ceph/src/rocksdb/tools/benchmark.sh

   1 #!/bin/bash
   2 # REQUIRE: db_bench binary exists in the current directory
   3
   4 if [ $# -ne 1 ]; then
   5   echo -n "./benchmark.sh [bulkload/fillseq/overwrite/filluniquerandom/"
   6   echo    "readrandom/readwhilewriting/readwhilemerging/updaterandom/"
   7   echo    "mergerandom/randomtransaction/compact]"
   8   exit 0
   9 fi
  10
  11 # Make it easier to run only the compaction test. Getting valid data requires
  12 # a number of iterations and having an ability to run the test separately from
  13 # rest of the benchmarks helps.
  14 if [ "$COMPACTION_TEST" == "1" -a "$1" != "universal_compaction" ]; then
  15   echo "Skipping $1 because it's not a compaction test."
  16   exit 0
  17 fi
  18
  19 # size constants
  20 K=1024
  21 M=$((1024 * K))
  22 G=$((1024 * M))
  23
  24 if [ -z $DB_DIR ]; then
  25   echo "DB_DIR is not defined"
  26   exit 0
  27 fi
  28
  29 if [ -z $WAL_DIR ]; then
  30   echo "WAL_DIR is not defined"
  31   exit 0
  32 fi
  33
  34 output_dir=${OUTPUT_DIR:-/tmp/}
  35 if [ ! -d $output_dir ]; then
  36   mkdir -p $output_dir
  37 fi
  38
  39 # all multithreaded tests run with sync=1 unless
  40 # $DB_BENCH_NO_SYNC is defined
  41 syncval="1"
  42 if [ ! -z $DB_BENCH_NO_SYNC ]; then
  43   echo "Turning sync off for all multithreaded tests"
  44   syncval="0";
  45 fi
  46
  47 num_threads=${NUM_THREADS:-16}
  48 mb_written_per_sec=${MB_WRITE_PER_SEC:-0}
  49 # Only for tests that do range scans
  50 num_nexts_per_seek=${NUM_NEXTS_PER_SEEK:-10}
  51 cache_size=${CACHE_SIZE:-$((1 * G))}
  52 compression_max_dict_bytes=${COMPRESSION_MAX_DICT_BYTES:-0}
  53 compression_type=${COMPRESSION_TYPE:-snappy}
  54 duration=${DURATION:-0}
  55
  56 num_keys=${NUM_KEYS:-$((1 * G))}
  57 key_size=${KEY_SIZE:-20}
  58 value_size=${VALUE_SIZE:-400}
  59 block_size=${BLOCK_SIZE:-8192}
  60
  61 const_params="
  62   --db=$DB_DIR \
  63   --wal_dir=$WAL_DIR \
  64   \
  65   --num=$num_keys \
  66   --num_levels=6 \
  67   --key_size=$key_size \
  68   --value_size=$value_size \
  69   --block_size=$block_size \
  70   --cache_size=$cache_size \
  71   --cache_numshardbits=6 \
  72   --compression_max_dict_bytes=$compression_max_dict_bytes \
  73   --compression_ratio=0.5 \
  74   --compression_type=$compression_type \
  75   --level_compaction_dynamic_level_bytes=true \
  76   --bytes_per_sync=$((8 * M)) \
  77   --cache_index_and_filter_blocks=0 \
  78   --pin_l0_filter_and_index_blocks_in_cache=1 \
  79   --benchmark_write_rate_limit=$(( 1024 * 1024 * $mb_written_per_sec )) \
  80   \
  81   --hard_rate_limit=3 \
  82   --rate_limit_delay_max_milliseconds=1000000 \
  83   --write_buffer_size=$((128 * M)) \
  84   --target_file_size_base=$((128 * M)) \
  85   --max_bytes_for_level_base=$((1 * G)) \
  86   \
  87   --verify_checksum=1 \
  88   --delete_obsolete_files_period_micros=$((60 * M)) \
  89   --max_bytes_for_level_multiplier=8 \
  90   \
  91   --statistics=0 \
  92   --stats_per_interval=1 \
  93   --stats_interval_seconds=60 \
  94   --histogram=1 \
  95   \
  96   --memtablerep=skip_list \
  97   --bloom_bits=10 \
  98   --open_files=-1"
  99
 100 l0_config="
 101   --level0_file_num_compaction_trigger=4 \
 102   --level0_slowdown_writes_trigger=12 \
 103   --level0_stop_writes_trigger=20"
 104
 105 if [ $duration -gt 0 ]; then
 106   const_params="$const_params --duration=$duration"
 107 fi
 108
 109 params_w="$const_params \
 110           $l0_config \
 111           --max_background_compactions=16 \
 112           --max_write_buffer_number=8 \
 113           --max_background_flushes=7"
 114
 115 params_bulkload="$const_params \
 116                  --max_background_compactions=16 \
 117                  --max_write_buffer_number=8 \
 118                  --max_background_flushes=7 \
 119                  --level0_file_num_compaction_trigger=$((10 * M)) \
 120                  --level0_slowdown_writes_trigger=$((10 * M)) \
 121                  --level0_stop_writes_trigger=$((10 * M))"
 122
 123 #
 124 # Tune values for level and universal compaction.
