RocksDB In Memory Workload Performance Benchmarks

The goal of these benchmarks is to measure the performance of RocksDB when the data resides in RAM. The system is configured to store transaction logs on persistent storage so that data is not lost on machine reboots.

Setup

All of the benchmarks are run on the same machine. Here are common setup used for following benchmark tests. Some configs vary based on tests and will be mentioned in the corresponding subsection.

• Intel Xeon E5-2660 @ 2.2GHz, 32 cores
• 20MB CPU cache, 144GB Ram
• CentOS release 6.3 (Kernel 3.2.51)
• Commit (c90d446ee7b87682b1e0ec7e0c778d25a90c6294) from master branch
• 2 max writer buffers of 128MB each
• Level style compaction
• PlainTable SST format, with BloomFilter enabled
• HashSkipList memtable format, with BloomFilter disabled
• WAL is enabled and kept on spinning disk. However, they are not archived and no DB backup is performed during the benchmark
• Each key is of size 20 bytes, each value is of size 100 bytes
• No compression
• Database is initially loaded with 500M unique keys by using db_bench's filluniquerandom mode, then read performance is measured for a 7200-second run in readwhilewriting mode with 32 reader threads. Each reader thread issues random key request, which is guaranteed to be found. Another dedicated writer thread issues write requests in the meantime.
• Performance are measured for light write rate (10K writes/sec => ~1.2MB/sec) and moderate write rate (80K writes/sec => ~9.6 MB/sec) respectively.
• Total database size is ~73GB at steady state, data files stored in tmpfs
• Statistics is disabled during the benchmark and perf_context level is set to 0
• jemalloc memory allocator

Test 1. Point Lookup

In this test, whole key is indexed for fast point lookup (i.e. prefix_extractor = NewFixedPrefixTransform(20)). When using this setup, range query is not supported. It provides maximum throughput for point lookup queries.

Below is the reported read performance when write rate is set to 80K writes/sec:

readwhilewriting :       0.220 micros/op 4553529 ops/sec; (1026765999 of 1026765999 found)

The actual sustained write rate is found to be at ~52K writes/sec

Here is the command for filling up the DB:

./db_bench --db=/mnt/db/rocksdb --num_levels=6 --key_size=20 --prefix_size=20 --keys_per_prefix=0 --value_size=100 --block_size=4096 --cache_size=17179869184 --cache_numshardbits=6 --compression_type=none --compression_ratio=1 --min_level_to_compress=-1 --disable_seek_compaction=1 --hard_rate_limit=2 --write_buffer_size=134217728 --max_write_buffer_number=2 --level0_file_num_compaction_trigger=8 --target_file_size_base=134217728 --max_bytes_for_level_base=1073741824 --disable_wal=0 --wal_dir=/data/users/rocksdb/WAL_LOG --sync=0 --disable_data_sync=1 --verify_checksum=1 --delete_obsolete_files_period_micros=314572800 --max_grandparent_overlap_factor=10 --max_background_compactions=4 --max_background_flushes=0 --level0_slowdown_writes_trigger=16 --level0_stop_writes_trigger=24 --statistics=0 --stats_per_interval=0 --stats_interval=1048576 --histogram=0 --use_plain_table=1 --open_files=-1 --mmap_read=1 --mmap_write=0 --memtablerep=prefix_hash --bloom_bits=10 --bloom_locality=1 --perf_level=0 --benchmarks=filluniquerandom --use_existing_db=0 --num=524288000 --threads=1

Here is the command for running readwhilewriting benchmark:

