Performance whitepaper of Redis 6.0-compatible Tair DRAM-based instances - Tair

This topic describes the method, environment, and tool that are used to test the performance of a Redis 6.0-compatible Tair DRAM-based instance and provides the test results.

Overview

Tair DRAM-based instances are suitable for scenarios that involve high concurrency and a large number of read and write operations. DRAM-based instances provide more benefits than ApsaraDB for Redis Community Edition instances.

DRAM-based instances use the multi-threading model. A DRAM-based instance provides the read and write performance approximately three times that of an ApsaraDB for Redis Community Edition instance with the same specifications.
DRAM-based instances provide multiple enhanced data modules, including TairString (including CAS and CAD commands), TairHash, TairGIS, TairBloom, TairDoc, TairTS, TairCpc, TairZset, TairRoaring, TairVector, and TairSearch. These modules eliminate your worries about data structures and timeliness for storage and allow you to focus on application development.

For more information, see DRAM-based instances.

Test environment

Database


Environment-related item	Description
Region and zone	Zone A in the China (Zhangjiakou) region
Storage type	A DRAM-based instance that is compatible with Redis 6.0
Engine version	6.2.1.1
Instance architecture	The standard master-replica architecture that does not have the cluster mode enabled. For more information, see Standard architecture.
Instance type	The test results are less susceptible to instance types. In this example, the tair.rdb.4g instance type is used. For more information, see DRAM-based instances.

Client


Environment-related item	Description
Host of the client	An Elastic Compute Service (ECS) instance of the ecs.g5ne.16xlarge instance type is used. For more information, see Overview of instance families.
Region and zone	Zone A in the China (Zhangjiakou) region
Operating system	CentOS 7.2 64-bit
Network	The virtual private cloud (VPC) where the Tair and ECS instances reside

Test tool

The redis-benchmark tool is used. It is the official test tool of open source Redis that can be used to benchmark Tair performance. In this example, redis-benchmark 7.0.6 is used.

Note To ensure that the --threads parameter is available, we recommend that you use the latest version of redis-benchmark. For more information, visit GitHub.

Method

Connect to the ECS instance, download and compile the Redis source code, and then generate a redis-benchmark file to start your test.

To test the SET command performance, you can run the following command that involves 16 threads, 256 connections, 3,000,000 keys, and a maximum value length of 64 bytes:

./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 10000000  -c 256 -t set -d 64 --threads 16

Table 1. Parameters
Parameter	Description
-h	The internal endpoint of the Tair instance. For more information, see View endpoints and port numbers.
-p	The service port number that is used to connect to the Tair instance. Default value: 6379.
-a	The password of the account that is used to connect to the Tair instance. Note If you are using an account that is created in the Tair console to connect to the instance, the password must be in the `<user>:<password>` format. For example, if the username of the account is admin and the password is Rp829dlwa, use `admin:Rp829dlwa` as the password to connect to the Tair instance.
-c	The number of concurrent connections.
-n	The total number of requests tested. Set this parameter to a large value to ensure continuous stress testing.
-t	The command used in testing, such as GET or SET.
-d	The value that you process by using a SET or GET command. Unit: bytes.
-r	The random range of the used keys. It specifies the number of keys to be used.
--threads	Runs a multi-threaded stress test and specifies the number of threads.

Note For more information about the parameters, see Redis benchmark.

Test results

This section provides the test results of basic ApsaraDB for Redis commands like SET and GET in different scenarios.

Test metrics


Test metric	Description
QPS	The number of read and write operations that are processed per second.
Latency	The average latency for a specified proportion of operations. Unit: milliseconds.
99th Percentile Latency	The maximum latency for the fastest 99% of all requests. Unit milliseconds. For example, a value of 0.5 indicates that 99% of requests can be processed within 0.5 milliseconds.

Strings

GET

Sample command that involves 16 threads, 256 connections, 3,000,000 keys, and a maximum value length of 16 bytes:

./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -d 16 -t get

Test results:


Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
16	523907.66	0.476	0.719
128	519363.59	0.476	0.871
256	517027.44	0.480	0.839
1024	487709.72	0.509	0.879
4096	385728.06	0.629	1.319

Note

Read requests are not affected by master-replica synchronization or data persistence. As a result, the CPU utilization is high.
In this test, append-only file (AOF) persistence is enabled for the Tair instance. If you disable this feature, the queries per second (QPS) can be improved by about 10%.

SET

Sample command that involves 16 threads, 256 connections, 3,000,000 keys, and a maximum value length of 16 bytes:

./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -d 16 -t set

Test results:


Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
16	260642.91	0.956	1.871
128	244758.09	1.035	2.015
256	235128.14	1.072	2.119
1024	173761.95	1.462	2.783
4096	66198.86	3.857	66.687

Note In this test, a standard Tair instance is used. The default maximum bandwidth is 96 MB/s. The throttling mechanism provided by Tair is triggered when the upper bandwidth limit is reached at a value length of 480 bytes and QPS of up to 200,000. QPS decrease in inverse proportion to the value length. If you have higher requirements for bandwidth, we recommend that you choose cluster or read/write splitting instances.

