Add scripts to run leveldb benchmark

Summary:
This runs a benchmark for LevelDB similar to what we have
in tools/run_flash_bench.sh. It requires changes to db_bench that I published
in a LevelDB fork on github.  Some results are at:
http://smalldatum.blogspot.com/2015/04/comparing-leveldb-and-rocksdb-take-2.html

Sample output:
ops/sec	mb/sec	usec/op	avg	p50	Test
525	16.4	1904.5	1904.5	111.0	fillseq.v32768
75187	15.5	13.3	13.3	4.4	fillseq.v200
28328	5.8	35.3	35.3	4.7	overwrite.t1.s0
175438	0.0	5.7	5.7	4.4	readrandom.t1
28490	5.9	35.1	35.1	4.7	overwrite.t1.s0
121951	0.0	8.2	8.2	5.7	readwhilewriting.t1

Task ID: #

Blame Rev:

Test Plan:
Revert Plan:

Database Impact:

Memcache Impact:

Other Notes:

EImportant:

- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -

Reviewers: igor

Reviewed By: igor

Subscribers: dhruba

Differential Revision: https://reviews.facebook.net/D37749
main
Mark Callaghan 10 years ago
parent 1bb4928da9
commit a087f80e9d
  1. 185
      tools/benchmark_leveldb.sh
  2. 174
      tools/run_leveldb.sh

@ -0,0 +1,185 @@
#!/bin/bash
# REQUIRE: db_bench binary exists in the current directory
#
# This should be used with the LevelDB fork listed here to use additional test options.
# For more details on the changes see the blog post listed below.
# https://github.com/mdcallag/leveldb-1
# http://smalldatum.blogspot.com/2015/04/comparing-leveldb-and-rocksdb-take-2.html
if [ $# -ne 1 ]; then
echo -n "./benchmark.sh [fillseq/overwrite/readrandom/readwhilewriting]"
exit 0
fi
# size constants
K=1024
M=$((1024 * K))
G=$((1024 * M))
if [ -z $DB_DIR ]; then
echo "DB_DIR is not defined"
exit 0
fi
output_dir=${OUTPUT_DIR:-/tmp/}
if [ ! -d $output_dir ]; then
mkdir -p $output_dir
fi
# all multithreaded tests run with sync=1 unless
# $DB_BENCH_NO_SYNC is defined
syncval="1"
if [ ! -z $DB_BENCH_NO_SYNC ]; then
echo "Turning sync off for all multithreaded tests"
syncval="0";
fi
num_threads=${NUM_THREADS:-16}
# Only for *whilewriting, *whilemerging
writes_per_second=${WRITES_PER_SECOND:-$((10 * K))}
cache_size=${CACHE_SIZE:-$((1 * G))}
num_keys=${NUM_KEYS:-$((1 * G))}
key_size=20
value_size=${VALUE_SIZE:-400}
block_size=${BLOCK_SIZE:-4096}
const_params="
--db=$DB_DIR \
\
--num=$num_keys \
--value_size=$value_size \
--cache_size=$cache_size \
--compression_ratio=0.5 \
\
--write_buffer_size=$((2 * M)) \
\
--histogram=1 \
\
--bloom_bits=10 \
--open_files=$((20 * K))"
params_w="$const_params "
function summarize_result {
test_out=$1
test_name=$2
bench_name=$3
nthr=$4
usecs_op=$( grep ^${bench_name} $test_out | awk '{ printf "%.1f", $3 }' )
mb_sec=$( grep ^${bench_name} $test_out | awk '{ printf "%.1f", $5 }' )
ops=$( grep "^Count:" $test_out | awk '{ print $2 }' )
ops_sec=$( echo "scale=0; (1000000.0 * $nthr) / $usecs_op" | bc )
avg=$( grep "^Count:" $test_out | awk '{ printf "%.1f", $4 }' )
p50=$( grep "^Min:" $test_out | awk '{ printf "%.1f", $4 }' )
echo -e "$ops_sec\t$mb_sec\t$usecs_op\t$avg\t$p50\t$test_name" \
>> $output_dir/report.txt
}
function run_fillseq {
# This runs with a vector memtable and the WAL disabled to load faster. It is still crash safe and the
# client can discover where to restart a load after a crash. I think this is a good way to load.
echo "Loading $num_keys keys sequentially"
cmd="./db_bench --benchmarks=fillseq \
--use_existing_db=0 \
--sync=0 \
$params_w \
--threads=1 \
--seed=$( date +%s ) \
2>&1 | tee -a $output_dir/benchmark_fillseq.v${value_size}.log"
echo $cmd | tee $output_dir/benchmark_fillseq.v${value_size}.log
eval $cmd
summarize_result $output_dir/benchmark_fillseq.v${value_size}.log fillseq.v${value_size} fillseq 1
}
function run_change {
operation=$1
echo "Do $num_keys random $operation"
out_name="benchmark_${operation}.t${num_threads}.s${syncval}.log"
cmd="./db_bench --benchmarks=$operation \
--use_existing_db=1 \
--sync=$syncval \
$params_w \
--threads=$num_threads \
--seed=$( date +%s ) \
2>&1 | tee -a $output_dir/${out_name}"
echo $cmd | tee $output_dir/${out_name}
eval $cmd
summarize_result $output_dir/${out_name} ${operation}.t${num_threads}.s${syncval} $operation $num_threads
}
function run_readrandom {
echo "Reading $num_keys random keys"
out_name="benchmark_readrandom.t${num_threads}.log"
cmd="./db_bench --benchmarks=readrandom \
--use_existing_db=1 \
$params_w \
--threads=$num_threads \
--seed=$( date +%s ) \
2>&1 | tee -a $output_dir/${out_name}"
echo $cmd | tee $output_dir/${out_name}
eval $cmd
summarize_result $output_dir/${out_name} readrandom.t${num_threads} readrandom $num_threads
}
function run_readwhile {
operation=$1
echo "Reading $num_keys random keys while $operation"
out_name="benchmark_readwhile${operation}.t${num_threads}.log"
cmd="./db_bench --benchmarks=readwhile${operation} \
--use_existing_db=1 \
--sync=$syncval \
$params_w \
--threads=$num_threads \
--writes_per_second=$writes_per_second \
--seed=$( date +%s ) \
2>&1 | tee -a $output_dir/${out_name}"
echo $cmd | tee $output_dir/${out_name}
eval $cmd
summarize_result $output_dir/${out_name} readwhile${operation}.t${num_threads} readwhile${operation} $num_threads
}
function now() {
echo `date +"%s"`
}
report="$output_dir/report.txt"
schedule="$output_dir/schedule.txt"
echo "===== Benchmark ====="
# Run!!!
IFS=',' read -a jobs <<< $1
for job in ${jobs[@]}; do
if [ $job != debug ]; then
echo "Start $job at `date`" | tee -a $schedule
fi
start=$(now)
if [ $job = fillseq ]; then
run_fillseq
elif [ $job = overwrite ]; then
run_change overwrite
elif [ $job = readrandom ]; then
run_readrandom
elif [ $job = readwhilewriting ]; then
run_readwhile writing
elif [ $job = debug ]; then
num_keys=1000; # debug
echo "Setting num_keys to $num_keys"
else
echo "unknown job $job"
exit
fi
end=$(now)
if [ $job != debug ]; then
echo "Complete $job in $((end-start)) seconds" | tee -a $schedule
fi
echo -e "ops/sec\tmb/sec\tusec/op\tavg\tp50\tTest"
tail -1 $output_dir/report.txt
done

