db_bench.cc 48.1 KB
Newer Older
J
jorlow@chromium.org 已提交
1 2 3 4 5 6 7 8 9
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.

#include <sys/types.h>
#include <stdio.h>
#include <stdlib.h>
#include "db/db_impl.h"
#include "db/version_set.h"
10
#include "db/db_statistics.h"
11 12 13 14
#include "leveldb/cache.h"
#include "leveldb/db.h"
#include "leveldb/env.h"
#include "leveldb/write_batch.h"
15
#include "leveldb/statistics.h"
J
jorlow@chromium.org 已提交
16 17
#include "port/port.h"
#include "util/crc32c.h"
J
jorlow@chromium.org 已提交
18
#include "util/histogram.h"
19
#include "util/mutexlock.h"
J
jorlow@chromium.org 已提交
20 21
#include "util/random.h"
#include "util/testutil.h"
22
#include "hdfs/env_hdfs.h"
J
jorlow@chromium.org 已提交
23 24 25

// Comma-separated list of operations to run in the specified order
//   Actual benchmarks:
26 27 28 29 30
//      fillseq       -- write N values in sequential key order in async mode
//      fillrandom    -- write N values in random key order in async mode
//      overwrite     -- overwrite N values in random key order in async mode
//      fillsync      -- write N/100 values in random key order in sync mode
//      fill100K      -- write N/1000 100K values in random order in async mode
S
Sanjay Ghemawat 已提交
31 32
//      deleteseq     -- delete N keys in sequential order
//      deleterandom  -- delete N keys in random order
33 34 35
//      readseq       -- read N times sequentially
//      readreverse   -- read N times in reverse order
//      readrandom    -- read N times in random order
S
Sanjay Ghemawat 已提交
36
//      readmissing   -- read N missing keys in random order
37
//      readhot       -- read N times in random order from 1% section of DB
S
Sanjay Ghemawat 已提交
38
//      seekrandom    -- N random seeks
J
jorlow@chromium.org 已提交
39
//      crc32c        -- repeated crc32c of 4K of data
40
//      acquireload   -- load N*1000 times
J
jorlow@chromium.org 已提交
41 42
//   Meta operations:
//      compact     -- Compact the entire DB
43
//      stats       -- Print DB stats
S
Sanjay Ghemawat 已提交
44
//      sstables    -- Print sstable info
J
jorlow@chromium.org 已提交
45 46
//      heapprofile -- Dump a heap profile (if supported by this port)
static const char* FLAGS_benchmarks =
47
    "fillseq,"
J
jorlow@chromium.org 已提交
48
    "fillsync,"
49 50
    "fillrandom,"
    "overwrite,"
J
jorlow@chromium.org 已提交
51 52
    "readrandom,"
    "readrandom,"  // Extra run to allow previous compactions to quiesce
J
jorlow@chromium.org 已提交
53
    "readseq,"
J
jorlow@chromium.org 已提交
54
    "readreverse,"
J
jorlow@chromium.org 已提交
55
    "compact,"
J
jorlow@chromium.org 已提交
56
    "readrandom,"
J
jorlow@chromium.org 已提交
57
    "readseq,"
J
jorlow@chromium.org 已提交
58
    "readreverse,"
59
    "readrandomwriterandom," // mix reads and writes based on FLAGS_readwritepercent
J
jorlow@chromium.org 已提交
60 61
    "fill100K,"
    "crc32c,"
62 63
    "snappycomp,"
    "snappyuncomp,"
64
    "acquireload,"
J
jorlow@chromium.org 已提交
65
    ;
J
jorlow@chromium.org 已提交
66 67

// Number of key/values to place in database
68
static long FLAGS_num = 1000000;
J
jorlow@chromium.org 已提交
69

70
// Number of read operations to do.  If negative, do FLAGS_num reads.
71
static long FLAGS_reads = -1;
72

73 74 75 76 77 78
// When ==1 reads use ::Get, when >1 reads use an iterator
static long FLAGS_read_range = 1;

// Seed base for random number generators. When 0 it is deterministic.
static long FLAGS_seed = 0;

79 80 81
// Number of concurrent threads to run.
static int FLAGS_threads = 1;

J
jorlow@chromium.org 已提交
82 83 84 85 86
// Size of each value
static int FLAGS_value_size = 100;

// Arrange to generate values that shrink to this fraction of
// their original size after compression
87
static double FLAGS_compression_ratio = 0.5;
J
jorlow@chromium.org 已提交
88 89 90 91 92

// Print histogram of operation timings
static bool FLAGS_histogram = false;

// Number of bytes to buffer in memtable before compacting
93 94 95
// (initialized to default value by "main")
static int FLAGS_write_buffer_size = 0;

A
Abhishek Kona 已提交
96
// The number of in-memory memtables.
97 98 99 100 101 102 103 104 105
// Each memtable is of size FLAGS_write_buffer_size.
// This is initialized to default value of 2 in "main" function.
static int FLAGS_max_write_buffer_number = 0;

// The maximum number of concurrent background compactions
// that can occur in parallel.
// This is initialized to default value of 1 in "main" function.
static int FLAGS_max_background_compactions = 0;

106 107
// Number of bytes to use as a cache of uncompressed data.
// Negative means use default settings.
D
Dhruba Borthakur 已提交
108
static long FLAGS_cache_size = -1;
J
jorlow@chromium.org 已提交
109

110 111 112
// Number of bytes in a block.
static int FLAGS_block_size = 0;

113 114 115
// Maximum number of files to keep open at the same time (use default if == 0)
static int FLAGS_open_files = 0;

S
Sanjay Ghemawat 已提交
116 117 118 119
// Bloom filter bits per key.
// Negative means use default settings.
static int FLAGS_bloom_bits = -1;

120 121 122 123 124
// If true, do not destroy the existing database.  If you set this
// flag and also specify a benchmark that wants a fresh database, that
// benchmark will fail.
static bool FLAGS_use_existing_db = false;

125
// Use the db with the following name.
H
heyongqiang 已提交
126
static const char* FLAGS_db = NULL;
127

128 129 130 131 132
// Number of shards for the block cache is 2 ** FLAGS_cache_numshardbits.
// Negative means use default settings. This is applied only
// if FLAGS_cache_size is non-negative.
static int FLAGS_cache_numshardbits = -1;

133 134 135
// Verify checksum for every block read from storage
static bool FLAGS_verify_checksum = false;

136 137 138 139
// Database statistics
static bool FLAGS_statistics = false;
static class leveldb::DBStatistics* dbstats = NULL;

140 141 142
// Number of write operations to do.  If negative, do FLAGS_num reads.
static long FLAGS_writes = -1;

H
heyongqiang 已提交
143 144
// These default values might change if the hardcoded

145 146 147
// Sync all writes to disk
static bool FLAGS_sync = false;

H
heyongqiang 已提交
148 149 150
// If true, do not wait until data is synced to disk.
static bool FLAGS_disable_data_sync = false;

151 152 153
// If true, issue fsync instead of fdatasync
static bool FLAGS_use_fsync = false;

H
heyongqiang 已提交
154 155 156
// If true, do not write WAL for write.
static bool FLAGS_disable_wal = false;

157
// The total number of levels
158
static unsigned int FLAGS_num_levels = 7;
159

H
heyongqiang 已提交
160 161 162 163 164 165
// Target level-0 file size for compaction
static int FLAGS_target_file_size_base = 2 * 1048576;

// A multiplier to compute targe level-N file size
static int FLAGS_target_file_size_multiplier = 1;

166
// Max bytes for level-1
167
static uint64_t FLAGS_max_bytes_for_level_base = 10 * 1048576;
H
heyongqiang 已提交
168 169 170 171

// A multiplier to compute max bytes for level-N
static int FLAGS_max_bytes_for_level_multiplier = 10;

H
heyongqiang 已提交
172 173 174
// Number of files in level-0 that will trigger put stop.
static int FLAGS_level0_stop_writes_trigger = 12;

175 176
// Number of files in level-0 that will slow down writes.
static int FLAGS_level0_slowdown_writes_trigger = 8;
H
heyongqiang 已提交
177

178 179 180
// Number of files in level-0 when compactions start
static int FLAGS_level0_file_num_compaction_trigger = 4;

181 182 183 184 185
// Ratio of reads to writes (expressed as a percentage)
// for the ReadRandomWriteRandom workload. The default
// setting is 9 gets for every 1 put.
static int FLAGS_readwritepercent = 90;

