined(__linux)
366 time_t now = time(nullptr);
367 fprintf(stderr, "Date: %s", ctime(&now)); // ctime() adds newline
368
369 FILE* cpuinfo = fopen("/proc/cpuinfo", "r");
370 if (cpuinfo != nullptr) {
371 char line[1000];
372 int num_cpus = 0;
373 std::string cpu_type;
374 std::string cache_size;
375 while (fgets(line, sizeof(line), cpuinfo) != nullptr) {
376 const char* sep = strchr(line, ':');
377 if (sep == nullptr) {
378 continue;
379 }
380 Slice key = TrimSpace(Slice(line, sep - 1 - line));
381 Slice val = TrimSpace(Slice(sep + 1));
382 if (key == "model name") {
383 ++num_cpus;
384 cpu_type = val.ToString();
385 } else if (key == "cache size") {
386 cache_size = val.ToString();
387 }
388 }
389 fclose(cpuinfo);
390 fprintf(stderr, "CPU: %d * %s\n", num_cpus, cpu_type.c_str());
391 fprintf(stderr, "CPUCache: %s\n", cache_size.c_str());
392 }
393 #endif
394 }
395
396 public:
397 Benchmark()
398 : cache_(FLAGS_cache_size >= 0 ? NewLRUCache(FLAGS_cache_size) : nullptr),
399 filter_policy_(FLAGS_bloom_bits >= 0
400 ? NewBloomFilterPolicy(FLAGS_bloom_bits)
401 : nullptr),
402 db_(nullptr),
403 num_(FLAGS_num),
404 value_size_(FLAGS_value_size),
405 entries_per_batch_(1),
406 reads_(FLAGS_reads < 0 ? FLAGS_num : FLAGS_reads),
407 heap_counter_(0) {
408 std::vector<std::string> files;
409 g_env->GetChildren(FLAGS_db, &files);
410 for (size_t i = 0; i < files.size(); i++) {
411 if (Slice(files[i]).starts_with("heap-")) {
412 g_env->DeleteFile(std::string(FLAGS_db) + "/" + files[i]);
413 }
414 }
415 if (!FLAGS_use_existing_db) {
416 DestroyDB(FLAGS_db, Options());
417 }
418 }
419
420 ~Benchmark() {
421 delete db_;
422 delete cache_;
423 delete filter_policy_;
424 }
425
426 void Run() {
427 PrintHeader();
428 Open();
429
430 const char* benchmarks = FLAGS_benchmarks;
431 while (benchmarks != nullptr) {
432 const char* sep = strchr(benchmarks, ',');
433 Slice name;
434 if (sep == nullptr) {
435 name = benchmarks;
436 benchmarks = nullptr;
437 } else {
438 name = Slice(benchmarks, sep - benchmarks);
439 benchmarks = sep + 1;
440 }
441
442 // Reset parameters that may be overridden below
443 num_ = FLAGS_num;
444 reads_ = (FLAGS_reads < 0 ? FLAGS_num : FLAGS_reads);
445 value_size_ = FLAGS_value_size;
446 entries_per_batch_ = 1;
447 write_options_ = WriteOptions();
448
449 void (Benchmark::*method)(ThreadState*) = nullptr;
450 bool fresh_db = false;
451 int num_threads = FLAGS_threads;
452
453 if (name == Slice("open")) {
454 method = &Benchmark::OpenBench;
455 num_ /= 10000;
456 if (num_ < 1) num_ = 1;
457 } else if (name == Slice("fillseq")) {
458 fresh_db = true;
459 method = &Benchmark::WriteSeq;
460 } else if (name == Slice("fillbatch")) {
461 fresh_db = true;
462 entries_per_batch_ = 1000;
463 method = &Benchmark::WriteSeq;
464 } else if (name == Slice("fillrandom")) {
465 fresh_db = true;
466 method = &Benchmark::WriteRandom;
467 } else if (name == Slice("overwrite")) {
468 fresh_db = false;