@ -0,0 +1,174 @@
#!/bin/bash
# REQUIRE: benchmark_leveldb.sh exists in the current directory
# After execution of this script, log files are generated in $output_dir.
# report.txt provides a high level statistics
#
# This should be used with the LevelDB fork listed here to use additional test options.
# For more details on the changes see the blog post listed below.
# https://github.com/mdcallag/leveldb-1
# http://smalldatum.blogspot.com/2015/04/comparing-leveldb-and-rocksdb-take-2.html
#
# This should be run from the parent of the tools directory. The command line is:
# [$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
#
# 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
# NWRITESPERSEC - the writes/second rate limit for the *whilewriting* tests.
# If this is too large then the non-writer threads can get
# starved.
# 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_leveldb.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
# 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
# This allows tests from steps 2, 3 to be repeated faster.
# SAVE_SETUP - saves a copy of the database at the end of step 1 to
# $DATA_DIR.bak.
# Size constants
K=1024
M=$((1024 * K))
G=$((1024 * M))
num_keys=${NKEYS:-$((1 * G))}
wps=${NWRITESPERSEC:-$((10 * K))}
vs=${VAL_SIZE:-400}
cs=${CACHE_BYTES:-$(( 1 * G ))}
bs=${BLOCK_LENGTH:-4096}
# 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/"}
do_setup=${DO_SETUP:-1}
save_setup=${SAVE_SETUP:-0}
output_dir="/tmp/output"
ARGS="\
OUTPUT_DIR=$output_dir \
NUM_KEYS=$num_keys \
DB_DIR=$db_dir \
VALUE_SIZE=$vs \
BLOCK_SIZE=$bs \
CACHE_SIZE=$cs"
mkdir -p $output_dir
echo -e "ops/sec\tmb/sec\tusec/op\tavg\tp50\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
###### Setup the database
if [[ $do_setup != 0 ]]; then
echo Doing setup
# 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_leveldb.sh fillseq
# Test 2b: sequential fill with the configured value size
env $ARGS ./tools/benchmark_leveldb.sh fillseq
# Test 3: single-threaded overwrite
env $ARGS NUM_THREADS=1 DB_BENCH_NO_SYNC=1 ./tools/benchmark_leveldb.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
fi
if [[ $save_setup != 0 ]]; then
echo Save database to ${db_dir}.bak
cp -p -r $db_dir ${db_dir}.bak
fi
###### Read-only tests
for num_thr in "${nthreads[@]}" ; do
# Test 4: random read
env $ARGS NUM_THREADS=$num_thr ./tools/benchmark_leveldb.sh readrandom
done
###### Non read-only tests
for num_thr in "${nthreads[@]}" ; do
# Test 7: overwrite with sync=0
env $ARGS NUM_THREADS=$num_thr DB_BENCH_NO_SYNC=1 \
./tools/benchmark_leveldb.sh overwrite
# Test 8: overwrite with sync=1
# Not run for now because LevelDB db_bench doesn't have an option to limit the
# test run to X seconds and doing sync-per-commit for --num can take too long.
# env $ARGS NUM_THREADS=$num_thr ./tools/benchmark_leveldb.sh overwrite
# Test 11: random read while writing
env $ARGS NUM_THREADS=$num_thr WRITES_PER_SECOND=$wps \
./tools/benchmark_leveldb.sh readwhilewriting
done
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
echo fillseq >> $output_dir/report2.txt
head -1 $output_dir/report.txt >> $output_dir/report2.txt
grep fillseq $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
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
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 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
cat $output_dir/report2.txt
Loading…
Cancel
Save