Why is grpc via dapr slower than http? #4866
Description
Note: If you have a general support question and are looking for a quicker response, please checkout our discord channel for answers from the community:
https://aka.ms/dapr-discord
In what area(s)?
/area runtime
/area operator
/area placement
/area docs
/area test-and-release
Ask your question here
A simplistic performance comparison of invoking a method via gRPC and http with and without dapr shows that without dapr the grpc calls are more efficient by factor 2 but via dapr less efficient by 10%. Also interesting is that grpc performance is nearly 5 times slower via dapr compared to direct calls. Is this expected?
The tests are in the makefile in this repo: https://github.com/wostzone/echo-dapr and the readme shows the numbers. Repeated here for convenience:
| Makefile | test | duration |
|---|---|---|
| run1 | 1000 plain grpc calls | 164 msec |
| run2 | 1000 plain http calls | 320 msec |
| run3 | 1000 dapr wrapped grpc calls | 986 msec |
| run4 | 1000 dapr wrapped http calls | 824 msec |
| run5 | 1000 grpc sdk calls | 772 msec |
| run6 | 1000 http sdk calls | 700 msec |
Please note that this is a simplistic performance test without additional middleware, just simple invoke.