TiDB Sysbench Performance Test Report -- v2.1 vs. v2.0
Test purpose
This test aims to compare the performance of TiDB 2.1 and TiDB 2.0 for OLTP where the working set fits in memory.
Test version, time, and place
TiDB version: v2.1.0-rc.2 vs. v2.0.6
Time: September, 2018
Place: Beijing, China
Test environment
IDC machine:
Type | Name |
---|---|
OS | Linux (CentOS 7.3.1611) |
CPU | 40 vCPUs, Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz |
RAM | 128GB |
DISK | Optane 500GB SSD * 1 |
Sysbench version: 1.1.0
Test plan
Use Sysbench to import 16 tables, with 10,000,000 rows in each table. With the HAProxy, requests are sent to the cluster at an incremental concurrent number. A single concurrent test lasts 5 minutes.
TiDB version information
v2.1.0-rc.2
Component | GitHash |
---|---|
TiDB | 08e56cd3bae166b2af3c2f52354fbc9818717f62 |
TiKV | 57e684016dafb17dc8a6837d30224be66cbc7246 |
PD | 6a7832d2d6e5b2923c79683183e63d030f954563 |
v2.0.6
Component | GitHash |
---|---|
TiDB | b13bc08462a584a085f377625a7bab0cc0351570 |
TiKV | 57c83dc4ebc93d38d77dc8f7d66db224760766cc |
PD | b64716707b7279a4ae822be767085ff17b5f3fea |
TiDB parameter configuration
The default TiDB configuration is used in both v2.1 and v2.0.
TiKV parameter configuration
The following TiKV configuration is used in both v2.1 and v2.0:
[readpool.storage]
normal-concurrency = 8
[server]
grpc-concurrency = 8
[raftstore]
sync-log = false
[rocksdb.defaultcf]
block-cache-size = "60GB"
[rocksdb.writecf]
block-cache-size = "20GB"
Cluster topology
Machine IP | Deployment instance |
---|---|
172.16.30.31 | 1*Sysbench 1*HAProxy |
172.16.30.32 | 1*TiDB 1*pd 1*TiKV |
172.16.30.33 | 1*TiDB 1*TiKV |
172.16.30.34 | 1*TiDB 1*TiKV |
Test result
Point Select
test
Version | Threads | QPS | 95% Latency (ms) |
---|---|---|---|
v2.1 | 64 | 111481.09 | 1.16 |
v2.1 | 128 | 145102.62 | 2.52 |
v2.1 | 256 | 161311.9 | 4.57 |
v2.1 | 512 | 184991.19 | 7.56 |
v2.1 | 1024 | 230282.74 | 10.84 |
v2.0 | 64 | 75285.87 | 1.93 |
v2.0 | 128 | 92141.79 | 3.68 |
v2.0 | 256 | 107464.93 | 6.67 |
v2.0 | 512 | 121350.61 | 11.65 |
v2.0 | 1024 | 150036.31 | 17.32 |
According to the statistics above, the Point Select
query performance of TiDB 2.1 has increased by 50% than that of TiDB 2.0.
Update Non-Index
test
Version | Threads | QPS | 95% Latency (ms) |
---|---|---|---|
v2.1 | 64 | 18946.09 | 5.77 |
v2.1 | 128 | 22022.82 | 12.08 |
v2.1 | 256 | 24679.68 | 25.74 |
v2.1 | 512 | 25107.1 | 51.94 |
v2.1 | 1024 | 27144.92 | 106.75 |
v2.0 | 64 | 16316.85 | 6.91 |
v2.0 | 128 | 20944.6 | 11.45 |
v2.0 | 256 | 24017.42 | 23.1 |
v2.0 | 512 | 25994.33 | 46.63 |
v2.0 | 1024 | 27917.52 | 92.42 |
According to the statistics above, the Update Non-Index
write performance of TiDB 2.1 and TiDB 2.0 is almost the same.
Update Index
test
Version | Threads | QPS | 95% Latency (ms) |
---|---|---|---|
v2.1 | 64 | 9934.49 | 12.08 |
v2.1 | 128 | 10505.95 | 25.28 |
v2.1 | 256 | 11007.7 | 55.82 |
v2.1 | 512 | 11198.81 | 106.75 |
v2.1 | 1024 | 11591.89 | 200.47 |
v2.0 | 64 | 9754.68 | 11.65 |
v2.0 | 128 | 10603.31 | 24.38 |
v2.0 | 256 | 11011.71 | 50.11 |
v2.0 | 512 | 11162.63 | 104.84 |
v2.0 | 1024 | 12067.63 | 179.94 |
According to the statistics above, the Update Index
write performance of TiDB 2.1 and TiDB 2.0 is almost the same.