469 method = &Benchmark::WriteRandom;
470 } else if (name == Slice("fillsync")) {
471 fresh_db = true;
472 num_ /= 1000;
473 write_options_.sync = true;
474 method = &Benchmark::WriteRandom;
475 } else if (name == Slice("fill100K")) {
476 fresh_db = true;
477 num_ /= 1000;
478 value_size_ = 100 * 1000;
479 method = &Benchmark::WriteRandom;
480 } else if (name == Slice("readseq")) {
481 method = &Benchmark::ReadSequential;
482 } else if (name == Slice("readreverse")) {
483 method = &Benchmark::ReadReverse;
484 } else if (name == Slice("readrandom")) {
485 method = &Benchmark::ReadRandom;
486 } else if (name == Slice("readmissing")) {
487 method = &Benchmark::ReadMissing;
488 } else if (name == Slice("seekrandom")) {
489 method = &Benchmark::SeekRandom;
490 } else if (name == Slice("readhot")) {
491 method = &Benchmark::ReadHot;
492 } else if (name == Slice("readrandomsmall")) {
493 reads_ /= 1000;
494 method = &Benchmark::ReadRandom;
495 } else if (name == Slice("deleteseq")) {
496 method = &Benchmark::DeleteSeq;
497 } else if (name == Slice("deleterandom")) {
498 method = &Benchmark::DeleteRandom;
499 } else if (name == Slice("readwhilewriting")) {
500 num_threads++; // Add extra thread for writing
501 method = &Benchmark::ReadWhileWriting;
502 } else if (name == Slice("compact")) {
503 method = &Benchmark::Compact;
504 } else if (name == Slice("crc32c")) {
505 method = &Benchmark::Crc32c;
506 } else if (name == Slice("snappycomp")) {
507 method = &Benchmark::SnappyCompress;
508 } else if (name == Slice("snappyuncomp")) {
509 method = &Benchmark::SnappyUncompress;
510 } else if (name == Slice("heapprofile")) {
511 HeapProfile();
512 } else if (name == Slice("stats")) {
513 PrintStats("leveldb.stats");
514 } else if (name == Slice("sstables")) {
515 PrintStats("leveldb.sstables");
516 } else {
517 if (!name.empty()) { // No error message for empty name
518 fprintf(stderr, "unknown benchmark '%s'\n", name.ToString().c_str());
519 }
520 }
521
522 if (fresh_db) {
523 if (FLAGS_use_existing_db) {
524 fprintf(stdout, "%-12s : skipped (--use_existing_db is true)\n",
525 name.ToString().c_str());
526 method = nullptr;
527 } else {
528 delete db_;
529 db_ = nullptr;
530 DestroyDB(FLAGS_db, Options());
531 Open();
532 }
533 }
534
535 if (method != nullptr) {
536 RunBenchmark(num_threads, name, method);
537 }
538 }
539 }
540
541 private:
542 struct ThreadArg {
543 Benchmark* bm;
544 SharedState* shared;
545 ThreadState* thread;
546 void (Benchmark::*method)(ThreadState*);
547 };
548
549 static void ThreadBody(void* v) {
550 ThreadArg* arg = reinterpret_cast<ThreadArg*>(v);
551 SharedState* shared = arg->shared;
552 ThreadState* thread = arg->thread;
553 {
554 MutexLock l(&shared->mu);
555 shared->num_initialized++;
556 if (shared->num_initialized >= shared->total) {
557 shared->cv.SignalAll();
558 }
559 while (!shared->start) {
560 shared->cv.Wait();
561 }
562 }
563
564 thread->stats.Start();
565 (arg->bm->*(arg->method))(thread);