 125 # For universal compaction, these level0_* options mean total sorted of runs in
 126 # LSM. In level-based compaction, it means number of L0 files.
 127 #
 128 params_level_compact="$const_params \
 129                 --max_background_flushes=4 \
 130                 --max_write_buffer_number=4 \
 131                 --level0_file_num_compaction_trigger=4 \
 132                 --level0_slowdown_writes_trigger=16 \
 133                 --level0_stop_writes_trigger=20"
 134
 135 params_univ_compact="$const_params \
 136                 --max_background_flushes=4 \
 137                 --max_write_buffer_number=4 \
 138                 --level0_file_num_compaction_trigger=8 \
 139                 --level0_slowdown_writes_trigger=16 \
 140                 --level0_stop_writes_trigger=20"
 141
 142 function summarize_result {
 143   test_out=$1
 144   test_name=$2
 145   bench_name=$3
 146
 147   # Note that this function assumes that the benchmark executes long enough so
 148   # that "Compaction Stats" is written to stdout at least once. If it won't
 149   # happen then empty output from grep when searching for "Sum" will cause
 150   # syntax errors.
 151   uptime=$( grep ^Uptime\(secs $test_out | tail -1 | awk '{ printf "%.0f", $2 }' )
 152   stall_time=$( grep "^Cumulative stall" $test_out | tail -1  | awk '{  print $3 }' )
 153   stall_pct=$( grep "^Cumulative stall" $test_out| tail -1  | awk '{  print $5 }' )
 154   ops_sec=$( grep ^${bench_name} $test_out | awk '{ print $5 }' )
 155   mb_sec=$( grep ^${bench_name} $test_out | awk '{ print $7 }' )
 156   lo_wgb=$( grep "^  L0" $test_out | tail -1 | awk '{ print $8 }' )
 157   sum_wgb=$( grep "^ Sum" $test_out | tail -1 | awk '{ print $8 }' )
 158   sum_size=$( grep "^ Sum" $test_out | tail -1 | awk '{ printf "%.1f", $3 / 1024.0 }' )
 159   wamp=$( echo "scale=1; $sum_wgb / $lo_wgb" | bc )
 160   wmb_ps=$( echo "scale=1; ( $sum_wgb * 1024.0 ) / $uptime" | bc )
 161   usecs_op=$( grep ^${bench_name} $test_out | awk '{ printf "%.1f", $3 }' )
 162   p50=$( grep "^Percentiles:" $test_out | tail -1 | awk '{ printf "%.1f", $3 }' )
 163   p75=$( grep "^Percentiles:" $test_out | tail -1 | awk '{ printf "%.1f", $5 }' )
 164   p99=$( grep "^Percentiles:" $test_out | tail -1 | awk '{ printf "%.0f", $7 }' )
 165   p999=$( grep "^Percentiles:" $test_out | tail -1 | awk '{ printf "%.0f", $9 }' )
 166   p9999=$( grep "^Percentiles:" $test_out | tail -1 | awk '{ printf "%.0f", $11 }' )
 167   echo -e "$ops_sec\t$mb_sec\t$sum_size\t$lo_wgb\t$sum_wgb\t$wamp\t$wmb_ps\t$usecs_op\t$p50\t$p75\t$p99\t$p999\t$p9999\t$uptime\t$stall_time\t$stall_pct\t$test_name" \
 168     >> $output_dir/report.txt
 169 }
 170
 171 function run_bulkload {
 172   # This runs with a vector memtable and the WAL disabled to load faster. It is still crash safe and the
 173   # client can discover where to restart a load after a crash. I think this is a good way to load.
