LevelDB Benchmarks

Google, July 2011

---

In order to test LevelDB's performance, we benchmark it against other well-established database implementations. We compare LevelDB (revision 39) against SQLite3 (version 3.7.6.3) and Kyoto Cabinet's (version 1.2.67) TreeDB (a B+Tree based key-value store). We would like to acknowledge Scott Hess and Mikio Hirabayashi for their suggestions and contributions to the SQLite3 and Kyoto Cabinet benchmarks, respectively.

Benchmarks were all performed on a six-core Intel(R) Xeon(R) CPU X5650 @ 2.67GHz, with 12288 KB of total L3 cache and 12 GB of DDR3 RAM at 1333 MHz. (Note that LevelDB uses at most two CPUs since the benchmarks are single threaded: one to run the benchmark, and one for background compactions.) We ran the benchmarks on two machines (with identical processors), one with an Ext3 file system and one with an Ext4 file system. The machine with the Ext3 file system has a SATA Hitachi HDS721050CLA362 hard drive. The machine with the Ext4 file system has a SATA Samsung HD502HJ hard drive. Both hard drives spin at 7200 RPM and have hard drive write-caching enabled (using `hdparm -W 1 [device]`). The numbers reported below are the median of three measurements.

Benchmark Source Code

We wrote benchmark tools for SQLite and Kyoto TreeDB based on LevelDB's db_bench. The code for each of the benchmarks resides here:

• LevelDB: db/db_bench.cc.
• SQLite: doc/bench/db_bench_sqlite3.cc.
• Kyoto TreeDB: doc/bench/db_bench_tree_db.cc.

Custom Build Specifications

• LevelDB: LevelDB was compiled with the tcmalloc library and the Snappy compression library (revision 33). Assertions were disabled.
• TreeDB: TreeDB was compiled using the LZO compression library (version 2.03). Furthermore, we enabled the TSMALL and TLINEAR options when opening the database in order to reduce the footprint of each record.
• SQLite: We tuned SQLite's performance, by setting its locking mode to exclusive. We also enabled SQLite's write-ahead logging.

1. Baseline Performance

This section gives the baseline performance of all thedatabases. Following sections show how performance changes as variousparameters are varied. For the baseline:

• Each database is allowed 4 MB of cache memory.
• Databases are opened in asynchronous write mode. (LevelDB's sync option, TreeDB's OAUTOSYNC option, and SQLite3's synchronous options are all turned off). I.e., every write is pushed to the operating system, but the benchmark does not wait for the write to reach the disk.
• Keys are 16 bytes each.
• Value are 100 bytes each (with enough redundancy so that a simple compressor shrinks them to 50% of their original size).
• Sequential reads/writes traverse the key space in increasing order.
• Random reads/writes traverse the key space in random order.

A. Sequential Reads

LevelDB 4,030,000 ops/sec  Kyoto TreeDB 1,010,000 ops/sec  SQLite3 383,000 ops/sec  

B. Random Reads

LevelDB 129,000 ops/sec  Kyoto TreeDB 151,000 ops/sec  SQLite3 134,000 ops/sec  

C. Sequential Writes

LevelDB 779,000 ops/sec  Kyoto TreeDB 342,000 ops/sec  SQLite3 48,600 ops/sec  

D. Random Writes

LevelDB 164,000 ops/sec  Kyoto TreeDB 88,500 ops/sec  SQLite3 9,860 ops/sec  

LevelDB outperforms both SQLite3 and TreeDB in sequential and random write operations and sequential read operations. Kyoto Cabinet has the fastest random read operations.

2. Write Performance under Different Configurations

A. Large Values

For this benchmark, we start with an empty database, and write 100,000 byte values (~50% compressible). To keep the benchmark running time reasonable, we stop after writing 1000 values.

Sequential Writes

LevelDB 1,100 ops/sec  Kyoto TreeDB 1,000 ops/sec  SQLite3 1,600 ops/sec  

Random Writes

LevelDB 480 ops/sec  Kyoto TreeDB 1,100 ops/sec  SQLite3 1,600 ops/sec  

LevelDB doesn't perform as well with large values of 100,000 bytes each. This is because LevelDB writes keys and values at least twice: first time to the transaction log, and second time (during a compaction) to a sorted file.With larger values, LevelDB's per-operation efficiency is swamped by thecost of extra copies of large values.

B. Batch Writes

A batch write is a set of writes that are applied atomically to the underlying database. A single batch of N writes may be significantly faster than N individual writes. The following benchmark writes one thousand batches where each batch contains one thousand 100-byte values. TreeDB does not support batch writes and is omitted from this benchmark.

Sequential Writes

LevelDB 840,000 entries/sec   (1.08x baseline)SQLite3 124,000 entries/sec   (2.55x baseline)

Random Writes

LevelDB 221,000 entries/sec   (1.35x baseline)SQLite3 22,000 entries/sec   (2.23x baseline)

Because of the way LevelDB persistent storage is organized, batches ofrandom writes are not much slower (only a factor of 4x) than batchesof sequential writes.

C. Synchronous Writes

In the following benchmark, we enable the synchronous writing modesof all of the databases. Since this change significantly slows down thebenchmark, we stop after 10,000 writes. For synchronous write tests, we'vedisabled hard drive write-caching (using `hdparm -W 0 [device]`).

• For LevelDB, we set WriteOptions.sync = true.
• In TreeDB, we enabled TreeDB's OAUTOSYNC option.
• For SQLite3, we set "PRAGMA synchronous = FULL".