./db_bench --db=/mnt/db/rocksdb --num_levels=6 --key_size=20 --prefix_size=20 --keys_per_prefix=0 --value_size=100 --block_size=4096 --cache_size=17179869184 --cache_numshardbits=6 --compression_type=none --compression_ratio=1 --min_level_to_compress=-1 --disable_seek_compaction=1 --hard_rate_limit=2 --write_buffer_size=134217728 --max_write_buffer_number=2 --level0_file_num_compaction_trigger=8 --target_file_size_base=134217728 --max_bytes_for_level_base=1073741824 --disable_wal=0 --wal_dir=/data/users/rocksdb/WAL_LOG --sync=0 --disable_data_sync=1 --verify_checksum=1 --delete_obsolete_files_period_micros=314572800 --max_grandparent_overlap_factor=10 --max_background_compactions=4 --max_background_flushes=0 --level0_slowdown_writes_trigger=16 --level0_stop_writes_trigger=24 --statistics=0 --stats_per_interval=0 --stats_interval=1048576 --histogram=0 --use_plain_table=1 --open_files=-1 --mmap_read=1 --mmap_write=0 --memtablerep=prefix_hash --bloom_bits=10 --bloom_locality=1 --perf_level=0 --duration=7200 --benchmarks=readwhilewriting --use_existing_db=1 --num=524288000 --threads=32 --writes_per_second=81920

Test 2. Prefix Range Query

In this test, a key prefix size of 12 bytes is indexed (i.e. prefix_extractor = NewFixedPrefixTransform(12)). db_bench is configured to generate approximately 10 keys per unique prefix. In this setting, Seek() can be performed within a given prefix as well as iteration on the returned iterator. However, behavior of scanning beyond the prefix boundary is not defined. This is a fairly common access pattern for graph data. Point lookup throughput is measured under this setting:

Below is the reported read performance when write rate is set to 80K writes/sec:

readwhilewriting :       0.251 micros/op 3979207 ops/sec; (893448999 of 893448999 found)

The actual sustained write rate is found to be at ~67K writes/sec

Here is the command for filling up the DB:

./db_bench --db=/mnt/db/rocksdb --num_levels=6 --key_size=20 --prefix_size=12 --keys_per_prefix=10 --value_size=100 --block_size=4096 --cache_size=17179869184 --cache_numshardbits=6 --compression_type=none --compression_ratio=1 --min_level_to_compress=-1 --disable_seek_compaction=1 --hard_rate_limit=2 --write_buffer_size=134217728 --max_write_buffer_number=2 --level0_file_num_compaction_trigger=8 --target_file_size_base=134217728 --max_bytes_for_level_base=1073741824 --disable_wal=0 --wal_dir=/data/users/rocksdb/WAL_LOG --sync=0 --disable_data_sync=1 --verify_checksum=1 --delete_obsolete_files_period_micros=314572800 --max_grandparent_overlap_factor=10 --max_background_compactions=4 --max_background_flushes=0 --level0_slowdown_writes_trigger=16 --level0_stop_writes_trigger=24 --statistics=0 --stats_per_interval=0 --stats_interval=1048576 --histogram=0 --use_plain_table=1 --open_files=-1 --mmap_read=1 --mmap_write=0 --memtablerep=prefix_hash --bloom_bits=10 --bloom_locality=1 --perf_level=0 --benchmarks=filluniquerandom --use_existing_db=0 --num=524288000 --threads=1

Here is the command for running readwhilewriting benchmark:

./db_bench --db=/mnt/db/rocksdb --num_levels=6 --key_size=20 --prefix_size=12 --keys_per_prefix=10 --value_size=100 --block_size=4096 --cache_size=17179869184 --cache_numshardbits=6 --compression_type=none --compression_ratio=1 --min_level_to_compress=-1 --disable_seek_compaction=1 --hard_rate_limit=2 --write_buffer_size=134217728 --max_write_buffer_number=2 --level0_file_num_compaction_trigger=8 --target_file_size_base=134217728 --max_bytes_for_level_base=1073741824 --disable_wal=0 --wal_dir=/data/users/rocksdb/WAL_LOG --sync=0 --disable_data_sync=1 --verify_checksum=1 --delete_obsolete_files_period_micros=314572800 --max_grandparent_overlap_factor=10 --max_background_compactions=4 --max_background_flushes=0 --level0_slowdown_writes_trigger=16 --level0_stop_writes_trigger=24 --statistics=0 --stats_per_interval=0 --stats_interval=1048576 --histogram=0 --use_plain_table=1 --open_files=-1 --mmap_read=1 --mmap_write=0 --memtablerep=prefix_hash --bloom_bits=10 --bloom_locality=1 --perf_level=0 --duration=7200 --benchmarks=readwhilewriting --use_existing_db=1 --num=524288000 --threads=32 --writes_per_second=81920