Introduction

These benchmarks are run on hardware provided by OSU OSL. The hardware configuration is:

• 24 Intel(R) Xeon(R) CPU X5650 @ 2.67GHz (2 physical CPUs)
• 144 GB RAM
• 400 GB Intel s3700 SSD
• 6+ disk SAS array using RAID-0 and xfs

Basic configurations:

The following tests are run using 500 million 200 byte records. All numbers are as reported by the benchmark - the throughput numbers do not necessarily reflect disk I/O.

The benchmarks run are based on the run_all script here: https://github.com/wiredtiger/rocksdb/blob/afb742f729fcffe41beb4b64862b215608555b29/run_all.sh and the subscripts it calls from rocksdb/runners/ at revision afb742f729fcffe41beb4b64862b215608555b29.

For my records: WiredTiger SSD is taken from wt_1386646429_500 RocksDB HDD is taken from rocks_1386718275_500 WiredTiger HDD is taken from wt_1386657211_500

Fill Random

Measurement	WiredTiger SSD	RocksDB HDD	WiredTiger HDD
Ops/sec	394474	807852	401782
micros/op	2.535	1.24	2.49
Throughput (MB/s)	81	166	82
99% Latency	6.31	1.98	6.26
99.99% Latency	39.5	7.3	41
Max Latency	10013	342197	10211
DB size (GB)	88	67	86

NOTES:

• I'm not entirely sure how the configuration options for RocksDB work, but I suspect RocksDB is allowed an additional 7.6GB of cache in this test via: `--write_buffer_size=268435456 --max_write_buffer_number=30". WiredTiger is configured with a single 1GB cache for the workload.

Fill Sequential

Measurement	WiredTiger SSD	RocksDB HDD	WiredTiger HDD
Ops/sec	691312	725649	643382
micros/op	1.45	1.38	1.55
Throughput (MB/s)	142	149	132
99% Latency	4.67	2.28	4.68
99.99% Latency	45.61	8.59	43.35
Max Latency	10014	306481	10538
DB size (GB)	73	106	74

WiredTiger and RocksDB deliver similar performance in this test, even though RocksDB is configured to load directly into a single level of the LSM tree (thus avoiding the overhead of any background merge operations).

Overwrite

Measurement	WiredTiger SSD	RocksDB HDD	WiredTiger HDD
Ops/sec	377543	135930	385427
micros/op	2.649	7.36	2.59
Throughput (MB/s)	77.8	28.0	79.4
99% Latency	6.52	11.39	6.44
99.99% Latency	41.04	1192	40
Max Latency	97248	341608	134240
DB size (GB)	147	117	147

NOTES:

• WiredTiger is 2.8x faster than RocksDB for this test. This is the closest to a true insert test of the benchmarks on this page, since overwrite in an LSM tree (at least as implemented here) is simply an insert operation.

Read Random

Read random is run with 32 threads. Database size is large because no compression is used.

WiredTiger executed 20 million operations per thread RocksDB executed 1 million operations per thread (because it takes so long)

The above difference puts WiredTiger at a distinct disadvantage when it comes to tail latency measurements in this test.

Measurement	WiredTiger SSD	RocksDB HDD	WiredTiger HDD
Ops/sec	273508	3747	173435
micros/op	3.66	266.8	5.77
99% Latency	1120	78658	1167
99.99% Latency	41444	208705	67045
Max Latency	6415292	1092593	7968930
DB size (GB)	207	107	209

Realistic read random

Database setup:

• No compression
• 180 byte items
• 16 bit bloom filter keys
• 16k leaf page size, 2 LSM merge threads for WiredTiger
• 2GB cache
• Sequential insert of 400 million. Allow LSM tree to merge.
• Overwrite 50 million entries
• Insert 8k items into the in-memory chunk
• Read items using 32 threads
• Reading 20 million items each in WiredTiger
• Reading 10 million items each in RocksDB

Measurement	WiredTiger HDD	RocksDB SSD
Ops/sec	135171	59783
micros/op	7.398	16.727
99% Latency	2904	7779
99.99% Latency	64157	24937
Max Latency	1297245	227393
DB size (GB)	90	86

Similar, except:

• 3 GB cache (to account for larger bloom filter size)
• 20 threads - to avoid process switching
• 20 million reads per thread (WiredTiger and RocksDB)
• Allow WiredTiger to use mmap
• Run on SSD

Measurement	WiredTiger SSD	RocksDB SSD
Ops/sec	253481	443969
micros/op	3.95	2.25
99% Latency	303	78
99.99% Latency	10667	417
Max Latency	2787967	363031
DB size (GB)	87	86