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.
8
#include "db/db_impl.h"
9
#include "db/version_set.h"
10
#include "leveldb/cache.h"
11
#include "leveldb/db.h"
12
#include "leveldb/env.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"
21
// Comma-separated list of operations to run in the specified order
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
// readseq -- read N times sequentially
29
// readreverse -- read N times in reverse order
30
// readrandom -- read N times in random order
31
// readhot -- read N times in random order from 1% section of DB
32
// crc32c -- repeated crc32c of 4K of data
33
// acquireload -- load N*1000 times
35
// compact -- Compact the entire DB
36
// stats -- Print DB stats
37
// heapprofile -- Dump a heap profile (if supported by this port)
38
static const char* FLAGS_benchmarks =
44
"readrandom," // Extra run to allow previous compactions to quiesce
58
// Number of key/values to place in database
59
static int FLAGS_num = 1000000;
61
// Number of read operations to do. If negative, do FLAGS_num reads.
62
static int FLAGS_reads = -1;
64
// Number of concurrent threads to run.
65
static int FLAGS_threads = 1;
68
static int FLAGS_value_size = 100;
70
// Arrange to generate values that shrink to this fraction of
71
// their original size after compression
72
static double FLAGS_compression_ratio = 0.5;
74
// Print histogram of operation timings
75
static bool FLAGS_histogram = false;
77
// Number of bytes to buffer in memtable before compacting
78
// (initialized to default value by "main")
79
static int FLAGS_write_buffer_size = 0;
81
// Number of bytes to use as a cache of uncompressed data.
82
// Negative means use default settings.
83
static int FLAGS_cache_size = -1;
85
// Maximum number of files to keep open at the same time (use default if == 0)
86
static int FLAGS_open_files = 0;
88
// If true, do not destroy the existing database. If you set this
89
// flag and also specify a benchmark that wants a fresh database, that
90
// benchmark will fail.
91
static bool FLAGS_use_existing_db = false;
93
// Use the db with the following name.
94
static const char* FLAGS_db = "/tmp/dbbench";
100
// Helper for quickly generating random data.
101
class RandomGenerator {
108
// We use a limited amount of data over and over again and ensure
109
// that it is larger than the compression window (32KB), and also
110
// large enough to serve all typical value sizes we want to write.
113
while (data_.size() < 1048576) {
114
// Add a short fragment that is as compressible as specified
115
// by FLAGS_compression_ratio.
116
test::CompressibleString(&rnd, FLAGS_compression_ratio, 100, &piece);
122
Slice Generate(int len) {
123
if (pos_ + len > data_.size()) {
125
assert(len < data_.size());
128
return Slice(data_.data() + pos_ - len, len);
132
static Slice TrimSpace(Slice s) {
134
while (start < s.size() && isspace(s[start])) {
137
int limit = s.size();
138
while (limit > start && isspace(s[limit-1])) {
141
return Slice(s.data() + start, limit - start);
144
static void AppendWithSpace(std::string* str, Slice msg) {
145
if (msg.empty()) return;
149
str->append(msg.data(), msg.size());
160
double last_op_finish_;
162
std::string message_;
169
last_op_finish_ = start_;
174
start_ = Env::Default()->NowMicros();
179
void Merge(const Stats& other) {
180
hist_.Merge(other.hist_);
181
done_ += other.done_;
182
bytes_ += other.bytes_;
183
seconds_ += other.seconds_;
184
if (other.start_ < start_) start_ = other.start_;
185
if (other.finish_ > finish_) finish_ = other.finish_;
187
// Just keep the messages from one thread
188
if (message_.empty()) message_ = other.message_;
192
finish_ = Env::Default()->NowMicros();
193
seconds_ = (finish_ - start_) * 1e-6;
196
void AddMessage(Slice msg) {
197
AppendWithSpace(&message_, msg);
200
void FinishedSingleOp() {
201
if (FLAGS_histogram) {
202
double now = Env::Default()->NowMicros();
203
double micros = now - last_op_finish_;
205
if (micros > 20000) {
206
fprintf(stderr, "long op: %.1f micros%30s\r", micros, "");
209
last_op_finish_ = now;
213
if (done_ >= next_report_) {
214
if (next_report_ < 1000) next_report_ += 100;
215
else if (next_report_ < 5000) next_report_ += 500;
216
else if (next_report_ < 10000) next_report_ += 1000;
217
else if (next_report_ < 50000) next_report_ += 5000;
218
else if (next_report_ < 100000) next_report_ += 10000;
219
else if (next_report_ < 500000) next_report_ += 50000;
220
else next_report_ += 100000;
221
fprintf(stderr, "... finished %d ops%30s\r", done_, "");
226
void AddBytes(int64_t n) {
230
void Report(const Slice& name) {
231
// Pretend at least one op was done in case we are running a benchmark
232
// that does not call FinishedSingleOp().
