ceph/src/rocksdb/tools/benchmark.sh

   1 #!/usr/bin/env 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           $l0_config \
 111           --max_background_jobs=20 \
 112           --max_write_buffer_number=8"
 113
 114 params_bulkload="$const_params \
 115                  --max_background_jobs=20 \
 116                  --max_write_buffer_number=8 \
 117                  --level0_file_num_compaction_trigger=$((10 * M)) \
 118                  --level0_slowdown_writes_trigger=$((10 * M)) \
 119                  --level0_stop_writes_trigger=$((10 * M))"
 120
 121 #
 122 # Tune values for level and universal compaction.
 123 # For universal compaction, these level0_* options mean total sorted of runs in
 124 # LSM. In level-based compaction, it means number of L0 files.
 125 #
 126 params_level_compact="$const_params \
 127                 --max_background_jobs=16 \
 128                 --max_write_buffer_number=4 \
 129                 --level0_file_num_compaction_trigger=4 \
 130                 --level0_slowdown_writes_trigger=16 \
 131                 --level0_stop_writes_trigger=20"
 132
 133 params_univ_compact="$const_params \
 134                 --max_background_jobs=20 \
 135                 --max_write_buffer_number=4 \
 136                 --level0_file_num_compaction_trigger=8 \
 137                 --level0_slowdown_writes_trigger=16 \
 138                 --level0_stop_writes_trigger=20"
 139
 140 function summarize_result {
 141   test_out=$1
 142   test_name=$2
 143   bench_name=$3
 144
 145   # Note that this function assumes that the benchmark executes long enough so
 146   # that "Compaction Stats" is written to stdout at least once. If it won't
 147   # happen then empty output from grep when searching for "Sum" will cause
 148   # syntax errors.
 149   uptime=$( grep ^Uptime\(secs $test_out | tail -1 | awk '{ printf "%.0f", $2 }' )
 150   stall_time=$( grep "^Cumulative stall" $test_out | tail -1  | awk '{  print $3 }' )
 151   stall_pct=$( grep "^Cumulative stall" $test_out| tail -1  | awk '{  print $5 }' )
 152   ops_sec=$( grep ^${bench_name} $test_out | awk '{ print $5 }' )
 153   mb_sec=$( grep ^${bench_name} $test_out | awk '{ print $7 }' )
 154   lo_wgb=$( grep "^  L0" $test_out | tail -1 | awk '{ print $8 }' )
 155   sum_wgb=$( grep "^ Sum" $test_out | tail -1 | awk '{ print $8 }' )
 156   sum_size=$( grep "^ Sum" $test_out | tail -1 | awk '{ printf "%.1f", $3 / 1024.0 }' )
 157   wamp=$( echo "scale=1; $sum_wgb / $lo_wgb" | bc )
 158   wmb_ps=$( echo "scale=1; ( $sum_wgb * 1024.0 ) / $uptime" | bc )
 159   usecs_op=$( grep ^${bench_name} $test_out | awk '{ printf "%.1f", $3 }' )
 160   p50=$( grep "^Percentiles:" $test_out | tail -1 | awk '{ printf "%.1f", $3 }' )
 161   p75=$( grep "^Percentiles:" $test_out | tail -1 | awk '{ printf "%.1f", $5 }' )
 162   p99=$( grep "^Percentiles:" $test_out | tail -1 | awk '{ printf "%.0f", $7 }' )
 163   p999=$( grep "^Percentiles:" $test_out | tail -1 | awk '{ printf "%.0f", $9 }' )
 164   p9999=$( grep "^Percentiles:" $test_out | tail -1 | awk '{ printf "%.0f", $11 }' )
 165   echo -e "$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" \
 166     >> $output_dir/report.txt
 167 }
 168
 169 function run_bulkload {
 170   # This runs with a vector memtable and the WAL disabled to load faster. It is still crash safe and the
 171   # client can discover where to restart a load after a crash. I think this is a good way to load.
 172   echo "Bulk loading $num_keys random keys"
 173   cmd="./db_bench --benchmarks=fillrandom \
 174        --use_existing_db=0 \
 175        --disable_auto_compactions=1 \
 176        --sync=0 \
 177        $params_bulkload \
 178        --threads=1 \
 179        --memtablerep=vector \
 180        --allow_concurrent_memtable_write=false \
 181        --disable_wal=1 \
 182        --seed=$( date +%s ) \
 183        2>&1 | tee -a $output_dir/benchmark_bulkload_fillrandom.log"
 184   echo $cmd | tee $output_dir/benchmark_bulkload_fillrandom.log
 185   eval $cmd
 186   summarize_result $output_dir/benchmark_bulkload_fillrandom.log bulkload fillrandom
 187   echo "Compacting..."
 188   cmd="./db_bench --benchmarks=compact \
 189        --use_existing_db=1 \
 190        --disable_auto_compactions=1 \
 191        --sync=0 \
 192        $params_w \
 193        --threads=1 \
 194        2>&1 | tee -a $output_dir/benchmark_bulkload_compact.log"
 195   echo $cmd | tee $output_dir/benchmark_bulkload_compact.log
 196   eval $cmd
 197 }
 198
 199 #
 200 # Parameter description:
 201 #
 202 # $1 - 1 if I/O statistics should be collected.