186 187 188
// Option to disable compation triggered by read.
static int FLAGS_disable_seek_compaction = false;

189 190 191 192 193
// Option to delete obsolete files periodically
// Default: 0 which means that obsolete files are
// deleted after every compaction run.
static uint64_t FLAGS_delete_obsolete_files_period_micros = 0;

194 195 196 197
// Algorithm to use to compress the database
static enum leveldb::CompressionType FLAGS_compression_type =
    leveldb::kSnappyCompression;

198 199
// Allows compression for levels 0 and 1 to be disabled when
// other levels are compressed
200
static int FLAGS_min_level_to_compress = -1;
201

202 203
static int FLAGS_table_cache_numshardbits = 4;

204 205 206
// posix or hdfs environment
static leveldb::Env* FLAGS_env = leveldb::Env::Default();

207 208 209 210
// Stats are reported every N operations when this is greater
// than zero. When 0 the interval grows over time.
static int FLAGS_stats_interval = 0;

211 212 213 214 215 216 217 218 219
// Reports additional stats per interval when this is greater
// than 0.
static int FLAGS_stats_per_interval = 0;

// When not equal to 0 this make threads sleep at each stats
// reporting interval until the compaction score for all levels is
// less than or equal to this value.
static double FLAGS_rate_limit = 0;

220 221
// Control maximum bytes of overlaps in grandparent (i.e., level+2) before we
// stop building a single file in a level->level+1 compaction.
222
static int FLAGS_max_grandparent_overlap_factor = 10;
223

H
heyongqiang 已提交
224 225 226
// Run read only benchmarks.
static bool FLAGS_read_only = false;

227 228 229
// Do not auto trigger compactions
static bool FLAGS_disable_auto_compactions = false;

230 231 232 233
// Cap the size of data in levelK for a compaction run
// that compacts Levelk with LevelK+1
static int FLAGS_source_compaction_factor = 1;

234 235 236
// Set the TTL for the WAL Files.
static uint64_t FLAGS_WAL_ttl_seconds = 0;

237
extern bool useOsBuffer;
238
extern bool useFsReadAhead;
239
extern bool useMmapRead;
240
extern bool useMmapWrite;
241

J
jorlow@chromium.org 已提交
242 243
namespace leveldb {

244
// Helper for quickly generating random data.
J
jorlow@chromium.org 已提交
245 246 247
class RandomGenerator {
 private:
  std::string data_;
248
  unsigned int pos_;
J
jorlow@chromium.org 已提交
249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273

 public:
  RandomGenerator() {
    // We use a limited amount of data over and over again and ensure
    // that it is larger than the compression window (32KB), and also
    // large enough to serve all typical value sizes we want to write.
    Random rnd(301);
    std::string piece;
    while (data_.size() < 1048576) {
      // Add a short fragment that is as compressible as specified
      // by FLAGS_compression_ratio.
      test::CompressibleString(&rnd, FLAGS_compression_ratio, 100, &piece);
      data_.append(piece);
    }
    pos_ = 0;
  }

  Slice Generate(int len) {
    if (pos_ + len > data_.size()) {
      pos_ = 0;
      assert(len < data_.size());
    }
    pos_ += len;
    return Slice(data_.data() + pos_ - len, len);
  }
274
};
275
static Slice TrimSpace(Slice s) {
276
  unsigned int start = 0;
277 278 279
  while (start < s.size() && isspace(s[start])) {
    start++;
  }
280
  unsigned int limit = s.size();
281 282 283 284 285 286
  while (limit > start && isspace(s[limit-1])) {
    limit--;
  }
  return Slice(s.data() + start, limit - start);
}

287 288 289 290 291 292 293 294 295 296
static void AppendWithSpace(std::string* str, Slice msg) {
  if (msg.empty()) return;
  if (!str->empty()) {
    str->push_back(' ');
  }
  str->append(msg.data(), msg.size());
}

class Stats {
 private:
297
  int id_;
298 299 300
  double start_;
  double finish_;
  double seconds_;
301
  long done_;
302
  long last_report_done_;
303 304 305
  int next_report_;
  int64_t bytes_;
  double last_op_finish_;
306
  double last_report_finish_;
307 308 309 310
  Histogram hist_;
  std::string message_;

 public:
311
  Stats() { Start(-1); }
312

313 314 315
  void Start(int id) {
    id_ = id;
    next_report_ = FLAGS_stats_interval ? FLAGS_stats_interval : 100;
316 317 318
    last_op_finish_ = start_;
    hist_.Clear();
    done_ = 0;
319
    last_report_done_ = 0;
320 321
    bytes_ = 0;
    seconds_ = 0;
322
    start_ = FLAGS_env->NowMicros();
323
    finish_ = start_;
324
    last_report_finish_ = start_;
325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340
    message_.clear();
  }

  void Merge(const Stats& other) {
    hist_.Merge(other.hist_);
    done_ += other.done_;
    bytes_ += other.bytes_;
    seconds_ += other.seconds_;
    if (other.start_ < start_) start_ = other.start_;
    if (other.finish_ > finish_) finish_ = other.finish_;

    // Just keep the messages from one thread
    if (message_.empty()) message_ = other.message_;
  }

  void Stop() {
341
    finish_ = FLAGS_env->NowMicros();
342 343 344 345 346 347 348
    seconds_ = (finish_ - start_) * 1e-6;
  }

  void AddMessage(Slice msg) {
    AppendWithSpace(&message_, msg);
  }

349 350
  void SetId(int id) { id_ = id; }

M
Mark Callaghan 已提交
351
  void FinishedSingleOp(DB* db) {
352
    if (FLAGS_histogram) {
353
      double now = FLAGS_env->NowMicros();
354 355
      double micros = now - last_op_finish_;
      hist_.Add(micros);
356
      if (micros > 20000 && !FLAGS_stats_interval) {
357 358 359 360 361 362 363 364
        fprintf(stderr, "long op: %.1f micros%30s\r", micros, "");
        fflush(stderr);
      }
      last_op_finish_ = now;
    }

    done_++;
    if (done_ >= next_report_) {
365 366 367 368 369 370 371 372 373 374 375 376 377
      if (!FLAGS_stats_interval) {
        if      (next_report_ < 1000)   next_report_ += 100;
        else if (next_report_ < 5000)   next_report_ += 500;
        else if (next_report_ < 10000)  next_report_ += 1000;
        else if (next_report_ < 50000)  next_report_ += 5000;
        else if (next_report_ < 100000) next_report_ += 10000;
        else if (next_report_ < 500000) next_report_ += 50000;
        else                            next_report_ += 100000;
        fprintf(stderr, "... finished %ld ops%30s\r", done_, "");
        fflush(stderr);
      } else {
        double now = FLAGS_env->NowMicros();
        fprintf(stderr,
378
                "%s ... thread %d: (%ld,%ld) ops and (%.1f,%.1f) ops/second in (%.6f,%.6f) seconds\n",
379
                FLAGS_env->TimeToString((uint64_t) now/1000000).c_str(),
380
                id_,
M
Mark Callaghan 已提交
381
                done_ - last_report_done_, done_,
382
                (done_ - last_report_done_) /
M
Mark Callaghan 已提交
383 384 385 386
                ((now - last_report_finish_) / 1000000.0),
                done_ / ((now - start_) / 1000000.0),
                (now - last_report_finish_) / 1000000.0,
                (now - start_) / 1000000.0);
M
Mark Callaghan 已提交
387

388 389 390
        if (FLAGS_stats_per_interval) {
          std::string stats;
          if (db && db->GetProperty("leveldb.stats", &stats))
391
            fprintf(stderr, "%s\n", stats.c_str());
392
        }
M
Mark Callaghan 已提交
393

394 395 396 397 398
        fflush(stderr);
        next_report_ += FLAGS_stats_interval;
        last_report_finish_ = now;
        last_report_done_ = done_;
      }
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421
    }
  }

  void AddBytes(int64_t n) {
    bytes_ += n;
  }

  void Report(const Slice& name) {
    // Pretend at least one op was done in case we are running a benchmark
    // that does not call FinishedSingleOp().
    if (done_ < 1) done_ = 1;

    std::string extra;
    if (bytes_ > 0) {
      // Rate is computed on actual elapsed time, not the sum of per-thread
      // elapsed times.
      double elapsed = (finish_ - start_) * 1e-6;
      char rate[100];
      snprintf(rate, sizeof(rate), "%6.1f MB/s",
               (bytes_ / 1048576.0) / elapsed);
      extra = rate;
    }
    AppendWithSpace(&extra, message_);
422 423
    double elapsed = (finish_ - start_) * 1e-6;
    double throughput = (double)done_/elapsed;
424

