Revision 4a21b1402cb454f6e68d4c6f288d73700b3ff9db authored by Islam AbdelRahman on 01 December 2016, 21:30:01 UTC, committed by Facebook Github Bot on 01 December 2016, 21:39:14 UTC
Summary: Reduce number of comparisons in heap by caching which child node in the first level is smallest (left_child or right_child) So next time we can compare directly against the smallest child I see that the total number of calls to comparator drops significantly when using this optimization Before caching (~2mil key comparison for iterating the DB) ``` $ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq" --db="/dev/shm/heap_opt" --use_existing_db --disable_auto_compactions --cache_size=1000000000 --perf_level=2 readseq : 0.338 micros/op 2959201 ops/sec; 327.4 MB/s user_key_comparison_count = 2000008 ``` After caching (~1mil key comparison for iterating the DB) ``` $ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq" --db="/dev/shm/heap_opt" --use_existing_db --disable_auto_compactions --cache_size=1000000000 --perf_level=2 readseq : 0.309 micros/op 3236801 ops/sec; 358.1 MB/s user_key_comparison_count = 1000011 ``` It also improves Closes https://github.com/facebook/rocksdb/pull/1600 Differential Revision: D4256027 Pulled By: IslamAbdelRahman fbshipit-source-id: 76fcc66
1 parent e39d080
run_flash_bench.sh
#!/bin/bash
# REQUIRE: benchmark.sh exists in the current directory
# After execution of this script, log files are generated in $output_dir.
# report.txt provides a high level statistics
# This should be run from the parent of the tools directory. The command line is:
# [$env_vars] tools/run_flash_bench.sh [list-of-threads]
#
# This runs a sequence of tests in the following sequence:
# step 1) load - bulkload, compact, fillseq, overwrite
# step 2) read-only for each number of threads
# step 3) read-write for each number of threads
# step 4) merge for each number of threads
#
# The list of threads is optional and when not set is equivalent to "24".
# Were list-of-threads specified as "1 2 4" then the tests in steps 2, 3 and
# 4 above would be repeated for 1, 2 and 4 threads. The tests in step 1 are
# only run for 1 thread.
# Test output is written to $OUTPUT_DIR, currently /tmp/output. The performance
# summary is in $OUTPUT_DIR/report.txt. There is one file in $OUTPUT_DIR per
# test and the tests are listed below.
#
# The environment variables are also optional. The variables are:
#
# NKEYS - number of key/value pairs to load
# BG_MBWRITEPERSEC - write rate limit in MB/second for tests in which
# there is one thread doing writes and stats are
# reported for read threads. "BG" stands for background.
# If this is too large then the non-writer threads can get
# starved. This is used for the "readwhile" tests.
# FG_MBWRITEPERSEC - write rate limit in MB/second for tests like overwrite
# where stats are reported for the write threads.
# NSECONDS - number of seconds for which to run each test in steps 2,
# 3 and 4. There are currently 15 tests in those steps and
# they are repeated for each entry in list-of-threads so
# this variable lets you control the total duration to
# finish the benchmark.
# RANGE_LIMIT - the number of rows to read per range query for tests that
# do range queries.
# VAL_SIZE - the length of the value in the key/value pairs loaded.