566 thread->stats.Stop();
567
568 {
569 MutexLock l(&shared->mu);
570 shared->num_done++;
571 if (shared->num_done >= shared->total) {
572 shared->cv.SignalAll();
573 }
574 }
575 }
576
577 void RunBenchmark(int n, Slice name,
578 void (Benchmark::*method)(ThreadState*)) {
579 SharedState shared(n);
580
581 ThreadArg* arg = new ThreadArg[n];
582 for (int i = 0; i < n; i++) {
583 arg[i].bm = this;
584 arg[i].method = method;
585 arg[i].shared = &shared;
586 arg[i].thread = new ThreadState(i);
587 arg[i].thread->shared = &shared;
588 g_env->StartThread(ThreadBody, &arg[i]);
589 }
590
591 shared.mu.Lock();
592 while (shared.num_initialized < n) {
593 shared.cv.Wait();
594 }
595
596 shared.start = true;
597 shared.cv.SignalAll();
598 while (shared.num_done < n) {
599 shared.cv.Wait();
600 }
601 shared.mu.Unlock();
602
603 for (int i = 1; i < n; i++) {
604 arg[0].thread->stats.Merge(arg[i].thread->stats);
605 }
606 arg[0].thread->stats.Report(name);
607
608 for (int i = 0; i < n; i++) {
609 delete arg[i].thread;
610 }
611 delete[] arg;
612 }
613
614 void Crc32c(ThreadState* thread) {
615 // Checksum about 500MB of data total
616 const int size = 4096;
617 const char* label = "(4K per op)";
618 std::string data(size, 'x');
619 int64_t bytes = 0;
620 uint32_t crc = 0;
621 while (bytes < 500 * 1048576) {
622 crc = crc32c::Value(data.data(), size);
623 thread->stats.FinishedSingleOp();
624 bytes += size;
625 }
626 // Print so result is not dead
627 fprintf(stderr, "... crc=0x%x\r", static_cast<unsigned int>(crc));
628
629 thread->stats.AddBytes(bytes);
630 thread->stats.AddMessage(label);
631 }
632
633 void SnappyCompress(ThreadState* thread) {
634 RandomGenerator gen;
635 Slice input = gen.Generate(Options().block_size);
636 int64_t bytes = 0;
637 int64_t produced = 0;
638 bool ok = true;
639 std::string compressed;
640 while (ok && bytes < 1024 * 1048576) { // Compress 1G
641 ok = port::Snappy_Compress(input.data(), input.size(), &compressed);
642 produced += compressed.size();
643 bytes += input.size();
644 thread->stats.FinishedSingleOp();
645 }
646
647 if (!ok) {
648 thread->stats.AddMessage("(snappy failure)");
649 } else {
650 char buf[100];
651 snprintf(buf, sizeof(buf), "(output: %.1f%%)",
652 (produced * 100.0) / bytes);
653 thread->stats.AddMessage(buf);
654 thread->stats.AddBytes(bytes);
655 }
656 }
657
658 void SnappyUncompress(ThreadState* thread) {
659 RandomGenerator gen;
660 Slice input = gen.Generate(Options().block_size);
661 std::string compressed;
662 bool ok = port::Snappy_Compress(input.data(), input.size(), &compressed);
663 int64_t bytes = 0;
664 char* uncompressed = new char[input.size()];
665 while (ok && bytes < 1024 * 1048576) { // Compress 1G
666 ok = port::Snappy_Uncompress(compressed.data(), compressed.size(),
667 uncompressed);
668 bytes += input.size();
669 thread->stats.FinishedSingleOp();
670 }
671 delete[] uncompressed;
672
673 if (!ok) {
674 thread->stats.AddMessage("(snappy failure)");