D
Dhruba Borthakur 已提交
425
    fprintf(stdout, "%-12s : %11.3f micros/op %ld ops/sec;%s%s\n",
426 427
            name.ToString().c_str(),
            seconds_ * 1e6 / done_,
D
Dhruba Borthakur 已提交
428
            (long)throughput,
429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449
            (extra.empty() ? "" : " "),
            extra.c_str());
    if (FLAGS_histogram) {
      fprintf(stdout, "Microseconds per op:\n%s\n", hist_.ToString().c_str());
    }
    fflush(stdout);
  }
};

// State shared by all concurrent executions of the same benchmark.
struct SharedState {
  port::Mutex mu;
  port::CondVar cv;
  int total;

  // Each thread goes through the following states:
  //    (1) initializing
  //    (2) waiting for others to be initialized
  //    (3) running
  //    (4) done

450 451
  long num_initialized;
  long num_done;
452 453 454 455 456 457 458 459 460 461
  bool start;

  SharedState() : cv(&mu) { }
};

// Per-thread state for concurrent executions of the same benchmark.
struct ThreadState {
  int tid;             // 0..n-1 when running in n threads
  Random rand;         // Has different seeds for different threads
  Stats stats;
462
  SharedState* shared;
463

A
Abhishek Kona 已提交
464
  /* implicit */ ThreadState(int index)
465
      : tid(index),
466
        rand((FLAGS_seed ? FLAGS_seed : 1000) + index) {
467 468 469
  }
};

J
jorlow@chromium.org 已提交
470 471 472
class Benchmark {
 private:
  Cache* cache_;
S
Sanjay Ghemawat 已提交
473
  const FilterPolicy* filter_policy_;
J
jorlow@chromium.org 已提交
474
  DB* db_;
475
  long num_;
476 477 478
  int value_size_;
  int entries_per_batch_;
  WriteOptions write_options_;
479
  long reads_;
480
  long writes_;
481
  long readwrites_;
J
jorlow@chromium.org 已提交
482 483
  int heap_counter_;

484 485 486 487 488 489 490
  void PrintHeader() {
    const int kKeySize = 16;
    PrintEnvironment();
    fprintf(stdout, "Keys:       %d bytes each\n", kKeySize);
    fprintf(stdout, "Values:     %d bytes each (%d bytes after compression)\n",
            FLAGS_value_size,
            static_cast<int>(FLAGS_value_size * FLAGS_compression_ratio + 0.5));
491
    fprintf(stdout, "Entries:    %ld\n", num_);
492
    fprintf(stdout, "RawSize:    %.1f MB (estimated)\n",
493 494
            ((static_cast<int64_t>(kKeySize + FLAGS_value_size) * num_)
             / 1048576.0));
495 496 497
    fprintf(stdout, "FileSize:   %.1f MB (estimated)\n",
            (((kKeySize + FLAGS_value_size * FLAGS_compression_ratio) * num_)
             / 1048576.0));
498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513

    switch (FLAGS_compression_type) {
      case leveldb::kNoCompression:
        fprintf(stdout, "Compression: none\n");
        break;
      case leveldb::kSnappyCompression:
        fprintf(stdout, "Compression: snappy\n");
        break;
      case leveldb::kZlibCompression:
        fprintf(stdout, "Compression: zlib\n");
        break;
      case leveldb::kBZip2Compression:
        fprintf(stdout, "Compression: bzip2\n");
        break;
    }

514 515 516 517 518 519 520 521 522 523 524 525 526 527
    PrintWarnings();
    fprintf(stdout, "------------------------------------------------\n");
  }

  void PrintWarnings() {
#if defined(__GNUC__) && !defined(__OPTIMIZE__)
    fprintf(stdout,
            "WARNING: Optimization is disabled: benchmarks unnecessarily slow\n"
            );
#endif
#ifndef NDEBUG
    fprintf(stdout,
            "WARNING: Assertions are enabled; benchmarks unnecessarily slow\n");
#endif
528

529 530 531 532 533 534 535 536 537 538 539 540 541
    if (FLAGS_compression_type != leveldb::kNoCompression) {
      // The test string should not be too small.
      const int len = FLAGS_block_size;
      char* text = (char*) malloc(len+1);
      bool result = true;
      const char* name = NULL;
      std::string compressed;

      memset(text, (int) 'y', len);
      text[len] = '\0';

      switch (FLAGS_compression_type) {
        case kSnappyCompression:
542 543
          result = port::Snappy_Compress(Options().compression_opts, text,
                                         strlen(text), &compressed);
544 545 546
          name = "Snappy";
          break;
        case kZlibCompression:
547 548
          result = port::Zlib_Compress(Options().compression_opts, text,
                                       strlen(text), &compressed);
549 550 551
          name = "Zlib";
          break;
        case kBZip2Compression:
552 553
          result = port::BZip2_Compress(Options().compression_opts, text,
                                        strlen(text), &compressed);
554 555
          name = "BZip2";
          break;
556 557 558
        case kNoCompression:
          assert(false); // cannot happen
          break;
559 560 561 562 563 564 565 566 567
      }

      if (!result) {
        fprintf(stdout, "WARNING: %s compression is not enabled\n", name);
      } else if (name && compressed.size() >= strlen(text)) {
        fprintf(stdout, "WARNING: %s compression is not effective\n", name);
      }

      free(text);
568
    }
569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605
  }

  void PrintEnvironment() {
    fprintf(stderr, "LevelDB:    version %d.%d\n",
            kMajorVersion, kMinorVersion);

#if defined(__linux)
    time_t now = time(NULL);
    fprintf(stderr, "Date:       %s", ctime(&now));  // ctime() adds newline

    FILE* cpuinfo = fopen("/proc/cpuinfo", "r");
    if (cpuinfo != NULL) {
      char line[1000];
      int num_cpus = 0;
      std::string cpu_type;
      std::string cache_size;
      while (fgets(line, sizeof(line), cpuinfo) != NULL) {
        const char* sep = strchr(line, ':');
        if (sep == NULL) {
          continue;
        }
        Slice key = TrimSpace(Slice(line, sep - 1 - line));
        Slice val = TrimSpace(Slice(sep + 1));
        if (key == "model name") {
          ++num_cpus;
          cpu_type = val.ToString();
        } else if (key == "cache size") {
          cache_size = val.ToString();
        }
      }
      fclose(cpuinfo);
      fprintf(stderr, "CPU:        %d * %s\n", num_cpus, cpu_type.c_str());
      fprintf(stderr, "CPUCache:   %s\n", cache_size.c_str());
    }
#endif
  }

606 607
  void PrintStatistics() {
    if (FLAGS_statistics) {
A
Abhishek Kona 已提交
608
      fprintf(stdout, "File opened:%ld closed:%ld errors:%ld\n"
609 610 611 612 613 614 615
          "Block Cache Hit Count:%ld Block Cache Miss Count:%ld\n"
          "Bloom Filter Useful: %ld \n"
          "Compaction key_drop_newer_entry: %ld key_drop_obsolete: %ld "
          "Compaction key_drop_user: %ld",
          dbstats->getNumFileOpens(),
          dbstats->getNumFileCloses(),
          dbstats->getNumFileErrors(),
A
Abhishek Kona 已提交
616
          dbstats->getTickerCount(BLOCK_CACHE_HIT),
617 618 619 620 621
          dbstats->getTickerCount(BLOCK_CACHE_MISS),
          dbstats->getTickerCount(BLOOM_FILTER_USEFUL),
          dbstats->getTickerCount(COMPACTION_KEY_DROP_NEWER_ENTRY),
          dbstats->getTickerCount(COMPACTION_KEY_DROP_OBSOLETE),
          dbstats->getTickerCount(COMPACTION_KEY_DROP_USER));
622 623 624
    }
  }

