We are running performance tests for vector tiles serving and rasterization. In these tests, tiles are served without any caching to obtain the raw server performance. Every VM runs a single MapTiler Server instance with default rasterization configuration. The load was generated from 50 virtual users.
Performance tests results
|
OS |
Azure VM type |
RAM |
CPUs |
Vector tiles |
Vector tiles |
Raster tiles hits/s |
Raster tiles AVG response time (ms) |
|---|---|---|---|---|---|---|---|
|
RHEL 8.2 |
Standard A2 v2 |
4 |
2 |
190 |
206 |
13 |
2910 |
|
Ubuntu 20.04 |
Standard D4s v3 |
16 |
4 |
490 |
76 |
34 |
1099 |
|
Ubuntu 20.04 |
Standard D2as v4 |
8 |
2 |
668 |
56 |
37 |
1031 |
Performance tips
-
configure caching in production is basically a must-have, see how to configure it in Nginx
-
more CPUs and RAM will be used mainly during map rasterization, if you serve vector/raster tiles, multiple CPUs will not help you much
-
rather invest into fast SSD drives
-
rasterization use internal caching massively, if you use it for the whole planet (lots of fonts etc) 16 RAM is minimum
-
rasterization performance can be tuned in server settings/rasterization. Setting renderer pool size higher than CPU count can slightly increase throughput but beware that every renderer allocates its own memory space.