MSET
Sample command that involves 16 threads, 256 connections, 3,000,000 keys, and a maximum value length of 16 bytes:
```
./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -d 16 -t mset
```
Test results:
Value length (bytes) QPS Average latency (milliseconds) 99th percentile latency (milliseconds)
16 66533.60 3.637 9.735
128 66533.60 3.756 12.679
256 70439.07 3.610 14.487
1024 38634.90 6.497 97.407
INCR
Sample command that involves 16 threads, 256 connections, and 3,000,000 keys:
```
./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -t incr
```
Test results:
Value length (bytes) QPS Average latency (milliseconds) 99th percentile latency (milliseconds)
None 244758.09 1.024 2.879

Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
16	66533.60	3.637	9.735
128	66533.60	3.756	12.679
256	70439.07	3.610	14.487
1024	38634.90	6.497	97.407

Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
None	244758.09	1.024	2.879

Lists

LPUSH

Sample command that involves 16 threads, 256 connections, 3,000,000 keys, and a maximum value length of 16 bytes:

./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -d 16 -t lpush

Test results:


Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
16	133220.84	1.911	3.943
128	137797.98	1.846	3.831
256	130332.78	1.956	4.087
1024	116400.88	2.183	4.439
4096	68101.34	3.732	8.727

LPOP
Sample command that involves 16 threads, 256 connections, 3,000,000 keys, and a maximum value length of 16 bytes:
```
./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -d 16 -t lpop
```
Test results:
Value length (bytes) QPS Average latency (milliseconds) 99th percentile latency (milliseconds)
16 122354.09 2.084 4.343
128 122324.16 2.085 4.359
256 113079.53 2.239 4.719
1024 74447.23 3.419 7.007
4096 74449.07 3.410 7.071
RPUSH
Sample command that involves 16 threads, 256 connections, 3,000,000 keys, and a maximum value length of 16 bytes:
```
./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -d 16 -t rpush
```
Test results:
Value length (bytes) QPS Average latency (milliseconds) 99th percentile latency (milliseconds)
16 119875.34 2.129 4.583
128 116391.85 2.187 4.631
256 105148.78 2.409 5.095
1024 82674.24 3.073 6.023
4096 69683.18 3.656 6.271
RPOP
Sample command that involves 16 threads, 256 connections, 3,000,000 keys, and a maximum value length of 16 bytes:
```
./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -d 16 -t rpop
```
Test results:
Value length (bytes) QPS Average latency (milliseconds) 99th percentile latency (milliseconds)
16 110996.01 2.281 4.919
128 101584.73 2.500 5.287
256 92205.56 2.764 5.783
1024 76324.23 3.339 6.759
4096 79377.68 3.207 6.267

Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
16	122354.09	2.084	4.343
128	122324.16	2.085	4.359
256	113079.53	2.239	4.719
1024	74447.23	3.419	7.007
4096	74449.07	3.410	7.071

Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
16	119875.34	2.129	4.583
128	116391.85	2.187	4.631
256	105148.78	2.409	5.095
1024	82674.24	3.073	6.023
4096	69683.18	3.656	6.271

Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
16	110996.01	2.281	4.919
128	101584.73	2.500	5.287
256	92205.56	2.764	5.783
1024	76324.23	3.339	6.759
4096	79377.68	3.207	6.267

Sets

SADD
Sample command that involves 16 threads, 256 connections, and 3,000,000 keys:
```
./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -t sadd
```
Test results:
Value length (bytes) QPS Average latency (milliseconds) 99th percentile latency (milliseconds)
None 111020.66 2.286 4.927
SPOP
Sample command that involves 16 threads, 256 connections, and 3,000,000 keys:
```
./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -t spop
```
Test results:
Value length (bytes) QPS Average latency (milliseconds) 99th percentile latency (milliseconds)
None 67315.89 3.784 7.847

Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
None	111020.66	2.286	4.927

Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
None	67315.89	3.784	7.847

Sorted sets

ZADD
Sample command that involves 16 threads, 256 connections, and 3,000,000 keys:
```
./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -t zadd
```
Test results:
Value length (bytes) QPS Average latency (milliseconds) 99th percentile latency (milliseconds)
None 53490.24 4.761 9.799
ZPOPMIN
Sample command that involves 16 threads, 256 connections, and 3,000,000 keys:
```
./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -t zpopmin
```
Test results:
Value length (bytes) QPS Average latency (milliseconds) 99th percentile latency (milliseconds)
None 114163.94 2.227 4.871

Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
None	53490.24	4.761	9.799

Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
None	114163.94	2.227	4.871

Hashes

HSET

Sample command that involves 16 threads, 256 connections, 3,000,000 keys, and a maximum value length of 16 bytes:

./src/redis-benchmark -h r-bp1s02ae14mr****.redis.rds.aliyuncs.com -p 6379 -a testaccount:Rp829dlwa -n 3000000 -r 30000000 -c 256 --threads 16 -d 16 -t hset

Test results:


Value length (bytes)	QPS	Average latency (milliseconds)	99th percentile latency (milliseconds)
16	108964.12	2.327	5.103
128	111012.43	2.293	5.039
256	108960.16	2.336	5.103
1024	103334.25	2.466	5.143
4096	61469.11	4.147	7.263

Summary

Tair DRAM-based instances use an in-house multi-thread model that delivers higher QPS at lower latency. In this context, Redis 6.0-compatible Tair DRAM-based instances demonstrate good performance in various scenarios.