The repository presents a set of custom benchmarks for tokio asynchronous runtime. А lower value is better.
In this section, the characteristics of the subsequent benchmarks will be given first, then one by one the benchmarks.
Each subsequent benchmark is an estimate of the average execution time of a fixed number of tasks using multi_threaded tokio runtime.
The exact current parameters can be found here.
This benchmark is aimed at evaluating the scenario of generating a large number of tasks from the small number of producers tasks. And it is implemented in two versions.
N_SPAWN tasks are divided approximately equally between nsplit actively spawning tasks.
That is, let's say for N_SPAWN = 100, where N_SPAWN is the number of spawning tasks, the partition with a factor of 4 will look like [25; 4].
You can find out more about the task split method here.
In this case, the N_SPAWN tasks are divided so that each subsequent producer gets nsplit times less. That is, splitting 1000 elements with nsplit == 3 will be as follows: [667, 222, 74, 25, 8, 3, 1].
You can find out more about the task split method here.
The source code can be found here
This benchmark represents a scenario for the production of a large number of tasks from a single producer's task in two varants.
Spawn from block_on async lambda |
Spawn from local async task |
|---|---|
 |
 |
The source code can be found here
As an experiment, the order of selecting the next task for the worker was changed in the tokio source code. Then, the workload benchmark was as follows.
| Default | Reversed |
|---|---|
|
 |
| Default | Reversed |
|---|---|
|
 |
The work was performed under an MIT license, the text of which can be found here.
Feel free to suggest any changes and criticize the current implementation!