 174   echo "Bulk loading $num_keys random keys"
 175   cmd="./db_bench --benchmarks=fillrandom \
 176        --use_existing_db=0 \
 177        --disable_auto_compactions=1 \
 178        --sync=0 \
 179        $params_bulkload \
 180        --threads=1 \
 181        --memtablerep=vector \
 182        --disable_wal=1 \
 183        --seed=$( date +%s ) \
 184        2>&1 | tee -a $output_dir/benchmark_bulkload_fillrandom.log"
 185   echo $cmd | tee $output_dir/benchmark_bulkload_fillrandom.log
 186   eval $cmd
 187   summarize_result $output_dir/benchmark_bulkload_fillrandom.log bulkload fillrandom
 188   echo "Compacting..."
 189   cmd="./db_bench --benchmarks=compact \
 190        --use_existing_db=1 \
 191        --disable_auto_compactions=1 \
 192        --sync=0 \
 193        $params_w \
 194        --threads=1 \
 195        2>&1 | tee -a $output_dir/benchmark_bulkload_compact.log"
 196   echo $cmd | tee $output_dir/benchmark_bulkload_compact.log
 197   eval $cmd
 198 }
 199
 200 #
 201 # Parameter description:
 202 #
 203 # $1 - 1 if I/O statistics should be collected.
 204 # $2 - compaction type to use (level=0, universal=1).
 205 # $3 - number of subcompactions.
 206 # $4 - number of maximum background compactions.
 207 #
 208 function run_manual_compaction_worker {
 209   # This runs with a vector memtable and the WAL disabled to load faster.
 210   # It is still crash safe and the client can discover where to restart a
 211   # load after a crash. I think this is a good way to load.
 212   echo "Bulk loading $num_keys random keys for manual compaction."
 213
 214   fillrandom_output_file=$output_dir/benchmark_man_compact_fillrandom_$3.log
 215   man_compact_output_log=$output_dir/benchmark_man_compact_$3.log
 216
 217   if [ "$2" == "1" ]; then
 218     extra_params=$params_univ_compact
 219   else
 220     extra_params=$params_level_compact
 221   fi
 222
 223   # Make sure that fillrandom uses the same compaction options as compact.
 224   cmd="./db_bench --benchmarks=fillrandom \
 225        --use_existing_db=0 \
 226        --disable_auto_compactions=0 \
 227        --sync=0 \
 228        $extra_params \
 229        --threads=$num_threads \
 230        --compaction_measure_io_stats=$1 \
 231        --compaction_style=$2 \
 232        --subcompactions=$3 \
 233        --memtablerep=vector \
 234        --disable_wal=1 \
 235        --max_background_compactions=$4 \
 236        --seed=$( date +%s ) \
 237        2>&1 | tee -a $fillrandom_output_file"
 238
 239   echo $cmd | tee $fillrandom_output_file
 240   eval $cmd
 241
 242   summarize_result $fillrandom_output_file man_compact_fillrandom_$3 fillrandom
 243
 244   echo "Compacting with $3 subcompactions specified ..."
 245
 246   # This is the part we're really interested in. Given that compact benchmark
 247   # doesn't output regular statistics then we'll just use the time command to
 248   # measure how long this step takes.
 249   cmd="{ \
 250        time ./db_bench --benchmarks=compact \
 251        --use_existing_db=1 \
 252        --disable_auto_compactions=0 \
 253        --sync=0 \
 254        $extra_params \
 255        --threads=$num_threads \
 256        --compaction_measure_io_stats=$1 \
 257        --compaction_style=$2 \
 258        --subcompactions=$3 \
 259        --max_background_compactions=$4 \
 260        ;}
 261        2>&1 | tee -a $man_compact_output_log"
 262
 263   echo $cmd | tee $man_compact_output_log
 264   eval $cmd
 265
 266   # Can't use summarize_result here. One way to analyze the results is to run
 267   # "grep real" on the resulting log files.
 268 }
 269
 270 function run_univ_compaction {
 271   # Always ask for I/O statistics to be measured.
 272   io_stats=1
 273
 274   # Values: kCompactionStyleLevel = 0x0, kCompactionStyleUniversal = 0x1.
 275   compaction_style=1
 276
 277   # Define a set of benchmarks.
 278   subcompactions=(1 2 4 8 16)
 279   max_background_compactions=(16 16 8 4 2)
 280
 281   i=0
 282   total=${#subcompactions[@]}
 283
 284   # Execute a set of benchmarks to cover variety of scenarios.
 285   while [ "$i" -lt "$total" ]
 286   do
 287     run_manual_compaction_worker $io_stats $compaction_style ${subcompactions[$i]} \
 288       ${max_background_compactions[$i]}
 289     ((i++))
 290   done
 291 }
 292
 293 function run_fillseq {
 294   # This runs with a vector memtable. WAL can be either disabled or enabled
 295   # depending on the input parameter (1 for disabled, 0 for enabled). The main
 296   # benefit behind disabling WAL is to make loading faster. It is still crash
 297   # safe and the client can discover where to restart a load after a crash. I
 298   # think this is a good way to load.
