performance.md

Performance：

0. Env

Disk: 2T NVME

CPU: Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz * 40

Memory: 256G

Network Card: 10-Gigabit

OS： CentOS Linux release 7.4

1. Single Db

1.1 Write binlog with one slave

data size: 64bytes

key num: 1,000,000

TEST	QPS
set	124347
get	283849

1.2 No binlog No slave

1.3 Benchmark Result

	WithBinlog&Slave QPS	NoBinlog&Slave QPS
PING_INLINE	262329	272479
PING_BULK	262467	270562
SET	124953	211327
GET	284900	292568
INCR	120004	213766
MSET (10 keys)	64863	111578
MGET (10 keys)	224416	223513
MGET (100 keys)	29935	29550
MGET (200 keys)	15128	14912
LPUSH	117799	205380
RPUSH	117481	205212
LPOP	112120	200320
RPOP	119932	207986
LRANGE_10 (first 10 elements)	277932	284414
LRANGE_100 (first 100 elements)	165118	164355
LRANGE_300 (first 300 elements)	54907	55096
LRANGE_450 (first 450 elements)	36656	36630
LRANGE_600 (first 600 elements)	27540	27510
SADD	126230	208768
SPOP	103135	166555
HSET	122443	214362
HINCRBY	114757	208942
HINCRBYFLOAT	114377	208550
HGET	284900	290951
HMSET (10 fields)	58937	111445
HMGET (10 fields)	203624	205592
HGETALL	166861	160797
ZADD	106780	189178
ZREM	112866	201938
PFADD	4708	4692
PFCOUNT	27412	27345
PFMERGE	478	494

1.4 Compare Wiht Redis

With Redis AOF configuration appendfsync everysec, redis basically write data into memeory. However, pika uses rocksdb, which writes WAL on ssd druing every write batch. That comparation becomes multiple threads sequential write into ssd and one thread write into memory.

Put the fairness or unfairness aside. Here is the performance.

2. Cluster (Codis)

2.1 Topology:

WithBInlog&Slave

4 machine * 2 pika instance (Master)

4 machine * 2 pika instance (Slave)

NoBinlog&Slave

4 machine * 2 pika instance (Master)

Slots Distribution:

2.2 WithBinlog&Slave

Command	QPS
Set	1,400,000+

2.3 NoBinlog&Slave

Command	QPS
Set	1,600,000+

2.4 Get Command

Command	QPS
Get	2,300,000+

With or without binlog, for Get command, QPS is approximately the same.