Sequential Writes

LevelDB 100 ops/sec   (0.003x baseline)Kyoto TreeDB 7 ops/sec   (0.0004x baseline)SQLite3 88 ops/sec   (0.002x baseline)

Random Writes

LevelDB 100 ops/sec   (0.015x baseline)Kyoto TreeDB 8 ops/sec   (0.001x baseline)SQLite3 88 ops/sec   (0.009x baseline)

Also see the ext4 performance numbers belowsince synchronous writes behave significantly differentlyon ext3 and ext4.

D. Turning Compression Off

In the baseline measurements, LevelDB and TreeDB were usinglight-weight compression(Snappy for LevelDB,and LZO forTreeDB). SQLite3, by default does not use compression. Theexperiments below show what happens when compression is disabled inall of the databases (the SQLite3 numbers are just a copy ofits baseline measurements):

Sequential Writes

LevelDB 594,000 ops/sec   (0.76x baseline)Kyoto TreeDB 485,000 ops/sec   (1.42x baseline)SQLite3 48,600 ops/sec   (1.00x baseline)

Random Writes

LevelDB 135,000 ops/sec   (0.82x baseline)Kyoto TreeDB 159,000 ops/sec   (1.80x baseline)SQLite3 9,860 ops/sec   (1.00x baseline)

LevelDB's write performance is better with compression than withoutsince compression decreases the amount of data that has to be writtento disk. Therefore LevelDB users can leave compression enabled inmost scenarios without having worry about a tradeoff between spaceusage and performance. TreeDB's performance on the other hand isbetter without compression than with compression. Presumably this isbecause TreeDB's compression library (LZO) is more expensive thanLevelDB's compression library (Snappy).

E. Using More Memory

We increased the overall cache size for each database to 128 MB. For LevelDB, we partitioned 128 MB into a 120 MB write buffer and 8 MB of cache (up from 2 MB of write buffer and 2 MB of cache). For SQLite3, we kept the page size at 1024 bytes, but increased the number of pages to 131,072 (up from 4096). For TreeDB, we also kept the page size at 1024 bytes, but increased the cache size to 128 MB (up from 4 MB).

Sequential Writes

LevelDB 812,000 ops/sec   (1.04x baseline)Kyoto TreeDB 321,000 ops/sec   (0.94x baseline)SQLite3 48,500 ops/sec   (1.00x baseline)

Random Writes

LevelDB 355,000 ops/sec   (2.16x baseline)Kyoto TreeDB 284,000 ops/sec   (3.21x baseline)SQLite3 9,670 ops/sec   (0.98x baseline)

SQLite's performance does not change substantially when compared tothe baseline, but the random write performance for both LevelDB andTreeDB increases significantly. LevelDB's performance improvesbecause a larger write buffer reduces the need to merge sorted files(since it creates a smaller number of larger sorted files). TreeDB'sperformance goes up because the entire database is available in memoryfor fast in-place updates.

3. Read Performance under Different Configurations

A. Larger Caches

We increased the overall memory usage to 128 MB for each database.For LevelDB, we allocated 8 MB to LevelDB's write buffer and 120 MBto LevelDB's cache. The other databases don't differentiate between awrite buffer and a cache, so we simply set their cache size to 128MB.

Sequential Reads

LevelDB 5,210,000 ops/sec   (1.29x baseline)Kyoto TreeDB 1,070,000 ops/sec   (1.06x baseline)SQLite3 609,000 ops/sec   (1.59x baseline)

Random Reads

LevelDB 190,000 ops/sec   (1.47x baseline)Kyoto TreeDB 463,000 ops/sec   (3.07x baseline)SQLite3 186,000 ops/sec   (1.39x baseline)

As expected, the read performance of all of the databases increaseswhen the caches are enlarged. In particular, TreeDB seems to makevery effective use of a cache that is large enough to hold the entiredatabase.

B. No Compression Reads

For this benchmark, we populated a database with 1 million entries consisting of 16 byte keys and 100 byte values. We compiled LevelDB and Kyoto Cabinet without compression support, so results that are read out from the database are already uncompressed. We've listed the SQLite3 baseline read performance as a point of comparison.

Sequential Reads

LevelDB 4,880,000 ops/sec   (1.21x baseline)Kyoto TreeDB 1,230,000 ops/sec   (3.60x baseline)SQLite3 383,000 ops/sec   (1.00x baseline)

Random Reads

LevelDB 149,000 ops/sec   (1.16x baseline)Kyoto TreeDB 175,000 ops/sec   (1.16x baseline)SQLite3 134,000 ops/sec   (1.00x baseline)

Performance of both LevelDB and TreeDB improves a small amount whencompression is disabled. Note however that under different workloads,performance may very well be better with compression if it allows moreof the working set to fit in memory.

Note about Ext4 Filesystems

The preceding numbers are for an ext3 file system. Synchronous writes are much slower under ext4 (LevelDB drops to ~31 writes / second and TreeDB drops to ~5 writes / second; SQLite3's synchronous writes do not noticeably drop) due to ext4's different handling of fsync / msync calls. Even LevelDB's asynchronous write performance drops somewhat since it spreads its storage across multiple files and issues fsync calls when switching to a new file.

Acknowledgements

Jeff Dean and Sanjay Ghemawat wrote LevelDB. Kevin Tseng wrote and compiled these benchmarks. Mikio Hirabayashi, Scott Hess, and Gabor Cselle provided help and advice.