675 } else {
676 thread->stats.AddBytes(bytes);
677 }
678 }
679
680 void Open() {
681 assert(db_ == nullptr);
682 Options options;
683 options.env = g_env;
684 options.create_if_missing = !FLAGS_use_existing_db;
685 options.block_cache = cache_;
686 options.write_buffer_size = FLAGS_write_buffer_size;
687 options.max_file_size = FLAGS_max_file_size;
688 options.block_size = FLAGS_block_size;
689 options.max_open_files = FLAGS_open_files;
690 options.filter_policy = filter_policy_;
691 options.reuse_logs = FLAGS_reuse_logs;
692 Status s = DB::Open(options, FLAGS_db, &db_);
693 if (!s.ok()) {
694 fprintf(stderr, "open error: %s\n", s.ToString().c_str());
695 exit(1);
696 }
697 }
698
699 void OpenBench(ThreadState* thread) {
700 for (int i = 0; i < num_; i++) {
701 delete db_;
702 Open();
703 thread->stats.FinishedSingleOp();
704 }
705 }
706
707 void WriteSeq(ThreadState* thread) { DoWrite(thread, true); }
708
709 void WriteRandom(ThreadState* thread) { DoWrite(thread, false); }
710
711 void DoWrite(ThreadState* thread, bool seq) {
712 if (num_ != FLAGS_num) {
713 char msg[100];
714 snprintf(msg, sizeof(msg), "(%d ops)", num_);
715 thread->stats.AddMessage(msg);
716 }
717
718 RandomGenerator gen;
719 WriteBatch batch;
720 Status s;
721 int64_t bytes = 0;
722 for (int i = 0; i < num_; i += entries_per_batch_) {
723 batch.Clear();
724 for (int j = 0; j < entries_per_batch_; j++) {
725 const int k = seq ? i + j : (thread->rand.Next() % FLAGS_num);
726 char key[100];
727 snprintf(key, sizeof(key), "%016d", k);
728 batch.Put(key, gen.Generate(value_size_));
729 bytes += value_size_ + strlen(key);
730 thread->stats.FinishedSingleOp();
731 }
732 s = db_->Write(write_options_, &batch);
733 if (!s.ok()) {
734 fprintf(stderr, "put error: %s\n", s.ToString().c_str());
735 exit(1);
736 }
737 }
738 thread->stats.AddBytes(bytes);
739 }
740
741 void ReadSequential(ThreadState* thread) {
742 Iterator* iter = db_->NewIterator(ReadOptions());
743 int i = 0;
744 int64_t bytes = 0;
745 for (iter->SeekToFirst(); i < reads_ && iter->Valid(); iter->Next()) {
746 bytes += iter->key().size() + iter->value().size();
747 thread->stats.FinishedSingleOp();
748 ++i;
749 }
750 delete iter;
751 thread->stats.AddBytes(bytes);
752 }
753
754 void ReadReverse(ThreadState* thread) {
755 Iterator* iter = db_->NewIterator(ReadOptions());
756 int i = 0;
757 int64_t bytes = 0;
758 for (iter->SeekToLast(); i < reads_ && iter->Valid(); iter->Prev()) {
759 bytes += iter->key().size() + iter->value().size();
760 thread->stats.FinishedSingleOp();
761 ++i;
762 }
763 delete iter;
764 thread->stats.AddBytes(bytes);
765 }
766
767 void ReadRandom(ThreadState* thread) {
768 ReadOptions options;
769 std::string value;
770 int found = 0;
771 for (int i = 0; i < reads_; i++) {
772 char key[100];
773 const int k = thread->rand.Next() % FLAGS_num;
774 snprintf(key, sizeof(key), "%016d", k);
775 if (db_->Get(options, key, &value).ok()) {
776 found++;