# You can estimate the size of the test database from this,
# NKEYS and the compression rate (--compression_ratio) set
# in tools/benchmark.sh
# BLOCK_LENGTH - value for db_bench --block_size
# CACHE_BYTES - the size of the RocksDB block cache in bytes
# DATA_DIR - directory in which to create database files
# LOG_DIR - directory in which to create WAL files, may be the same
# as DATA_DIR
# DO_SETUP - when set to 0 then a backup of the database is copied from
# $DATA_DIR.bak to $DATA_DIR and the load tests from step 1
# The WAL directory is also copied from a backup if
# DATA_DIR != LOG_DIR. This allows tests from steps 2, 3, 4
# to be repeated faster.
# SAVE_SETUP - saves a copy of the database at the end of step 1 to
# $DATA_DIR.bak. When LOG_DIR != DATA_DIR then it is copied
# to $LOG_DIR.bak.
# SKIP_LOW_PRI_TESTS - skip some of the tests which aren't crucial for getting
# actionable benchmarking data (look for keywords "bulkload",
# "sync=1", and "while merging").
#
# Size constants
K=1024
M=$((1024 * K))
G=$((1024 * M))
num_keys=${NKEYS:-$((1 * G))}
# write rate for readwhile... tests
bg_mbwps=${BG_MBWRITEPERSEC:-4}
# write rate for tests other than readwhile, 0 means no limit
fg_mbwps=${FG_MBWRITEPERSEC:-0}
duration=${NSECONDS:-$((60 * 60))}
nps=${RANGE_LIMIT:-10}
vs=${VAL_SIZE:-400}
cs=${CACHE_BYTES:-$(( 1 * G ))}
bs=${BLOCK_LENGTH:-8192}
# If no command line arguments then run for 24 threads.
if [[ $# -eq 0 ]]; then
nthreads=( 24 )
else
nthreads=( "$@" )
fi
for num_thr in "${nthreads[@]}" ; do
echo Will run for $num_thr threads
done
# Update these parameters before execution !!!
db_dir=${DATA_DIR:-"/tmp/rocksdb/"}
wal_dir=${LOG_DIR:-"/tmp/rocksdb/"}
do_setup=${DO_SETUP:-1}
save_setup=${SAVE_SETUP:-0}
# By default we'll run all the tests. Set this to skip a set of tests which
# aren't critical for getting key metrics.
skip_low_pri_tests=${SKIP_LOW_PRI_TESTS:-0}
if [[ $skip_low_pri_tests == 1 ]]; then
echo "Skipping some non-critical tests because SKIP_LOW_PRI_TESTS is set."
fi
output_dir="/tmp/output"
ARGS="\
OUTPUT_DIR=$output_dir \
NUM_KEYS=$num_keys \
DB_DIR=$db_dir \
WAL_DIR=$wal_dir \
VALUE_SIZE=$vs \
BLOCK_SIZE=$bs \
CACHE_SIZE=$cs"
mkdir -p $output_dir
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" \
> $output_dir/report.txt
# Notes on test sequence:
# step 1) Setup database via sequential fill followed by overwrite to fragment it.
# Done without setting DURATION to make sure that overwrite does $num_keys writes
# step 2) read-only tests for all levels of concurrency requested
# step 3) non read-only tests for all levels of concurrency requested
# step 4) merge tests for all levels of concurrency requested. These must come last.
###### Setup the database
if [[ $do_setup != 0 ]]; then
echo Doing setup
if [[ $skip_low_pri_tests != 1 ]]; then
# Test 1: bulk load
env $ARGS ./tools/benchmark.sh bulkload
fi
# Test 2a: sequential fill with large values to get peak ingest
# adjust NUM_KEYS given the use of larger values
env $ARGS BLOCK_SIZE=$((1 * M)) VALUE_SIZE=$((32 * K)) NUM_KEYS=$(( num_keys / 64 )) \
./tools/benchmark.sh fillseq_disable_wal
# Test 2b: sequential fill with the configured value size
env $ARGS ./tools/benchmark.sh fillseq_disable_wal
# Test 2c: same as 2a, but with WAL being enabled.
env $ARGS BLOCK_SIZE=$((1 * M)) VALUE_SIZE=$((32 * K)) NUM_KEYS=$(( num_keys / 64 )) \