233
if (done_ < 1) done_ = 1;
237
// Rate is computed on actual elapsed time, not the sum of per-thread
239
double elapsed = (finish_ - start_) * 1e-6;
241
snprintf(rate, sizeof(rate), "%6.1f MB/s",
242
(bytes_ / 1048576.0) / elapsed);
245
AppendWithSpace(&extra, message_);
247
fprintf(stdout, "%-12s : %11.3f micros/op;%s%s\n",
248
name.ToString().c_str(),
249
seconds_ * 1e6 / done_,
250
(extra.empty() ? "" : " "),
252
if (FLAGS_histogram) {
253
fprintf(stdout, "Microseconds per op:\n%s\n", hist_.ToString().c_str());
259
// State shared by all concurrent executions of the same benchmark.
265
// Each thread goes through the following states:
267
// (2) waiting for others to be initialized
275
SharedState() : cv(&mu) { }
278
// Per-thread state for concurrent executions of the same benchmark.
280
int tid; // 0..n-1 when running in n threads
281
Random rand; // Has different seeds for different threads
285
ThreadState(int index)
299
int entries_per_batch_;
300
WriteOptions write_options_;
305
const int kKeySize = 16;
307
fprintf(stdout, "Keys: %d bytes each\n", kKeySize);
308
fprintf(stdout, "Values: %d bytes each (%d bytes after compression)\n",
310
static_cast<int>(FLAGS_value_size * FLAGS_compression_ratio + 0.5));
311
fprintf(stdout, "Entries: %d\n", num_);
312
fprintf(stdout, "RawSize: %.1f MB (estimated)\n",
313
((static_cast<int64_t>(kKeySize + FLAGS_value_size) * num_)
315
fprintf(stdout, "FileSize: %.1f MB (estimated)\n",
316
(((kKeySize + FLAGS_value_size * FLAGS_compression_ratio) * num_)
319
fprintf(stdout, "------------------------------------------------\n");
322
void PrintWarnings() {
323
#if defined(__GNUC__) && !defined(__OPTIMIZE__)
325
"WARNING: Optimization is disabled: benchmarks unnecessarily slow\n"
330
"WARNING: Assertions are enabled; benchmarks unnecessarily slow\n");
333
// See if snappy is working by attempting to compress a compressible string
334
const char text[] = "yyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyy";
335
std::string compressed;
336
if (!port::Snappy_Compress(text, sizeof(text), &compressed)) {
337
fprintf(stdout, "WARNING: Snappy compression is not enabled\n");
338
} else if (compressed.size() >= sizeof(text)) {
339
fprintf(stdout, "WARNING: Snappy compression is not effective\n");
343
void PrintEnvironment() {
344
fprintf(stderr, "LevelDB: version %d.%d\n",
345
kMajorVersion, kMinorVersion);
348
time_t now = time(NULL);
349
fprintf(stderr, "Date: %s", ctime(&now)); // ctime() adds newline
351
FILE* cpuinfo = fopen("/proc/cpuinfo", "r");
352
if (cpuinfo != NULL) {
355
std::string cpu_type;
356
std::string cache_size;
357
while (fgets(line, sizeof(line), cpuinfo) != NULL) {
358
const char* sep = strchr(line, ':');
362
Slice key = TrimSpace(Slice(line, sep - 1 - line));
363
Slice val = TrimSpace(Slice(sep + 1));
364
if (key == "model name") {
366
cpu_type = val.ToString();
367
} else if (key == "cache size") {
368
cache_size = val.ToString();
372
fprintf(stderr, "CPU: %d * %s\n", num_cpus, cpu_type.c_str());
373
fprintf(stderr, "CPUCache: %s\n", cache_size.c_str());
380
: cache_(FLAGS_cache_size >= 0 ? NewLRUCache(FLAGS_cache_size) : NULL),
383
value_size_(FLAGS_value_size),
384
entries_per_batch_(1),
385
reads_(FLAGS_reads < 0 ? FLAGS_num : FLAGS_reads),
387
std::vector<std::string> files;
388
Env::Default()->GetChildren(FLAGS_db, &files);
389
for (int i = 0; i < files.size(); i++) {
390
if (Slice(files[i]).starts_with("heap-")) {
391
Env::Default()->DeleteFile(std::string(FLAGS_db) + "/" + files[i]);
394
if (!FLAGS_use_existing_db) {
395
DestroyDB(FLAGS_db, Options());
408
const char* benchmarks = FLAGS_benchmarks;
409
while (benchmarks != NULL) {
410
const char* sep = strchr(benchmarks, ',');
416
name = Slice(benchmarks, sep - benchmarks);
417
benchmarks = sep + 1;
420
// Reset parameters that may be overriddden bwlow
422
reads_ = (FLAGS_reads < 0 ? FLAGS_num : FLAGS_reads);
423
value_size_ = FLAGS_value_size;
424
entries_per_batch_ = 1;
425
write_options_ = WriteOptions();
427
void (Benchmark::*method)(ThreadState*) = NULL;
428
bool fresh_db = false;
429
int num_threads = FLAGS_threads;