J
jorlow@chromium.org 已提交
625
 public:
626
  Benchmark()
627 628 629
  : cache_(FLAGS_cache_size >= 0 ?
           (FLAGS_cache_numshardbits >= 1 ?
            NewLRUCache(FLAGS_cache_size, FLAGS_cache_numshardbits) :
630
            NewLRUCache(FLAGS_cache_size)) : NULL),
S
Sanjay Ghemawat 已提交
631 632 633
    filter_policy_(FLAGS_bloom_bits >= 0
                   ? NewBloomFilterPolicy(FLAGS_bloom_bits)
                   : NULL),
634 635
    db_(NULL),
    num_(FLAGS_num),
636 637
    value_size_(FLAGS_value_size),
    entries_per_batch_(1),
638
    reads_(FLAGS_reads < 0 ? FLAGS_num : FLAGS_reads),
639
    writes_(FLAGS_writes < 0 ? FLAGS_num : FLAGS_writes),
640
    readwrites_((FLAGS_writes < 0  && FLAGS_reads < 0)? FLAGS_num :
641
                ((FLAGS_writes > FLAGS_reads) ? FLAGS_writes : FLAGS_reads)
642
               ),
643
    heap_counter_(0) {
J
jorlow@chromium.org 已提交
644
    std::vector<std::string> files;
645
    FLAGS_env->GetChildren(FLAGS_db, &files);
646
    for (unsigned int i = 0; i < files.size(); i++) {
J
jorlow@chromium.org 已提交
647
      if (Slice(files[i]).starts_with("heap-")) {
648
        FLAGS_env->DeleteFile(std::string(FLAGS_db) + "/" + files[i]);
J
jorlow@chromium.org 已提交
649 650
      }
    }
651
    if (!FLAGS_use_existing_db) {
652
      DestroyDB(FLAGS_db, Options());
653
    }
J
jorlow@chromium.org 已提交
654 655 656 657 658
  }

  ~Benchmark() {
    delete db_;
    delete cache_;
S
Sanjay Ghemawat 已提交
659
    delete filter_policy_;
J
jorlow@chromium.org 已提交
660 661 662
  }

  void Run() {
663 664
    PrintHeader();
    Open();
J
jorlow@chromium.org 已提交
665 666 667 668 669 670 671 672 673 674 675 676 677

    const char* benchmarks = FLAGS_benchmarks;
    while (benchmarks != NULL) {
      const char* sep = strchr(benchmarks, ',');
      Slice name;
      if (sep == NULL) {
        name = benchmarks;
        benchmarks = NULL;
      } else {
        name = Slice(benchmarks, sep - benchmarks);
        benchmarks = sep + 1;
      }

678 679
      // Reset parameters that may be overriddden bwlow
      num_ = FLAGS_num;
680
      reads_ = (FLAGS_reads < 0 ? FLAGS_num : FLAGS_reads);
681
      writes_ = (FLAGS_writes < 0 ? FLAGS_num : FLAGS_writes);
682 683 684
      value_size_ = FLAGS_value_size;
      entries_per_batch_ = 1;
      write_options_ = WriteOptions();
685 686 687
      if (FLAGS_sync) {
        write_options_.sync = true;
      }
688

H
heyongqiang 已提交
689 690
      write_options_.disableWAL = FLAGS_disable_wal;

691 692
      void (Benchmark::*method)(ThreadState*) = NULL;
      bool fresh_db = false;
693
      int num_threads = FLAGS_threads;
694 695

      if (name == Slice("fillseq")) {
696 697
        fresh_db = true;
        method = &Benchmark::WriteSeq;
698
      } else if (name == Slice("fillbatch")) {
699 700 701
        fresh_db = true;
        entries_per_batch_ = 1000;
        method = &Benchmark::WriteSeq;
702
      } else if (name == Slice("fillrandom")) {
703 704
        fresh_db = true;
        method = &Benchmark::WriteRandom;
705
      } else if (name == Slice("overwrite")) {
706 707
        fresh_db = false;
        method = &Benchmark::WriteRandom;
708
      } else if (name == Slice("fillsync")) {
709 710 711 712
        fresh_db = true;
        num_ /= 1000;
        write_options_.sync = true;
        method = &Benchmark::WriteRandom;
713
      } else if (name == Slice("fill100K")) {
714 715 716 717
        fresh_db = true;
        num_ /= 1000;
        value_size_ = 100 * 1000;
        method = &Benchmark::WriteRandom;
J
jorlow@chromium.org 已提交
718
      } else if (name == Slice("readseq")) {
719
        method = &Benchmark::ReadSequential;
J
jorlow@chromium.org 已提交
720
      } else if (name == Slice("readreverse")) {
721
        method = &Benchmark::ReadReverse;
J
jorlow@chromium.org 已提交
722
      } else if (name == Slice("readrandom")) {
723
        method = &Benchmark::ReadRandom;
S
Sanjay Ghemawat 已提交
724 725 726 727
      } else if (name == Slice("readmissing")) {
        method = &Benchmark::ReadMissing;
      } else if (name == Slice("seekrandom")) {
        method = &Benchmark::SeekRandom;
728
      } else if (name == Slice("readhot")) {
729
        method = &Benchmark::ReadHot;
730
      } else if (name == Slice("readrandomsmall")) {
731
        reads_ /= 1000;
732
        method = &Benchmark::ReadRandom;
S
Sanjay Ghemawat 已提交
733 734 735 736
      } else if (name == Slice("deleteseq")) {
        method = &Benchmark::DeleteSeq;
      } else if (name == Slice("deleterandom")) {
        method = &Benchmark::DeleteRandom;
737 738 739
      } else if (name == Slice("readwhilewriting")) {
        num_threads++;  // Add extra thread for writing
        method = &Benchmark::ReadWhileWriting;
740 741
      } else if (name == Slice("readrandomwriterandom")) {
        method = &Benchmark::ReadRandomWriteRandom;
J
jorlow@chromium.org 已提交
742
      } else if (name == Slice("compact")) {
743
        method = &Benchmark::Compact;
J
jorlow@chromium.org 已提交
744
      } else if (name == Slice("crc32c")) {
745
        method = &Benchmark::Crc32c;
746
      } else if (name == Slice("acquireload")) {
747
        method = &Benchmark::AcquireLoad;
748
      } else if (name == Slice("snappycomp")) {
749
        method = &Benchmark::SnappyCompress;
750
      } else if (name == Slice("snappyuncomp")) {
751
        method = &Benchmark::SnappyUncompress;
J
jorlow@chromium.org 已提交
752 753
      } else if (name == Slice("heapprofile")) {
        HeapProfile();
754
      } else if (name == Slice("stats")) {
S
Sanjay Ghemawat 已提交
755 756 757
        PrintStats("leveldb.stats");
      } else if (name == Slice("sstables")) {
        PrintStats("leveldb.sstables");
J
jorlow@chromium.org 已提交
758
      } else {
759 760 761 762
        if (name != Slice()) {  // No error message for empty name
          fprintf(stderr, "unknown benchmark '%s'\n", name.ToString().c_str());
        }
      }
763 764 765 766 767 768 769 770 771 772 773 774 775 776 777

      if (fresh_db) {
        if (FLAGS_use_existing_db) {
          fprintf(stdout, "%-12s : skipped (--use_existing_db is true)\n",
                  name.ToString().c_str());
          method = NULL;
        } else {
          delete db_;
          db_ = NULL;
          DestroyDB(FLAGS_db, Options());
          Open();
        }
      }

      if (method != NULL) {
778
        RunBenchmark(num_threads, name, method);
J
jorlow@chromium.org 已提交
779 780
      }
    }
781
    PrintStatistics();
J
jorlow@chromium.org 已提交
782 783
  }

784
 private:
785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806
  struct ThreadArg {
    Benchmark* bm;
    SharedState* shared;
    ThreadState* thread;
    void (Benchmark::*method)(ThreadState*);
  };

  static void ThreadBody(void* v) {
    ThreadArg* arg = reinterpret_cast<ThreadArg*>(v);
    SharedState* shared = arg->shared;
    ThreadState* thread = arg->thread;
    {
      MutexLock l(&shared->mu);
      shared->num_initialized++;
      if (shared->num_initialized >= shared->total) {
        shared->cv.SignalAll();
      }
      while (!shared->start) {
        shared->cv.Wait();
      }
    }

807
    thread->stats.Start(thread->tid);
808 809 810 811 812 813 814 815 816 817 818 819
    (arg->bm->*(arg->method))(thread);
    thread->stats.Stop();

    {
      MutexLock l(&shared->mu);
      shared->num_done++;
      if (shared->num_done >= shared->total) {
        shared->cv.SignalAll();
      }
    }
  }