 203 # $2 - compaction type to use (level=0, universal=1).
 204 # $3 - number of subcompactions.
 205 # $4 - number of maximum background compactions.
 206 #
 207 function run_manual_compaction_worker {
 208   # This runs with a vector memtable and the WAL disabled to load faster.
 209   # It is still crash safe and the client can discover where to restart a
 210   # load after a crash. I think this is a good way to load.
 211   echo "Bulk loading $num_keys random keys for manual compaction."
 212
 213   fillrandom_output_file=$output_dir/benchmark_man_compact_fillrandom_$3.log
 214   man_compact_output_log=$output_dir/benchmark_man_compact_$3.log
 215
 216   if [ "$2" == "1" ]; then
 217     extra_params=$params_univ_compact
 218   else
 219     extra_params=$params_level_compact
 220   fi
 221
 222   # Make sure that fillrandom uses the same compaction options as compact.
 223   cmd="./db_bench --benchmarks=fillrandom \
 224        --use_existing_db=0 \
 225        --disable_auto_compactions=0 \
 226        --sync=0 \
 227        $extra_params \
 228        --threads=$num_threads \
 229        --compaction_measure_io_stats=$1 \
 230        --compaction_style=$2 \
 231        --subcompactions=$3 \
 232        --memtablerep=vector \
 233        --allow_concurrent_memtable_write=false \
 234        --disable_wal=1 \
 235        --max_background_jobs=$4 \
 236        --seed=$( 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 $cmd
 265
 266   # Can't use summarize_result here. One way to analyze the results is to run
 267   # "grep real" on the resulting log files.
 268 }
 269
 270 function run_univ_compaction {
 271   # Always ask for I/O statistics to be measured.
 272   io_stats=1
 273
 274   # Values: kCompactionStyleLevel = 0x0, kCompactionStyleUniversal = 0x1.
 275   compaction_style=1
 276
 277   # Define a set of benchmarks.
 278   subcompactions=(1 2 4 8 16)
 279   max_background_jobs=(20 20 10 5 4)
 280
 281   i=0
 282   total=${#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       ${max_background_jobs[$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        --allow_concurrent_memtable_write=false \
 319        --disable_wal=$1 \
 320        --seed=$( date +%s ) \
 321        2>&1 | tee -a $log_file_name"
 322   echo $cmd | tee $log_file_name
 323   eval $cmd
 324
 325   # The constant "fillseq" which we pass to db_bench is the benchmark name.
 326   summarize_result $log_file_name $test_name fillseq
 327 }
 328
 329 function run_change {
 330   operation=$1
 331   echo "Do $num_keys random $operation"
 332   out_name="benchmark_${operation}.t${num_threads}.s${syncval}.log"
 333   cmd="./db_bench --benchmarks=$operation \
 334        --use_existing_db=1 \
 335        --sync=$syncval \
 336        $params_w \
 337        --threads=$num_threads \
 338        --merge_operator=\"put\" \
 339        --seed=$( date +%s ) \
 340        2>&1 | tee -a $output_dir/${out_name}"
 341   echo $cmd | tee $output_dir/${out_name}
 342   eval $cmd
 343   summarize_result $output_dir/${out_name} ${operation}.t${num_threads}.s${syncval} $operation
 344 }
 345
 346 function run_filluniquerandom {
 347   echo "Loading $num_keys unique keys randomly"
 348   cmd="./db_bench --benchmarks=filluniquerandom \
 349        --use_existing_db=0 \
 350        --sync=0 \
 351        $params_w \
 352        --threads=1 \
 353        --seed=$( date +%s ) \
 354        2>&1 | tee -a $output_dir/benchmark_filluniquerandom.log"
 355   echo $cmd | tee $output_dir/benchmark_filluniquerandom.log
 356   eval $cmd
 357   summarize_result $output_dir/benchmark_filluniquerandom.log filluniquerandom filluniquerandom
 358 }
 359
 360 function run_readrandom {
 361   echo "Reading $num_keys random keys"
 362   out_name="benchmark_readrandom.t${num_threads}.log"
 363   cmd="./db_bench --benchmarks=readrandom \
 364        --use_existing_db=1 \
 365        $params_w \
 366        --threads=$num_threads \
 367        --seed=$( date +%s ) \
 368        2>&1 | tee -a $output_dir/${out_name}"
 369   echo $cmd | tee $output_dir/${out_name}
 370   eval $cmd
 371   summarize_result $output_dir/${out_name} readrandom.t${num_threads} readrandom
 372 }
 373
 374 function run_readwhile {
 375   operation=$1
 376   echo "Reading $num_keys random keys while $operation"
 377   out_name="benchmark_readwhile${operation}.t${num_threads}.log"
 378   cmd="./db_bench --benchmarks=readwhile${operation} \
 379        --use_existing_db=1 \
 380        --sync=$syncval \
 381        $params_w \