 299
 300   # Make sure that we'll have unique names for all the files so that data won't
 301   # be overwritten.
 302   if [ $1 == 1 ]; then
 303     log_file_name=$output_dir/benchmark_fillseq.wal_disabled.v${value_size}.log
 304     test_name=fillseq.wal_disabled.v${value_size}
 305   else
 306     log_file_name=$output_dir/benchmark_fillseq.wal_enabled.v${value_size}.log
 307     test_name=fillseq.wal_enabled.v${value_size}
 308   fi
 309
 310   echo "Loading $num_keys keys sequentially"
 311   cmd="./db_bench --benchmarks=fillseq \
 312        --use_existing_db=0 \
 313        --sync=0 \
 314        $params_w \
 315        --min_level_to_compress=0 \
 316        --threads=1 \
 317        --memtablerep=vector \
 318        --disable_wal=$1 \
 319        --seed=$( date +%s ) \
 320        2>&1 | tee -a $log_file_name"
 321   echo $cmd | tee $log_file_name
 322   eval $cmd
 323
 324   # The constant "fillseq" which we pass to db_bench is the benchmark name.
 325   summarize_result $log_file_name $test_name fillseq
 326 }
 327
 328 function run_change {
 329   operation=$1
 330   echo "Do $num_keys random $operation"
 331   out_name="benchmark_${operation}.t${num_threads}.s${syncval}.log"
 332   cmd="./db_bench --benchmarks=$operation \
 333        --use_existing_db=1 \
 334        --sync=$syncval \
 335        $params_w \
 336        --threads=$num_threads \
 337        --merge_operator=\"put\" \
 338        --seed=$( date +%s ) \
 339        2>&1 | tee -a $output_dir/${out_name}"
 340   echo $cmd | tee $output_dir/${out_name}
 341   eval $cmd
 342   summarize_result $output_dir/${out_name} ${operation}.t${num_threads}.s${syncval} $operation
 343 }
 344
 345 function run_filluniquerandom {
 346   echo "Loading $num_keys unique keys randomly"
 347   cmd="./db_bench --benchmarks=filluniquerandom \
 348        --use_existing_db=0 \
 349        --sync=0 \
 350        $params_w \
 351        --threads=1 \
 352        --seed=$( date +%s ) \
 353        2>&1 | tee -a $output_dir/benchmark_filluniquerandom.log"
 354   echo $cmd | tee $output_dir/benchmark_filluniquerandom.log
 355   eval $cmd
 356   summarize_result $output_dir/benchmark_filluniquerandom.log filluniquerandom filluniquerandom
 357 }
 358
 359 function run_readrandom {
 360   echo "Reading $num_keys random keys"
 361   out_name="benchmark_readrandom.t${num_threads}.log"
 362   cmd="./db_bench --benchmarks=readrandom \
 363        --use_existing_db=1 \
 364        $params_w \
 365        --threads=$num_threads \
 366        --seed=$( date +%s ) \
 367        2>&1 | tee -a $output_dir/${out_name}"
 368   echo $cmd | tee $output_dir/${out_name}
 369   eval $cmd
 370   summarize_result $output_dir/${out_name} readrandom.t${num_threads} readrandom
 371 }
 372
 373 function run_readwhile {
 374   operation=$1
 375   echo "Reading $num_keys random keys while $operation"
 376   out_name="benchmark_readwhile${operation}.t${num_threads}.log"
 377   cmd="./db_bench --benchmarks=readwhile${operation} \
 378        --use_existing_db=1 \
 379        --sync=$syncval \
 380        $params_w \
 381        --threads=$num_threads \
 382        --merge_operator=\"put\" \
 383        --seed=$( date +%s ) \
 384        2>&1 | tee -a $output_dir/${out_name}"
 385   echo $cmd | tee $output_dir/${out_name}
 386   eval $cmd
 387   summarize_result $output_dir/${out_name} readwhile${operation}.t${num_threads} readwhile${operation}
 388 }
 389
 390 function run_rangewhile {
 391   operation=$1
 392   full_name=$2
 393   reverse_arg=$3
 394   out_name="benchmark_${full_name}.t${num_threads}.log"
 395   echo "Range scan $num_keys random keys while ${operation} for reverse_iter=${reverse_arg}"
 396   cmd="./db_bench --benchmarks=seekrandomwhile${operation} \
 397        --use_existing_db=1 \
 398        --sync=$syncval \
 399        $params_w \