820 821
  void RunBenchmark(int n, Slice name,
                    void (Benchmark::*method)(ThreadState*)) {
822 823 824 825 826 827 828 829 830 831 832 833
    SharedState shared;
    shared.total = n;
    shared.num_initialized = 0;
    shared.num_done = 0;
    shared.start = false;

    ThreadArg* arg = new ThreadArg[n];
    for (int i = 0; i < n; i++) {
      arg[i].bm = this;
      arg[i].method = method;
      arg[i].shared = &shared;
      arg[i].thread = new ThreadState(i);
834
      arg[i].thread->shared = &shared;
835
      FLAGS_env->StartThread(ThreadBody, &arg[i]);
836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861
    }

    shared.mu.Lock();
    while (shared.num_initialized < n) {
      shared.cv.Wait();
    }

    shared.start = true;
    shared.cv.SignalAll();
    while (shared.num_done < n) {
      shared.cv.Wait();
    }
    shared.mu.Unlock();

    for (int i = 1; i < n; i++) {
      arg[0].thread->stats.Merge(arg[i].thread->stats);
    }
    arg[0].thread->stats.Report(name);

    for (int i = 0; i < n; i++) {
      delete arg[i].thread;
    }
    delete[] arg;
  }

  void Crc32c(ThreadState* thread) {
J
jorlow@chromium.org 已提交
862
    // Checksum about 500MB of data total
863 864
    const int size = 4096;
    const char* label = "(4K per op)";
J
jorlow@chromium.org 已提交
865
    std::string data(size, 'x');
J
jorlow@chromium.org 已提交
866 867 868 869
    int64_t bytes = 0;
    uint32_t crc = 0;
    while (bytes < 500 * 1048576) {
      crc = crc32c::Value(data.data(), size);
M
Mark Callaghan 已提交
870
      thread->stats.FinishedSingleOp(NULL);
J
jorlow@chromium.org 已提交
871 872 873 874 875
      bytes += size;
    }
    // Print so result is not dead
    fprintf(stderr, "... crc=0x%x\r", static_cast<unsigned int>(crc));

876 877
    thread->stats.AddBytes(bytes);
    thread->stats.AddMessage(label);
J
jorlow@chromium.org 已提交
878 879
  }

880
  void AcquireLoad(ThreadState* thread) {
881 882 883 884
    int dummy;
    port::AtomicPointer ap(&dummy);
    int count = 0;
    void *ptr = NULL;
885
    thread->stats.AddMessage("(each op is 1000 loads)");
886 887 888 889 890
    while (count < 100000) {
      for (int i = 0; i < 1000; i++) {
        ptr = ap.Acquire_Load();
      }
      count++;
M
Mark Callaghan 已提交
891
      thread->stats.FinishedSingleOp(NULL);
892 893 894 895
    }
    if (ptr == NULL) exit(1); // Disable unused variable warning.
  }

896 897 898
  void SnappyCompress(ThreadState* thread) {
    RandomGenerator gen;
    Slice input = gen.Generate(Options().block_size);
899 900 901 902 903
    int64_t bytes = 0;
    int64_t produced = 0;
    bool ok = true;
    std::string compressed;
    while (ok && bytes < 1024 * 1048576) {  // Compress 1G
904 905
      ok = port::Snappy_Compress(Options().compression_opts, input.data(),
                                 input.size(), &compressed);
906 907
      produced += compressed.size();
      bytes += input.size();
M
Mark Callaghan 已提交
908
      thread->stats.FinishedSingleOp(NULL);
909 910 911
    }

    if (!ok) {
912
      thread->stats.AddMessage("(snappy failure)");
913 914 915 916
    } else {
      char buf[100];
      snprintf(buf, sizeof(buf), "(output: %.1f%%)",
               (produced * 100.0) / bytes);
917 918
      thread->stats.AddMessage(buf);
      thread->stats.AddBytes(bytes);
919 920 921
    }
  }

922 923 924
  void SnappyUncompress(ThreadState* thread) {
    RandomGenerator gen;
    Slice input = gen.Generate(Options().block_size);
925
    std::string compressed;
926 927
    bool ok = port::Snappy_Compress(Options().compression_opts, input.data(),
                                    input.size(), &compressed);
928
    int64_t bytes = 0;
929
    char* uncompressed = new char[input.size()];
930 931
    while (ok && bytes < 1024 * 1048576) {  // Compress 1G
      ok =  port::Snappy_Uncompress(compressed.data(), compressed.size(),
932 933
                                    uncompressed);
      bytes += input.size();
M
Mark Callaghan 已提交
934
      thread->stats.FinishedSingleOp(NULL);
935
    }
936
    delete[] uncompressed;
937 938

    if (!ok) {
939
      thread->stats.AddMessage("(snappy failure)");
940
    } else {
941
      thread->stats.AddBytes(bytes);
942 943 944
    }
  }

945 946 947
  void Open() {
    assert(db_ == NULL);
    Options options;
948
    options.create_if_missing = !FLAGS_use_existing_db;
949
    options.block_cache = cache_;
950 951 952
    if (cache_ == NULL) {
      options.no_block_cache = true;
    }
953
    options.write_buffer_size = FLAGS_write_buffer_size;
954 955
    options.max_write_buffer_number = FLAGS_max_write_buffer_number;
    options.max_background_compactions = FLAGS_max_background_compactions;
956
    options.block_size = FLAGS_block_size;
S
Sanjay Ghemawat 已提交
957
    options.filter_policy = filter_policy_;
958 959
    options.max_open_files = FLAGS_open_files;
    options.statistics = dbstats;
960
    options.env = FLAGS_env;
H
heyongqiang 已提交
961
    options.disableDataSync = FLAGS_disable_data_sync;
962
    options.use_fsync = FLAGS_use_fsync;
963
    options.num_levels = FLAGS_num_levels;
H
heyongqiang 已提交
964 965 966 967 968
    options.target_file_size_base = FLAGS_target_file_size_base;
    options.target_file_size_multiplier = FLAGS_target_file_size_multiplier;
    options.max_bytes_for_level_base = FLAGS_max_bytes_for_level_base;
    options.max_bytes_for_level_multiplier =
        FLAGS_max_bytes_for_level_multiplier;
H
heyongqiang 已提交
969
    options.level0_stop_writes_trigger = FLAGS_level0_stop_writes_trigger;
M
Mark Callaghan 已提交
970
    options.level0_file_num_compaction_trigger =
971
        FLAGS_level0_file_num_compaction_trigger;
H
heyongqiang 已提交
972 973
    options.level0_slowdown_writes_trigger =
      FLAGS_level0_slowdown_writes_trigger;
974
    options.compression = FLAGS_compression_type;
975
    options.WAL_ttl_seconds = FLAGS_WAL_ttl_seconds;
976 977 978
    if (FLAGS_min_level_to_compress >= 0) {
      assert(FLAGS_min_level_to_compress <= FLAGS_num_levels);
      options.compression_per_level = new CompressionType[FLAGS_num_levels];
979
      for (int i = 0; i < FLAGS_min_level_to_compress; i++) {
980 981
        options.compression_per_level[i] = kNoCompression;
      }
982
      for (unsigned int i = FLAGS_min_level_to_compress;
983 984 985 986
           i < FLAGS_num_levels; i++) {
        options.compression_per_level[i] = FLAGS_compression_type;
      }
    }
987
    options.disable_seek_compaction = FLAGS_disable_seek_compaction;
988 989
    options.delete_obsolete_files_period_micros =
      FLAGS_delete_obsolete_files_period_micros;
990
    options.rate_limit = FLAGS_rate_limit;
991
    options.table_cache_numshardbits = FLAGS_table_cache_numshardbits;
A
Abhishek Kona 已提交
992
    options.max_grandparent_overlap_factor =
993
      FLAGS_max_grandparent_overlap_factor;
994
    options.disable_auto_compactions = FLAGS_disable_auto_compactions;
995
    options.source_compaction_factor = FLAGS_source_compaction_factor;
H
heyongqiang 已提交
996 997 998 999 1000 1001
    Status s;
    if(FLAGS_read_only) {
      s = DB::OpenForReadOnly(options, FLAGS_db, &db_);
    } else {
      s = DB::Open(options, FLAGS_db, &db_);
    }
1002 1003 1004 1005
    if (!s.ok()) {
      fprintf(stderr, "open error: %s\n", s.ToString().c_str());
      exit(1);
    }
1006 1007 1008
    if (FLAGS_min_level_to_compress >= 0) {
      delete options.compression_per_level;
    }
1009 1010
  }