431
if (name == Slice("fillseq")) {
433
method = &Benchmark::WriteSeq;
434
} else if (name == Slice("fillbatch")) {
436
entries_per_batch_ = 1000;
437
method = &Benchmark::WriteSeq;
438
} else if (name == Slice("fillrandom")) {
440
method = &Benchmark::WriteRandom;
441
} else if (name == Slice("overwrite")) {
443
method = &Benchmark::WriteRandom;
444
} else if (name == Slice("fillsync")) {
447
write_options_.sync = true;
448
method = &Benchmark::WriteRandom;
449
} else if (name == Slice("fill100K")) {
452
value_size_ = 100 * 1000;
453
method = &Benchmark::WriteRandom;
454
} else if (name == Slice("readseq")) {
455
method = &Benchmark::ReadSequential;
456
} else if (name == Slice("readreverse")) {
457
method = &Benchmark::ReadReverse;
458
} else if (name == Slice("readrandom")) {
459
method = &Benchmark::ReadRandom;
460
} else if (name == Slice("readhot")) {
461
method = &Benchmark::ReadHot;
462
} else if (name == Slice("readrandomsmall")) {
464
method = &Benchmark::ReadRandom;
465
} else if (name == Slice("readwhilewriting")) {
466
num_threads++; // Add extra thread for writing
467
method = &Benchmark::ReadWhileWriting;
468
} else if (name == Slice("compact")) {
469
method = &Benchmark::Compact;
470
} else if (name == Slice("crc32c")) {
471
method = &Benchmark::Crc32c;
472
} else if (name == Slice("acquireload")) {
473
method = &Benchmark::AcquireLoad;
474
} else if (name == Slice("snappycomp")) {
475
method = &Benchmark::SnappyCompress;
476
} else if (name == Slice("snappyuncomp")) {
477
method = &Benchmark::SnappyUncompress;
478
} else if (name == Slice("heapprofile")) {
480
} else if (name == Slice("stats")) {
483
if (name != Slice()) { // No error message for empty name
484
fprintf(stderr, "unknown benchmark '%s'\n", name.ToString().c_str());
489
if (FLAGS_use_existing_db) {
490
fprintf(stdout, "%-12s : skipped (--use_existing_db is true)\n",
491
name.ToString().c_str());
496
DestroyDB(FLAGS_db, Options());
501
if (method != NULL) {
502
RunBenchmark(num_threads, name, method);
512
void (Benchmark::*method)(ThreadState*);
515
static void ThreadBody(void* v) {
516
ThreadArg* arg = reinterpret_cast<ThreadArg*>(v);
517
SharedState* shared = arg->shared;
518
ThreadState* thread = arg->thread;
520
MutexLock l(&shared->mu);
521
shared->num_initialized++;
522
if (shared->num_initialized >= shared->total) {
523
shared->cv.SignalAll();
525
while (!shared->start) {
530
thread->stats.Start();
531
(arg->bm->*(arg->method))(thread);
532
thread->stats.Stop();
535
MutexLock l(&shared->mu);
537
if (shared->num_done >= shared->total) {
538
shared->cv.SignalAll();
543
void RunBenchmark(int n, Slice name,
544
void (Benchmark::*method)(ThreadState*)) {
547
shared.num_initialized = 0;
549
shared.start = false;
551
ThreadArg* arg = new ThreadArg[n];
552
for (int i = 0; i < n; i++) {
554
arg[i].method = method;
555
arg[i].shared = &shared;
556
arg[i].thread = new ThreadState(i);
557
arg[i].thread->shared = &shared;
558
Env::Default()->StartThread(ThreadBody, &arg[i]);
562
while (shared.num_initialized < n) {
567
shared.cv.SignalAll();
568
while (shared.num_done < n) {
573
for (int i = 1; i < n; i++) {
574
arg[0].thread->stats.Merge(arg[i].thread->stats);
576
arg[0].thread->stats.Report(name);
578
for (int i = 0; i < n; i++) {
579
delete arg[i].thread;
584
void Crc32c(ThreadState* thread) {
585
// Checksum about 500MB of data total
586
const int size = 4096;
587
const char* label = "(4K per op)";
588
std::string data(size, 'x');
591
while (bytes < 500 * 1048576) {
592
crc = crc32c::Value(data.data(), size);
593
thread->stats.FinishedSingleOp();
596
// Print so result is not dead
597
fprintf(stderr, "... crc=0x%x\r", static_cast<unsigned int>(crc));
599
thread->stats.AddBytes(bytes);
600
thread->stats.AddMessage(label);
603
void AcquireLoad(ThreadState* thread) {
605
port::AtomicPointer ap(&dummy);
608
thread->stats.AddMessage("(each op is 1000 loads)");
609
while (count < 100000) {
610
for (int i = 0; i < 1000; i++) {
611
ptr = ap.Acquire_Load();
614
thread->stats.FinishedSingleOp();
616
if (ptr == NULL) exit(1); // Disable unused variable warning.