 400        --threads=$num_threads \
 401        --merge_operator=\"put\" \
 402        --seek_nexts=$num_nexts_per_seek \
 403        --reverse_iterator=$reverse_arg \
 404        --seed=$( date +%s ) \
 405        2>&1 | tee -a $output_dir/${out_name}"
 406   echo $cmd | tee $output_dir/${out_name}
 407   eval $cmd
 408   summarize_result $output_dir/${out_name} ${full_name}.t${num_threads} seekrandomwhile${operation}
 409 }
 410
 411 function run_range {
 412   full_name=$1
 413   reverse_arg=$2
 414   out_name="benchmark_${full_name}.t${num_threads}.log"
 415   echo "Range scan $num_keys random keys for reverse_iter=${reverse_arg}"
 416   cmd="./db_bench --benchmarks=seekrandom \
 417        --use_existing_db=1 \
 418        $params_w \
 419        --threads=$num_threads \
 420        --seek_nexts=$num_nexts_per_seek \
 421        --reverse_iterator=$reverse_arg \
 422        --seed=$( date +%s ) \
 423        2>&1 | tee -a $output_dir/${out_name}"
 424   echo $cmd | tee $output_dir/${out_name}
 425   eval $cmd
 426   summarize_result $output_dir/${out_name} ${full_name}.t${num_threads} seekrandom
 427 }
 428
 429 function run_randomtransaction {
 430   echo "..."
 431   cmd="./db_bench $params_r --benchmarks=randomtransaction \
 432        --num=$num_keys \
 433        --transaction_db \
 434        --threads=5 \
 435        --transaction_sets=5 \
 436        2>&1 | tee $output_dir/benchmark_randomtransaction.log"
 437   echo $cmd | tee $output_dir/benchmark_rangescanwhilewriting.log
 438   eval $cmd
 439 }
 440
 441 function now() {
 442   echo `date +"%s"`
 443 }
 444
 445 report="$output_dir/report.txt"
 446 schedule="$output_dir/schedule.txt"
 447
 448 echo "===== Benchmark ====="
 449
 450 # Run!!!
 451 IFS=',' read -a jobs <<< $1
 452 for job in ${jobs[@]}; do
 453
 454   if [ $job != debug ]; then
 455     echo "Start $job at `date`" | tee -a $schedule
 456   fi
 457
 458   start=$(now)
 459   if [ $job = bulkload ]; then
 460     run_bulkload
 461   elif [ $job = fillseq_disable_wal ]; then
 462     run_fillseq 1
 463   elif [ $job = fillseq_enable_wal ]; then
 464     run_fillseq 0
 465   elif [ $job = overwrite ]; then
 466     run_change overwrite
 467   elif [ $job = updaterandom ]; then
 468     run_change updaterandom
 469   elif [ $job = mergerandom ]; then
 470     run_change mergerandom
 471   elif [ $job = filluniquerandom ]; then
 472     run_filluniquerandom
 473   elif [ $job = readrandom ]; then
 474     run_readrandom
 475   elif [ $job = fwdrange ]; then
 476     run_range $job false
 477   elif [ $job = revrange ]; then
 478     run_range $job true
 479   elif [ $job = readwhilewriting ]; then
 480     run_readwhile writing
 481   elif [ $job = readwhilemerging ]; then
 482     run_readwhile merging
 483   elif [ $job = fwdrangewhilewriting ]; then
 484     run_rangewhile writing $job false
 485   elif [ $job = revrangewhilewriting ]; then
 486     run_rangewhile writing $job true
 487   elif [ $job = fwdrangewhilemerging ]; then
 488     run_rangewhile merging $job false
 489   elif [ $job = revrangewhilemerging ]; then
 490     run_rangewhile merging $job true
 491   elif [ $job = randomtransaction ]; then
 492     run_randomtransaction
 493   elif [ $job = universal_compaction ]; then
 494     run_univ_compaction
 495   elif [ $job = debug ]; then
 496     num_keys=1000; # debug
 497     echo "Setting num_keys to $num_keys"
 498   else
 499     echo "unknown job $job"
 500     exit
 501   fi
 502   end=$(now)
 503
 504   if [ $job != debug ]; then
 505     echo "Complete $job in $((end-start)) seconds" | tee -a $schedule
 506   fi
 507
 508   echo -e "ops/sec\tmb/sec\tSize-GB\tL0_GB\tSum_GB\tW-Amp\tW-MB/s\tusec/op\tp50\tp75\tp99\tp99.9\tp99.99\tUptime\tStall-time\tStall%\tTest"
 509   tail -1 $output_dir/report.txt
 510
 511 done