1011 1012 1013
  void WriteSeq(ThreadState* thread) {
    DoWrite(thread, true);
  }
1014

1015 1016 1017 1018 1019 1020
  void WriteRandom(ThreadState* thread) {
    DoWrite(thread, false);
  }

  void DoWrite(ThreadState* thread, bool seq) {
    if (num_ != FLAGS_num) {
1021
      char msg[100];
1022
      snprintf(msg, sizeof(msg), "(%ld ops)", num_);
1023
      thread->stats.AddMessage(msg);
1024 1025
    }

1026
    RandomGenerator gen;
J
jorlow@chromium.org 已提交
1027 1028
    WriteBatch batch;
    Status s;
1029
    int64_t bytes = 0;
1030
    for (int i = 0; i < writes_; i += entries_per_batch_) {
J
jorlow@chromium.org 已提交
1031
      batch.Clear();
1032 1033
      for (int j = 0; j < entries_per_batch_; j++) {
        const int k = seq ? i+j : (thread->rand.Next() % FLAGS_num);
1034 1035
        char key[100];
        snprintf(key, sizeof(key), "%016d", k);
1036 1037
        batch.Put(key, gen.Generate(value_size_));
        bytes += value_size_ + strlen(key);
M
Mark Callaghan 已提交
1038
        thread->stats.FinishedSingleOp(db_);
1039
      }
1040
      s = db_->Write(write_options_, &batch);
J
jorlow@chromium.org 已提交
1041 1042 1043 1044 1045
      if (!s.ok()) {
        fprintf(stderr, "put error: %s\n", s.ToString().c_str());
        exit(1);
      }
    }
1046
    thread->stats.AddBytes(bytes);
J
jorlow@chromium.org 已提交
1047 1048
  }

1049
  void ReadSequential(ThreadState* thread) {
1050
    Iterator* iter = db_->NewIterator(ReadOptions(FLAGS_verify_checksum, true));
1051
    long i = 0;
1052
    int64_t bytes = 0;
1053
    for (iter->SeekToFirst(); i < reads_ && iter->Valid(); iter->Next()) {
1054
      bytes += iter->key().size() + iter->value().size();
M
Mark Callaghan 已提交
1055
      thread->stats.FinishedSingleOp(db_);
1056 1057 1058
      ++i;
    }
    delete iter;
1059
    thread->stats.AddBytes(bytes);
1060 1061
  }

1062
  void ReadReverse(ThreadState* thread) {
1063
    Iterator* iter = db_->NewIterator(ReadOptions(FLAGS_verify_checksum, true));
1064
    long i = 0;
1065
    int64_t bytes = 0;
1066
    for (iter->SeekToLast(); i < reads_ && iter->Valid(); iter->Prev()) {
1067
      bytes += iter->key().size() + iter->value().size();
M
Mark Callaghan 已提交
1068
      thread->stats.FinishedSingleOp(db_);
1069 1070 1071
      ++i;
    }
    delete iter;
1072
    thread->stats.AddBytes(bytes);
1073 1074
  }

1075
  void ReadRandom(ThreadState* thread) {
1076
    ReadOptions options(FLAGS_verify_checksum, true);
1077 1078
    Iterator* iter = db_->NewIterator(options);

1079
    std::string value;
1080
    long found = 0;
1081
    for (long i = 0; i < reads_; i++) {
1082
      char key[100];
1083
      const int k = thread->rand.Next() % FLAGS_num;
1084
      snprintf(key, sizeof(key), "%016d", k);
1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097

      if (FLAGS_read_range < 2) {
        if (db_->Get(options, key, &value).ok()) {
          found++;
        }
      } else {
        Slice skey(key);
        int count = 1;
        for (iter->Seek(skey);
             iter->Valid() && count <= FLAGS_read_range;
             ++count, iter->Next()) {
          found++;
        }
S
Sanjay Ghemawat 已提交
1098
      }
1099

M
Mark Callaghan 已提交
1100
      thread->stats.FinishedSingleOp(db_);
S
Sanjay Ghemawat 已提交
1101
    }
1102
    delete iter;
S
Sanjay Ghemawat 已提交
1103
    char msg[100];
1104
    snprintf(msg, sizeof(msg), "(%ld of %ld found)", found, num_);
S
Sanjay Ghemawat 已提交
1105 1106 1107 1108
    thread->stats.AddMessage(msg);
  }

  void ReadMissing(ThreadState* thread) {
1109
    ReadOptions options(FLAGS_verify_checksum, true);
S
Sanjay Ghemawat 已提交
1110
    std::string value;
1111
    for (long i = 0; i < reads_; i++) {
S
Sanjay Ghemawat 已提交
1112 1113 1114
      char key[100];
      const int k = thread->rand.Next() % FLAGS_num;
      snprintf(key, sizeof(key), "%016d.", k);
1115
      db_->Get(options, key, &value);
M
Mark Callaghan 已提交
1116
      thread->stats.FinishedSingleOp(db_);
J
jorlow@chromium.org 已提交
1117 1118 1119
    }
  }

1120
  void ReadHot(ThreadState* thread) {
1121
    ReadOptions options(FLAGS_verify_checksum, true);
1122
    std::string value;
1123 1124
    const long range = (FLAGS_num + 99) / 100;
    for (long i = 0; i < reads_; i++) {
1125
      char key[100];
1126
      const int k = thread->rand.Next() % range;
1127 1128
      snprintf(key, sizeof(key), "%016d", k);
      db_->Get(options, key, &value);
M
Mark Callaghan 已提交
1129
      thread->stats.FinishedSingleOp(db_);
1130 1131 1132
    }
  }

S
Sanjay Ghemawat 已提交
1133
  void SeekRandom(ThreadState* thread) {
1134
    ReadOptions options(FLAGS_verify_checksum, true);
S
Sanjay Ghemawat 已提交
1135
    std::string value;
1136
    long found = 0;
1137
    for (long i = 0; i < reads_; i++) {
S
Sanjay Ghemawat 已提交
1138 1139 1140 1141 1142 1143 1144
      Iterator* iter = db_->NewIterator(options);
      char key[100];
      const int k = thread->rand.Next() % FLAGS_num;
      snprintf(key, sizeof(key), "%016d", k);
      iter->Seek(key);
      if (iter->Valid() && iter->key() == key) found++;
      delete iter;
M
Mark Callaghan 已提交
1145
      thread->stats.FinishedSingleOp(db_);
S
Sanjay Ghemawat 已提交
1146 1147
    }
    char msg[100];
1148
    snprintf(msg, sizeof(msg), "(%ld of %ld found)", found, num_);
S
Sanjay Ghemawat 已提交
1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162
    thread->stats.AddMessage(msg);
  }

  void DoDelete(ThreadState* thread, bool seq) {
    RandomGenerator gen;
    WriteBatch batch;
    Status s;
    for (int i = 0; i < num_; i += entries_per_batch_) {
      batch.Clear();
      for (int j = 0; j < entries_per_batch_; j++) {
        const int k = seq ? i+j : (thread->rand.Next() % FLAGS_num);
        char key[100];
        snprintf(key, sizeof(key), "%016d", k);
        batch.Delete(key);
M
Mark Callaghan 已提交
1163
        thread->stats.FinishedSingleOp(db_);
S
Sanjay Ghemawat 已提交
1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180
      }
      s = db_->Write(write_options_, &batch);
      if (!s.ok()) {
        fprintf(stderr, "del error: %s\n", s.ToString().c_str());
        exit(1);
      }
    }
  }

  void DeleteSeq(ThreadState* thread) {
    DoDelete(thread, true);
  }

  void DeleteRandom(ThreadState* thread) {
    DoDelete(thread, false);
  }

1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206
  void ReadWhileWriting(ThreadState* thread) {
    if (thread->tid > 0) {
      ReadRandom(thread);
    } else {
      // Special thread that keeps writing until other threads are done.
      RandomGenerator gen;
      while (true) {
        {
          MutexLock l(&thread->shared->mu);
          if (thread->shared->num_done + 1 >= thread->shared->num_initialized) {
            // Other threads have finished
            break;
          }
        }

        const int k = thread->rand.Next() % FLAGS_num;
        char key[100];
        snprintf(key, sizeof(key), "%016d", k);
        Status s = db_->Put(write_options_, key, gen.Generate(value_size_));
        if (!s.ok()) {
          fprintf(stderr, "put error: %s\n", s.ToString().c_str());
          exit(1);
        }
      }