777 }
778 thread->stats.FinishedSingleOp();
779 }
780 char msg[100];
781 snprintf(msg, sizeof(msg), "(%d of %d found)", found, num_);
782 thread->stats.AddMessage(msg);
783 }
784
785 void ReadMissing(ThreadState* thread) {
786 ReadOptions options;
787 std::string value;
788 for (int i = 0; i < reads_; i++) {
789 char key[100];
790 const int k = thread->rand.Next() % FLAGS_num;
791 snprintf(key, sizeof(key), "%016d.", k);
792 db_->Get(options, key, &value);
793 thread->stats.FinishedSingleOp();
794 }
795 }
796
797 void ReadHot(ThreadState* thread) {
798 ReadOptions options;
799 std::string value;
800 const int range = (FLAGS_num + 99) / 100;
801 for (int i = 0; i < reads_; i++) {
802 char key[100];
803 const int k = thread->rand.Next() % range;
804 snprintf(key, sizeof(key), "%016d", k);
805 db_->Get(options, key, &value);
806 thread->stats.FinishedSingleOp();
807 }
808 }
809
810 void SeekRandom(ThreadState* thread) {
811 ReadOptions options;
812 int found = 0;
813 for (int i = 0; i < reads_; i++) {
814 Iterator* iter = db_->NewIterator(options);
815 char key[100];
816 const int k = thread->rand.Next() % FLAGS_num;
817 snprintf(key, sizeof(key), "%016d", k);
818 iter->Seek(key);
819 if (iter->Valid() && iter->key() == key) found++;
820 delete iter;
821 thread->stats.FinishedSingleOp();
822 }
823 char msg[100];
824 snprintf(msg, sizeof(msg), "(%d of %d found)", found, num_);
825 thread->stats.AddMessage(msg);
826 }
827
828 void DoDelete(ThreadState* thread, bool seq) {
829 RandomGenerator gen;
830 WriteBatch batch;
831 Status s;
832 for (int i = 0; i < num_; i += entries_per_batch_) {
833 batch.Clear();
834 for (int j = 0; j < entries_per_batch_; j++) {
835 const int k = seq ? i + j : (thread->rand.Next() % FLAGS_num);
836 char key[100];
837 snprintf(key, sizeof(key), "%016d", k);
838 batch.Delete(key);
839 thread->stats.FinishedSingleOp();
840 }
841 s = db_->Write(write_options_, &batch);
842 if (!s.ok()) {
843 fprintf(stderr, "del error: %s\n", s.ToString().c_str());
844 exit(1);
845 }
846 }
847 }
848
849 void DeleteSeq(ThreadState* thread) { DoDelete(thread, true); }
850
851 void DeleteRandom(ThreadState* thread) { DoDelete(thread, false); }
852
853 void ReadWhileWriting(ThreadState* thread) {
854 if (thread->tid > 0) {
855 ReadRandom(thread);
856 } else {
857 // Special thread that keeps writing until other threads are done.
858 RandomGenerator gen;
859 while (true) {
860 {
861 MutexLock l(&thread->shared->mu);
862 if (thread->shared->num_done + 1 >= thread->shared->num_initialized) {
863 // Other threads have finished
864 break;
865 }
866 }
867
868 const int k = thread->rand.Next() % FLAGS_num;
869 char key[100];
870 snprintf(key, sizeof(key), "%016d", k);
871 Status s = db_->Put(write_options_, key, gen.Generate(value_size_));
872 if (!s.ok()) {
873 fprintf(stderr, "put error: %s\n", s.ToString().c_str());
874 exit(1);
875 }
876 }
877
878 // Do not count any of the preceding work/delay in stats.