./tools/benchmark.sh fillseq_enable_wal
# Test 2d: same as 2b, but with WAL being enabled.
env $ARGS ./tools/benchmark.sh fillseq_enable_wal
# Test 3: single-threaded overwrite
env $ARGS NUM_THREADS=1 DB_BENCH_NO_SYNC=1 ./tools/benchmark.sh overwrite
else
echo Restoring from backup
rm -rf $db_dir
if [ ! -d ${db_dir}.bak ]; then
echo Database backup does not exist at ${db_dir}.bak
exit -1
fi
echo Restore database from ${db_dir}.bak
cp -p -r ${db_dir}.bak $db_dir
if [[ $db_dir != $wal_dir ]]; then
rm -rf $wal_dir
if [ ! -d ${wal_dir}.bak ]; then
echo WAL backup does not exist at ${wal_dir}.bak
exit -1
fi
echo Restore WAL from ${wal_dir}.bak
cp -p -r ${wal_dir}.bak $wal_dir
fi
fi
if [[ $save_setup != 0 ]]; then
echo Save database to ${db_dir}.bak
cp -p -r $db_dir ${db_dir}.bak
if [[ $db_dir != $wal_dir ]]; then
echo Save WAL to ${wal_dir}.bak
cp -p -r $wal_dir ${wal_dir}.bak
fi
fi
###### Read-only tests
for num_thr in "${nthreads[@]}" ; do
# Test 4: random read
env $ARGS DURATION=$duration NUM_THREADS=$num_thr ./tools/benchmark.sh readrandom
# Test 5: random range scans
env $ARGS DURATION=$duration NUM_THREADS=$num_thr NUM_NEXTS_PER_SEEK=$nps \
./tools/benchmark.sh fwdrange
# Test 6: random reverse range scans
env $ARGS DURATION=$duration NUM_THREADS=$num_thr NUM_NEXTS_PER_SEEK=$nps \
./tools/benchmark.sh revrange
done
###### Non read-only tests
for num_thr in "${nthreads[@]}" ; do
# Test 7: overwrite with sync=0
env $ARGS DURATION=$duration NUM_THREADS=$num_thr MB_WRITE_PER_SEC=$fg_mbwps \
DB_BENCH_NO_SYNC=1 ./tools/benchmark.sh overwrite
if [[ $skip_low_pri_tests != 1 ]]; then
# Test 8: overwrite with sync=1
env $ARGS DURATION=$duration NUM_THREADS=$num_thr MB_WRITE_PER_SEC=$fg_mbwps \
./tools/benchmark.sh overwrite
fi
# Test 9: random update with sync=0
env $ARGS DURATION=$duration NUM_THREADS=$num_thr DB_BENCH_NO_SYNC=1 \
./tools/benchmark.sh updaterandom
if [[ $skip_low_pri_tests != 1 ]]; then
# Test 10: random update with sync=1
env $ARGS DURATION=$duration NUM_THREADS=$num_thr ./tools/benchmark.sh updaterandom
fi
# Test 11: random read while writing
env $ARGS DURATION=$duration NUM_THREADS=$num_thr MB_WRITE_PER_SEC=$bg_mbwps \
DB_BENCH_NO_SYNC=1 ./tools/benchmark.sh readwhilewriting
# Test 12: range scan while writing
env $ARGS DURATION=$duration NUM_THREADS=$num_thr MB_WRITE_PER_SEC=$bg_mbwps \
DB_BENCH_NO_SYNC=1 NUM_NEXTS_PER_SEEK=$nps ./tools/benchmark.sh fwdrangewhilewriting
# Test 13: reverse range scan while writing
env $ARGS DURATION=$duration NUM_THREADS=$num_thr MB_WRITE_PER_SEC=$bg_mbwps \
DB_BENCH_NO_SYNC=1 NUM_NEXTS_PER_SEEK=$nps ./tools/benchmark.sh revrangewhilewriting
done
###### Merge tests
for num_thr in "${nthreads[@]}" ; do
# Test 14: random merge with sync=0
env $ARGS DURATION=$duration NUM_THREADS=$num_thr MB_WRITE_PER_SEC=$fg_mbwps \
DB_BENCH_NO_SYNC=1 ./tools/benchmark.sh mergerandom
if [[ $skip_low_pri_tests != 1 ]]; then
# Test 15: random merge with sync=1
env $ARGS DURATION=$duration NUM_THREADS=$num_thr MB_WRITE_PER_SEC=$fg_mbwps \
./tools/benchmark.sh mergerandom
# Test 16: random read while merging
env $ARGS DURATION=$duration NUM_THREADS=$num_thr MB_WRITE_PER_SEC=$bg_mbwps \
DB_BENCH_NO_SYNC=1 ./tools/benchmark.sh readwhilemerging
# Test 17: range scan while merging
env $ARGS DURATION=$duration NUM_THREADS=$num_thr MB_WRITE_PER_SEC=$bg_mbwps \
DB_BENCH_NO_SYNC=1 NUM_NEXTS_PER_SEEK=$nps ./tools/benchmark.sh fwdrangewhilemerging
# Test 18: reverse range scan while merging
env $ARGS DURATION=$duration NUM_THREADS=$num_thr MB_WRITE_PER_SEC=$bg_mbwps \
DB_BENCH_NO_SYNC=1 NUM_NEXTS_PER_SEEK=$nps ./tools/benchmark.sh revrangewhilemerging