619
void SnappyCompress(ThreadState* thread) {
621
Slice input = gen.Generate(Options().block_size);
623
int64_t produced = 0;
625
std::string compressed;
626
while (ok && bytes < 1024 * 1048576) { // Compress 1G
627
ok = port::Snappy_Compress(input.data(), input.size(), &compressed);
628
produced += compressed.size();
629
bytes += input.size();
630
thread->stats.FinishedSingleOp();
634
thread->stats.AddMessage("(snappy failure)");
637
snprintf(buf, sizeof(buf), "(output: %.1f%%)",
638
(produced * 100.0) / bytes);
639
thread->stats.AddMessage(buf);
640
thread->stats.AddBytes(bytes);
644
void SnappyUncompress(ThreadState* thread) {
646
Slice input = gen.Generate(Options().block_size);
647
std::string compressed;
648
bool ok = port::Snappy_Compress(input.data(), input.size(), &compressed);
650
char* uncompressed = new char[input.size()];
651
while (ok && bytes < 1024 * 1048576) { // Compress 1G
652
ok = port::Snappy_Uncompress(compressed.data(), compressed.size(),
654
bytes += input.size();
655
thread->stats.FinishedSingleOp();
657
delete[] uncompressed;
660
thread->stats.AddMessage("(snappy failure)");
662
thread->stats.AddBytes(bytes);
669
options.create_if_missing = !FLAGS_use_existing_db;
670
options.block_cache = cache_;
671
options.write_buffer_size = FLAGS_write_buffer_size;
672
Status s = DB::Open(options, FLAGS_db, &db_);
674
fprintf(stderr, "open error: %s\n", s.ToString().c_str());
679
void WriteSeq(ThreadState* thread) {
680
DoWrite(thread, true);
683
void WriteRandom(ThreadState* thread) {
684
DoWrite(thread, false);
687
void DoWrite(ThreadState* thread, bool seq) {
688
if (num_ != FLAGS_num) {
690
snprintf(msg, sizeof(msg), "(%d ops)", num_);
691
thread->stats.AddMessage(msg);
698
for (int i = 0; i < num_; i += entries_per_batch_) {
700
for (int j = 0; j < entries_per_batch_; j++) {
701
const int k = seq ? i+j : (thread->rand.Next() % FLAGS_num);
703
snprintf(key, sizeof(key), "%016d", k);
704
batch.Put(key, gen.Generate(value_size_));
705
bytes += value_size_ + strlen(key);
706
thread->stats.FinishedSingleOp();
708
s = db_->Write(write_options_, &batch);
710
fprintf(stderr, "put error: %s\n", s.ToString().c_str());
714
thread->stats.AddBytes(bytes);
717
void ReadSequential(ThreadState* thread) {
718
Iterator* iter = db_->NewIterator(ReadOptions());
721
for (iter->SeekToFirst(); i < reads_ && iter->Valid(); iter->Next()) {
722
bytes += iter->key().size() + iter->value().size();
723
thread->stats.FinishedSingleOp();
727
thread->stats.AddBytes(bytes);
730
void ReadReverse(ThreadState* thread) {
731
Iterator* iter = db_->NewIterator(ReadOptions());
734
for (iter->SeekToLast(); i < reads_ && iter->Valid(); iter->Prev()) {
735
bytes += iter->key().size() + iter->value().size();
736
thread->stats.FinishedSingleOp();
740
thread->stats.AddBytes(bytes);
743
void ReadRandom(ThreadState* thread) {
746
for (int i = 0; i < reads_; i++) {
748
const int k = thread->rand.Next() % FLAGS_num;
749
snprintf(key, sizeof(key), "%016d", k);
750
db_->Get(options, key, &value);
751
thread->stats.FinishedSingleOp();
755
void ReadHot(ThreadState* thread) {
758
const int range = (FLAGS_num + 99) / 100;
759
for (int i = 0; i < reads_; i++) {
761
const int k = thread->rand.Next() % range;
762
snprintf(key, sizeof(key), "%016d", k);
763
db_->Get(options, key, &value);
764
thread->stats.FinishedSingleOp();
768
void ReadWhileWriting(ThreadState* thread) {
769
if (thread->tid > 0) {
772
// Special thread that keeps writing until other threads are done.