879 thread->stats.Start();
880 }
881 }
882
883 void Compact(ThreadState* thread) { db_->CompactRange(nullptr, nullptr); }
884
885 void PrintStats(const char* key) {
886 std::string stats;
887 if (!db_->GetProperty(key, &stats)) {
888 stats = "(failed)";
889 }
890 fprintf(stdout, "\n%s\n", stats.c_str());
891 }
892
893 static void WriteToFile(void* arg, const char* buf, int n) {
894 reinterpret_cast<WritableFile*>(arg)->Append(Slice(buf, n));
895 }
896
897 void HeapProfile() {
898 char fname[100];
899 snprintf(fname, sizeof(fname), "%s/heap-%04d", FLAGS_db, ++heap_counter_);
900 WritableFile* file;
901 Status s = g_env->NewWritableFile(fname, &file);
902 if (!s.ok()) {
903 fprintf(stderr, "%s\n", s.ToString().c_str());
904 return;
905 }
906 bool ok = port::GetHeapProfile(WriteToFile, file);
907 delete file;
908 if (!ok) {
909 fprintf(stderr, "heap profiling not supported\n");
910 g_env->DeleteFile(fname);
911 }
912 }
913 };
914
915 } // namespace leveldb
916
917 int main(int argc, char** argv) {
918 FLAGS_write_buffer_size = leveldb::Options().write_buffer_size;
919 FLAGS_max_file_size = leveldb::Options().max_file_size;
920 FLAGS_block_size = leveldb::Options().block_size;
921 FLAGS_open_files = leveldb::Options().max_open_files;
922 std::string default_db_path;
923
924 for (int i = 1; i < argc; i++) {
925 double d;
926 int n;
927 char junk;
928 if (leveldb::Slice(argv[i]).starts_with("--benchmarks=")) {
929 FLAGS_benchmarks = argv[i] + strlen("--benchmarks=");
930 } else if (sscanf(argv[i], "--compression_ratio=%lf%c", &d, &junk) == 1) {
931 FLAGS_compression_ratio = d;
932 } else if (sscanf(argv[i], "--histogram=%d%c", &n, &junk) == 1 &&
933 (n == 0 || n == 1)) {
934 FLAGS_histogram = n;
935 } else if (sscanf(argv[i], "--use_existing_db=%d%c", &n, &junk) == 1 &&
936 (n == 0 || n == 1)) {
937 FLAGS_use_existing_db = n;
938 } else if (sscanf(argv[i], "--reuse_logs=%d%c", &n, &junk) == 1 &&
939 (n == 0 || n == 1)) {
940 FLAGS_reuse_logs = n;
941 } else if (sscanf(argv[i], "--num=%d%c", &n, &junk) == 1) {
942 FLAGS_num = n;
943 } else if (sscanf(argv[i], "--reads=%d%c", &n, &junk) == 1) {
944 FLAGS_reads = n;
945 } else if (sscanf(argv[i], "--threads=%d%c", &n, &junk) == 1) {
946 FLAGS_threads = n;
947 } else if (sscanf(argv[i], "--value_size=%d%c", &n, &junk) == 1) {
948 FLAGS_value_size = n;
949 } else if (sscanf(argv[i], "--write_buffer_size=%d%c", &n, &junk) == 1) {
950 FLAGS_write_buffer_size = n;
951 } else if (sscanf(argv[i], "--max_file_size=%d%c", &n, &junk) == 1) {
952 FLAGS_max_file_size = n;
953 } else if (sscanf(argv[i], "--block_size=%d%c", &n, &junk) == 1) {
954 FLAGS_block_size = n;
955 } else if (sscanf(argv[i], "--cache_size=%d%c", &n, &junk) == 1) {
956 FLAGS_cache_size = n;
957 } else if (sscanf(argv[i], "--bloom_bits=%d%c", &n, &junk) == 1) {
958 FLAGS_bloom_bits = n;
959 } else if (sscanf(argv[i], "--open_files=%d%c", &n, &junk) == 1) {
960 FLAGS_open_files = n;
961 } else if (strncmp(argv[i], "--db=", 5) == 0) {
962 FLAGS_db = argv[i] + 5;
963 } else {
964 fprintf(stderr, "Invalid flag '%s'\n", argv[i]);
965 exit(1);
966 }
967 }
968
969 leveldb::g_env = leveldb::Env::Default();
970
971 // Choose a location for the test database if none given with --db=<path>
972 if (FLAGS_db == nullptr) {
973 leveldb::g_env->GetTestDirectory(&default_db_path);
974 default_db_path += "/dbbench";
975 FLAGS_db = default_db_path.c_str();
976 }
977
978 leveldb::Benchmark benchmark;
979 benchmark.Run();
980 return 0;
981 }
982