      // Do not count any of the preceding work/delay in stats.
1207
      thread->stats.Start(thread->tid);
1208 1209 1210
    }
  }

1211 1212
  //
  // This is diffferent from ReadWhileWriting because it does not use
1213
  // an extra thread.
1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224
  //
  void ReadRandomWriteRandom(ThreadState* thread) {
    ReadOptions options(FLAGS_verify_checksum, true);
    RandomGenerator gen;
    std::string value;
    long found = 0;
    int get_weight = 0;
    int put_weight = 0;
    long reads_done = 0;
    long writes_done = 0;
    // the number of iterations is the larger of read_ or write_
1225
    for (long i = 0; i < readwrites_; i++) {
1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251
      char key[100];
      const int k = thread->rand.Next() % FLAGS_num;
      snprintf(key, sizeof(key), "%016d", k);
      if (get_weight == 0 && put_weight == 0) {
        // one batch complated, reinitialize for next batch
        get_weight = FLAGS_readwritepercent;
        put_weight = 100 - get_weight;
      }
      if (get_weight > 0) {
        // do all the gets first
        if (db_->Get(options, key, &value).ok()) {
          found++;
        }
        get_weight--;
        reads_done++;
      } else  if (put_weight > 0) {
        // then do all the corresponding number of puts
        // for all the gets we have done earlier
        Status s = db_->Put(write_options_, key, gen.Generate(value_size_));
        if (!s.ok()) {
          fprintf(stderr, "put error: %s\n", s.ToString().c_str());
          exit(1);
        }
        put_weight--;
        writes_done++;
      }
M
Mark Callaghan 已提交
1252
      thread->stats.FinishedSingleOp(db_);
1253 1254
    }
    char msg[100];
1255
    snprintf(msg, sizeof(msg), "( reads:%ld writes:%ld total:%ld )",
1256
             reads_done, writes_done, readwrites_);
1257 1258 1259
    thread->stats.AddMessage(msg);
  }

1260
  void Compact(ThreadState* thread) {
G
Gabor Cselle 已提交
1261
    db_->CompactRange(NULL, NULL);
J
jorlow@chromium.org 已提交
1262 1263
  }

S
Sanjay Ghemawat 已提交
1264
  void PrintStats(const char* key) {
1265
    std::string stats;
S
Sanjay Ghemawat 已提交
1266
    if (!db_->GetProperty(key, &stats)) {
1267
      stats = "(failed)";
1268
    }
1269
    fprintf(stdout, "\n%s\n", stats.c_str());
1270 1271
  }

J
jorlow@chromium.org 已提交
1272 1273 1274 1275 1276 1277
  static void WriteToFile(void* arg, const char* buf, int n) {
    reinterpret_cast<WritableFile*>(arg)->Append(Slice(buf, n));
  }

  void HeapProfile() {
    char fname[100];
1278
    snprintf(fname, sizeof(fname), "%s/heap-%04d", FLAGS_db, ++heap_counter_);
J
jorlow@chromium.org 已提交
1279
    WritableFile* file;
1280
    Status s = FLAGS_env->NewWritableFile(fname, &file);
J
jorlow@chromium.org 已提交
1281
    if (!s.ok()) {
1282
      fprintf(stderr, "%s\n", s.ToString().c_str());
J
jorlow@chromium.org 已提交
1283 1284 1285 1286 1287
      return;
    }
    bool ok = port::GetHeapProfile(WriteToFile, file);
    delete file;
    if (!ok) {
1288
      fprintf(stderr, "heap profiling not supported\n");
1289
      FLAGS_env->DeleteFile(fname);
J
jorlow@chromium.org 已提交
1290 1291 1292 1293
    }
  }
};

H
Hans Wennborg 已提交
1294
}  // namespace leveldb
J
jorlow@chromium.org 已提交
1295 1296

int main(int argc, char** argv) {
1297
  FLAGS_write_buffer_size = leveldb::Options().write_buffer_size;
1298
  FLAGS_max_write_buffer_number = leveldb::Options().max_write_buffer_number;
1299
  FLAGS_open_files = leveldb::Options().max_open_files;
A
Abhishek Kona 已提交
1300
  FLAGS_max_background_compactions =
1301
    leveldb::Options().max_background_compactions;
1302 1303
  // Compression test code above refers to FLAGS_block_size
  FLAGS_block_size = leveldb::Options().block_size;
H
heyongqiang 已提交
1304
  std::string default_db_path;
1305