fi
done
###### Universal compaction tests.
# Use a single thread to reduce the variability in the benchmark.
env $ARGS COMPACTION_TEST=1 NUM_THREADS=1 ./tools/benchmark.sh universal_compaction
if [[ $skip_low_pri_tests != 1 ]]; then
echo bulkload > $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep bulkload $output_dir/report.txt >> $output_dir/report2.txt
fi
echo fillseq_wal_disabled >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep fillseq.wal_disabled $output_dir/report.txt >> $output_dir/report2.txt
echo fillseq_wal_enabled >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep fillseq.wal_enabled $output_dir/report.txt >> $output_dir/report2.txt
echo overwrite sync=0 >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep overwrite $output_dir/report.txt | grep \.s0 >> $output_dir/report2.txt
if [[ $skip_low_pri_tests != 1 ]]; then
echo overwrite sync=1 >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep overwrite $output_dir/report.txt | grep \.s1 >> $output_dir/report2.txt
fi
echo updaterandom sync=0 >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep updaterandom $output_dir/report.txt | grep \.s0 >> $output_dir/report2.txt
if [[ $skip_low_pri_tests != 1 ]]; then
echo updaterandom sync=1 >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep updaterandom $output_dir/report.txt | grep \.s1 >> $output_dir/report2.txt
fi
echo mergerandom sync=0 >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep mergerandom $output_dir/report.txt | grep \.s0 >> $output_dir/report2.txt
if [[ $skip_low_pri_tests != 1 ]]; then
echo mergerandom sync=1 >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep mergerandom $output_dir/report.txt | grep \.s1 >> $output_dir/report2.txt
fi
echo readrandom >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep readrandom $output_dir/report.txt >> $output_dir/report2.txt
echo fwdrange >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep fwdrange\.t $output_dir/report.txt >> $output_dir/report2.txt
echo revrange >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep revrange\.t $output_dir/report.txt >> $output_dir/report2.txt
echo readwhile >> $output_dir/report2.txt >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep readwhilewriting $output_dir/report.txt >> $output_dir/report2.txt
if [[ $skip_low_pri_tests != 1 ]]; then
echo readwhile >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep readwhilemerging $output_dir/report.txt >> $output_dir/report2.txt
fi
echo fwdreadwhilewriting >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep fwdrangewhilewriting $output_dir/report.txt >> $output_dir/report2.txt
if [[ $skip_low_pri_tests != 1 ]]; then
echo fwdreadwhilemerging >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep fwdrangewhilemerg $output_dir/report.txt >> $output_dir/report2.txt
fi
echo revreadwhilewriting >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep revrangewhilewriting $output_dir/report.txt >> $output_dir/report2.txt
if [[ $skip_low_pri_tests != 1 ]]; then
echo revreadwhilemerging >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep revrangewhilemerg $output_dir/report.txt >> $output_dir/report2.txt
fi
cat $output_dir/report2.txt
![swh spinner](/static/img/swh-spinner.gif)
Computing file changes ...