776
MutexLock l(&thread->shared->mu);
777
if (thread->shared->num_done + 1 >= thread->shared->num_initialized) {
778
// Other threads have finished
783
const int k = thread->rand.Next() % FLAGS_num;
785
snprintf(key, sizeof(key), "%016d", k);
786
Status s = db_->Put(write_options_, key, gen.Generate(value_size_));
788
fprintf(stderr, "put error: %s\n", s.ToString().c_str());
793
// Do not count any of the preceding work/delay in stats.
794
thread->stats.Start();
798
void Compact(ThreadState* thread) {
799
db_->CompactRange(NULL, NULL);
804
if (!db_->GetProperty("leveldb.stats", &stats)) {
807
fprintf(stdout, "\n%s\n", stats.c_str());
810
static void WriteToFile(void* arg, const char* buf, int n) {
811
reinterpret_cast<WritableFile*>(arg)->Append(Slice(buf, n));
816
snprintf(fname, sizeof(fname), "%s/heap-%04d", FLAGS_db, ++heap_counter_);
818
Status s = Env::Default()->NewWritableFile(fname, &file);
820
fprintf(stderr, "%s\n", s.ToString().c_str());
823
bool ok = port::GetHeapProfile(WriteToFile, file);
826
fprintf(stderr, "heap profiling not supported\n");
827
Env::Default()->DeleteFile(fname);
832
} // namespace leveldb
834
int main(int argc, char** argv) {
835
FLAGS_write_buffer_size = leveldb::Options().write_buffer_size;
836
FLAGS_open_files = leveldb::Options().max_open_files;
838
for (int i = 1; i < argc; i++) {
842
if (leveldb::Slice(argv[i]).starts_with("--benchmarks=")) {
843
FLAGS_benchmarks = argv[i] + strlen("--benchmarks=");
844
} else if (sscanf(argv[i], "--compression_ratio=%lf%c", &d, &junk) == 1) {
845
FLAGS_compression_ratio = d;
846
} else if (sscanf(argv[i], "--histogram=%d%c", &n, &junk) == 1 &&
847
(n == 0 || n == 1)) {
849
} else if (sscanf(argv[i], "--use_existing_db=%d%c", &n, &junk) == 1 &&
850
(n == 0 || n == 1)) {
851
FLAGS_use_existing_db = n;
852
} else if (sscanf(argv[i], "--num=%d%c", &n, &junk) == 1) {
854
} else if (sscanf(argv[i], "--reads=%d%c", &n, &junk) == 1) {
856
} else if (sscanf(argv[i], "--threads=%d%c", &n, &junk) == 1) {
858
} else if (sscanf(argv[i], "--value_size=%d%c", &n, &junk) == 1) {
859
FLAGS_value_size = n;
860
} else if (sscanf(argv[i], "--write_buffer_size=%d%c", &n, &junk) == 1) {
861
FLAGS_write_buffer_size = n;
862
} else if (sscanf(argv[i], "--cache_size=%d%c", &n, &junk) == 1) {
863
FLAGS_cache_size = n;
864
} else if (sscanf(argv[i], "--open_files=%d%c", &n, &junk) == 1) {
865
FLAGS_open_files = n;
866
} else if (strncmp(argv[i], "--db=", 5) == 0) {
867
FLAGS_db = argv[i] + 5;
869
fprintf(stderr, "Invalid flag '%s'\n", argv[i]);
874
leveldb::Benchmark benchmark;
added
removed
src/leveldb/db/db_bench.cc