 382        --threads=$num_threads \
 383        --merge_operator=\"put\" \
 384        --seed=$( date +%s ) \
 385        2>&1 | tee -a $output_dir/${out_name}"
 386   echo $cmd | tee $output_dir/${out_name}
 387   eval $cmd
 388   summarize_result $output_dir/${out_name} readwhile${operation}.t${num_threads} readwhile${operation}
 389 }
 390
 391 function run_rangewhile {
 392   operation=$1
 393   full_name=$2
 394   reverse_arg=$3
 395   out_name="benchmark_${full_name}.t${num_threads}.log"
 396   echo "Range scan $num_keys random keys while ${operation} for reverse_iter=${reverse_arg}"
 397   cmd="./db_bench --benchmarks=seekrandomwhile${operation} \
 398        --use_existing_db=1 \
 399        --sync=$syncval \
 400        $params_w \
 401        --threads=$num_threads \
 402        --merge_operator=\"put\" \
 403        --seek_nexts=$num_nexts_per_seek \
 404        --reverse_iterator=$reverse_arg \
 405        --seed=$( date +%s ) \
 406        2>&1 | tee -a $output_dir/${out_name}"
 407   echo $cmd | tee $output_dir/${out_name}
 408   eval $cmd
 409   summarize_result $output_dir/${out_name} ${full_name}.t${num_threads} seekrandomwhile${operation}
 410 }
 411
 412 function run_range {
 413   full_name=$1
 414   reverse_arg=$2
 415   out_name="benchmark_${full_name}.t${num_threads}.log"
 416   echo "Range scan $num_keys random keys for reverse_iter=${reverse_arg}"
 417   cmd="./db_bench --benchmarks=seekrandom \
 418        --use_existing_db=1 \
 419        $params_w \
 420        --threads=$num_threads \
 421        --seek_nexts=$num_nexts_per_seek \
 422        --reverse_iterator=$reverse_arg \
 423        --seed=$( date +%s ) \
 424        2>&1 | tee -a $output_dir/${out_name}"
 425   echo $cmd | tee $output_dir/${out_name}
 426   eval $cmd
 427   summarize_result $output_dir/${out_name} ${full_name}.t${num_threads} seekrandom
 428 }
 429
 430 function run_randomtransaction {
 431   echo "..."
 432   cmd="./db_bench $params_r --benchmarks=randomtransaction \
 433        --num=$num_keys \
 434        --transaction_db \
 435        --threads=5 \
 436        --transaction_sets=5 \
 437        2>&1 | tee $output_dir/benchmark_randomtransaction.log"
 438   echo $cmd | tee $output_dir/benchmark_rangescanwhilewriting.log
 439   eval $cmd
 440 }
 441
 442 function now() {
 443   echo `date +"%s"`
 444 }
 445
 446 report="$output_dir/report.txt"
 447 schedule="$output_dir/schedule.txt"
 448
 449 echo "===== Benchmark ====="
 450
 451 # Run!!!
 452 IFS=',' read -a jobs <<< $1
 453 # shellcheck disable=SC2068
 454 for job in ${jobs[@]}; do
 455
 456   if [ $job != debug ]; then
 457     echo "Start $job at `date`" | tee -a $schedule
 458   fi
 459
 460   start=$(now)
 461   if [ $job = bulkload ]; then
 462     run_bulkload
 463   elif [ $job = fillseq_disable_wal ]; then
 464     run_fillseq 1
 465   elif [ $job = fillseq_enable_wal ]; then
 466     run_fillseq 0
 467   elif [ $job = overwrite ]; then
 468     run_change overwrite
 469   elif [ $job = updaterandom ]; then
 470     run_change updaterandom
 471   elif [ $job = mergerandom ]; then
 472     run_change mergerandom
 473   elif [ $job = filluniquerandom ]; then
 474     run_filluniquerandom
 475   elif [ $job = readrandom ]; then
 476     run_readrandom
 477   elif [ $job = fwdrange ]; then
 478     run_range $job false
 479   elif [ $job = revrange ]; then
 480     run_range $job true
 481   elif [ $job = readwhilewriting ]; then
 482     run_readwhile writing
 483   elif [ $job = readwhilemerging ]; then
 484     run_readwhile merging
 485   elif [ $job = fwdrangewhilewriting ]; then
 486     run_rangewhile writing $job false
 487   elif [ $job = revrangewhilewriting ]; then
 488     run_rangewhile writing $job true
 489   elif [ $job = fwdrangewhilemerging ]; then
 490     run_rangewhile merging $job false
 491   elif [ $job = revrangewhilemerging ]; then
 492     run_rangewhile merging $job true
 493   elif [ $job = randomtransaction ]; then
 494     run_randomtransaction
 495   elif [ $job = universal_compaction ]; then
 496     run_univ_compaction
 497   elif [ $job = debug ]; then
 498     num_keys=1000; # debug
 499     echo "Setting num_keys to $num_keys"
 500   else
 501     echo "unknown job $job"
 502     exit
 503   fi
 504   end=$(now)
 505
 506   if [ $job != debug ]; then
 507     echo "Complete $job in $((end-start)) seconds" | tee -a $schedule
 508   fi
 509
 510   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"
 511   tail -1 $output_dir/report.txt
 512
 513 done