J
jorlow@chromium.org 已提交
1306 1307 1308
  for (int i = 1; i < argc; i++) {
    double d;
    int n;
1309
    long l;
J
jorlow@chromium.org 已提交
1310
    char junk;
1311
    char hdfsname[2048];
J
jorlow@chromium.org 已提交
1312 1313 1314 1315 1316 1317 1318
    if (leveldb::Slice(argv[i]).starts_with("--benchmarks=")) {
      FLAGS_benchmarks = argv[i] + strlen("--benchmarks=");
    } else if (sscanf(argv[i], "--compression_ratio=%lf%c", &d, &junk) == 1) {
      FLAGS_compression_ratio = d;
    } else if (sscanf(argv[i], "--histogram=%d%c", &n, &junk) == 1 &&
               (n == 0 || n == 1)) {
      FLAGS_histogram = n;
1319 1320 1321
    } else if (sscanf(argv[i], "--use_existing_db=%d%c", &n, &junk) == 1 &&
               (n == 0 || n == 1)) {
      FLAGS_use_existing_db = n;
1322 1323
    } else if (sscanf(argv[i], "--num=%ld%c", &l, &junk) == 1) {
      FLAGS_num = l;
1324 1325
    } else if (sscanf(argv[i], "--reads=%d%c", &n, &junk) == 1) {
      FLAGS_reads = n;
1326 1327 1328 1329
    } else if (sscanf(argv[i], "--read_range=%d%c", &n, &junk) == 1) {
      FLAGS_read_range = n;
    } else if (sscanf(argv[i], "--seed=%ld%c", &l, &junk) == 1) {
      FLAGS_seed = l;
1330 1331
    } else if (sscanf(argv[i], "--threads=%d%c", &n, &junk) == 1) {
      FLAGS_threads = n;
J
jorlow@chromium.org 已提交
1332 1333 1334 1335
    } else if (sscanf(argv[i], "--value_size=%d%c", &n, &junk) == 1) {
      FLAGS_value_size = n;
    } else if (sscanf(argv[i], "--write_buffer_size=%d%c", &n, &junk) == 1) {
      FLAGS_write_buffer_size = n;
1336 1337 1338 1339
    } else if (sscanf(argv[i], "--max_write_buffer_number=%d%c", &n, &junk) == 1) {
      FLAGS_max_write_buffer_number = n;
    } else if (sscanf(argv[i], "--max_background_compactions=%d%c", &n, &junk) == 1) {
      FLAGS_max_background_compactions = n;
1340 1341
    } else if (sscanf(argv[i], "--cache_size=%ld%c", &l, &junk) == 1) {
      FLAGS_cache_size = l;
1342 1343
    } else if (sscanf(argv[i], "--block_size=%d%c", &n, &junk) == 1) {
      FLAGS_block_size = n;
1344
    } else if (sscanf(argv[i], "--cache_numshardbits=%d%c", &n, &junk) == 1) {
1345 1346 1347 1348 1349 1350
      if (n < 20) {
        FLAGS_cache_numshardbits = n;
      } else {
        fprintf(stderr, "The cache cannot be sharded into 2**%d pieces\n", n);
        exit(1);
      }
1351
    } else if (sscanf(argv[i], "--table_cache_numshardbits=%d%c",
A
Abhishek Kona 已提交
1352
          &n, &junk) == 1) {
1353 1354 1355 1356 1357
      if (n <= 0 || n > 20) {
        fprintf(stderr, "The cache cannot be sharded into 2**%d pieces\n", n);
        exit(1);
      }
      FLAGS_table_cache_numshardbits = n;
S
Sanjay Ghemawat 已提交
1358 1359
    } else if (sscanf(argv[i], "--bloom_bits=%d%c", &n, &junk) == 1) {
      FLAGS_bloom_bits = n;
1360 1361
    } else if (sscanf(argv[i], "--open_files=%d%c", &n, &junk) == 1) {
      FLAGS_open_files = n;
1362 1363
    } else if (strncmp(argv[i], "--db=", 5) == 0) {
      FLAGS_db = argv[i] + 5;
1364 1365 1366
    } else if (sscanf(argv[i], "--verify_checksum=%d%c", &n, &junk) == 1 &&
               (n == 0 || n == 1)) {
      FLAGS_verify_checksum = n;
1367 1368 1369
    } else if (sscanf(argv[i], "--bufferedio=%d%c", &n, &junk) == 1 &&
               (n == 0 || n == 1)) {
      useOsBuffer = n;
1370 1371 1372
    } else if (sscanf(argv[i], "--mmap_read=%d%c", &n, &junk) == 1 &&
               (n == 0 || n == 1)) {
      useMmapRead = n;
1373 1374 1375
    } else if (sscanf(argv[i], "--mmap_write=%d%c", &n, &junk) == 1 &&
               (n == 0 || n == 1)) {
      useMmapWrite = n;
1376
    } else if (sscanf(argv[i], "--readahead=%d%c", &n, &junk) == 1 &&
1377 1378
               (n == 0 || n == 1)) {
      useFsReadAhead = n;
1379 1380 1381 1382 1383 1384
    } else if (sscanf(argv[i], "--statistics=%d%c", &n, &junk) == 1 &&
               (n == 0 || n == 1)) {
      if (n == 1) {
        dbstats = new leveldb::DBStatistics();
        FLAGS_statistics = true;
      }
1385 1386 1387 1388 1389
    } else if (sscanf(argv[i], "--writes=%d%c", &n, &junk) == 1) {
      FLAGS_writes = n;
    } else if (sscanf(argv[i], "--sync=%d%c", &n, &junk) == 1 &&
               (n == 0 || n == 1)) {
      FLAGS_sync = n;
1390
    } else if (sscanf(argv[i], "--readwritepercent=%d%c", &n, &junk) == 1 &&
1391
               n > 0 && n < 100) {
1392
      FLAGS_readwritepercent = n;
H
heyongqiang 已提交
1393 1394 1395
    } else if (sscanf(argv[i], "--disable_data_sync=%d%c", &n, &junk) == 1 &&
        (n == 0 || n == 1)) {
      FLAGS_disable_data_sync = n;
1396 1397 1398
    } else if (sscanf(argv[i], "--use_fsync=%d%c", &n, &junk) == 1 &&
        (n == 0 || n == 1)) {
      FLAGS_use_fsync = n;
H
heyongqiang 已提交
1399
    } else if (sscanf(argv[i], "--disable_wal=%d%c", &n, &junk) == 1 &&
1400
        (n == 0 || n == 1)) {
H
heyongqiang 已提交
1401
      FLAGS_disable_wal = n;
1402
    } else if (sscanf(argv[i], "--hdfs=%s", hdfsname) == 1) {
1403
      FLAGS_env  = new leveldb::HdfsEnv(hdfsname);
1404 1405 1406
    } else if (sscanf(argv[i], "--num_levels=%d%c",
        &n, &junk) == 1) {
      FLAGS_num_levels = n;
H
heyongqiang 已提交
1407 1408 1409 1410 1411 1412 1413
    } else if (sscanf(argv[i], "--target_file_size_base=%d%c",
        &n, &junk) == 1) {
      FLAGS_target_file_size_base = n;
    } else if ( sscanf(argv[i], "--target_file_size_multiplier=%d%c",
        &n, &junk) == 1) {
      FLAGS_target_file_size_multiplier = n;
    } else if (
1414 1415
        sscanf(argv[i], "--max_bytes_for_level_base=%ld%c", &l, &junk) == 1) {
      FLAGS_max_bytes_for_level_base = l;
H
heyongqiang 已提交
1416 1417 1418
    } else if (sscanf(argv[i], "--max_bytes_for_level_multiplier=%d%c",
        &n, &junk) == 1) {
      FLAGS_max_bytes_for_level_multiplier = n;
H
heyongqiang 已提交
1419 1420 1421 1422 1423 1424
    } else if (sscanf(argv[i],"--level0_stop_writes_trigger=%d%c",
        &n, &junk) == 1) {
      FLAGS_level0_stop_writes_trigger = n;
    } else if (sscanf(argv[i],"--level0_slowdown_writes_trigger=%d%c",
        &n, &junk) == 1) {
      FLAGS_level0_slowdown_writes_trigger = n;
1425 1426 1427
    } else if (sscanf(argv[i],"--level0_file_num_compaction_trigger=%d%c",
        &n, &junk) == 1) {
      FLAGS_level0_file_num_compaction_trigger = n;
1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440
    } else if (strncmp(argv[i], "--compression_type=", 19) == 0) {
      const char* ctype = argv[i] + 19;
      if (!strcasecmp(ctype, "none"))
        FLAGS_compression_type = leveldb::kNoCompression;
      else if (!strcasecmp(ctype, "snappy"))
        FLAGS_compression_type = leveldb::kSnappyCompression;
      else if (!strcasecmp(ctype, "zlib"))
        FLAGS_compression_type = leveldb::kZlibCompression;
      else if (!strcasecmp(ctype, "bzip2"))
        FLAGS_compression_type = leveldb::kBZip2Compression;
      else {
        fprintf(stdout, "Cannot parse %s\n", argv[i]);
      }
1441 1442
    } else if (sscanf(argv[i], "--min_level_to_compress=%d%c", &n, &junk) == 1
        && n >= 0) {
1443
      FLAGS_min_level_to_compress = n;
1444 1445 1446
    } else if (sscanf(argv[i], "--disable_seek_compaction=%d%c", &n, &junk) == 1
        && (n == 0 || n == 1)) {
      FLAGS_disable_seek_compaction = n;
1447 1448
    } else if (sscanf(argv[i], "--delete_obsolete_files_period_micros=%ld%c",
                      &l, &junk) == 1) {
1449
      FLAGS_delete_obsolete_files_period_micros = l;
1450 1451 1452
    } else if (sscanf(argv[i], "--stats_interval=%d%c", &n, &junk) == 1 &&
               n >= 0 && n < 2000000000) {
      FLAGS_stats_interval = n;
1453 1454 1455 1456 1457 1458
    } else if (sscanf(argv[i], "--stats_per_interval=%d%c", &n, &junk) == 1
        && (n == 0 || n == 1)) {
      FLAGS_stats_per_interval = n;
    } else if (sscanf(argv[i], "--rate_limit=%lf%c", &d, &junk) == 1 &&
               d > 1.0) {
      FLAGS_rate_limit = d;
H
heyongqiang 已提交
1459 1460 1461
    } else if (sscanf(argv[i], "--readonly=%d%c", &n, &junk) == 1 &&
        (n == 0 || n ==1 )) {
      FLAGS_read_only = n;
1462 1463 1464
    } else if (sscanf(argv[i], "--max_grandparent_overlap_factor=%d%c",
               &n, &junk) == 1) {
      FLAGS_max_grandparent_overlap_factor = n;
A
Abhishek Kona 已提交
1465
    } else if (sscanf(argv[i], "--disable_auto_compactions=%d%c",
1466 1467
               &n, &junk) == 1 && (n == 0 || n ==1)) {
      FLAGS_disable_auto_compactions = n;
A
Abhishek Kona 已提交
1468
    } else if (sscanf(argv[i], "--source_compaction_factor=%d%c",
1469 1470
               &n, &junk) == 1 && n > 0) {
      FLAGS_source_compaction_factor = n;
1471 1472
    } else if (sscanf(argv[i], "--wal_ttl=%d%c", &n, &junk) == 1) {
      FLAGS_WAL_ttl_seconds = static_cast<uint64_t>(n);
1473
    } else {
J
jorlow@chromium.org 已提交
1474 1475 1476 1477 1478
      fprintf(stderr, "Invalid flag '%s'\n", argv[i]);
      exit(1);
    }
  }

1479 1480 1481 1482
  // The number of background threads should be at least as much the
  // max number of concurrent compactions.
  FLAGS_env->SetBackgroundThreads(FLAGS_max_background_compactions);

H
heyongqiang 已提交
1483 1484 1485 1486 1487 1488 1489
  // Choose a location for the test database if none given with --db=<path>
  if (FLAGS_db == NULL) {
      leveldb::Env::Default()->GetTestDirectory(&default_db_path);
      default_db_path += "/dbbench";
      FLAGS_db = default_db_path.c_str();
  }

J
jorlow@chromium.org 已提交
1490 1491 1492 1493
  leveldb::Benchmark benchmark;
  benchmark.Run();
  return 0;
}