Shumai is a multi-thread benchmarking framework that produces accurate and reproducible results.
Shumai was developed as part of the Alchemy project to fullfil its academic requirements of accurate and reproducible benchmarking. Shumai put reproducibility as the first priority by automatically collecting the system information, benchmark configurations, and benchmark results. All of this data will be stored together in a json file. The benchmark configurations are also stored in a toml file which should be kept under source control.
```toml
[[Foo]] name = "foo" threads = [1, 2, 3] time = 1 parameter = [1, 2] ```
```rust // bench_config.rs
pub struct Foo {
pub name: String,
pub threads: Vec
impl ShumaiBench for TestBench { type Result = usize; type Config = Foo;
fn load(&self) -> Option<Value> {
None
}
fn run(&self, context: BenchContext<Foo>) -> Self::Result {
// Barrier to ensure all threads start at the same time
context.wait_for_start();
// start benchmark
todo!()
}
fn cleanup(&self) -> Option<Value> {
None
}
}
fn main() { let config = Foo::load() .expect("Failed to parse config!"); let repeat = 3;
for c in config.iter() {
let benchmark = TestBench::default();
let result = shumai::run(&benchmark, c, repeat);
result.to_json() // save results to a json file
}
}
```
With the above setup, Shumai will write the benchmark results to json files:
json
{
"config": {
"name": "foo-foo-1",
"threads": [
1,
2,
3
],
"time": 1,
"a": 1
},
"load_results": null,
"env": {
"os_release": "5.10.60.1-microsoft-standard-WSL2",
"rustc_version": "1.59.0",
"hostname": "DESKTOP-DPOIAG6",
"cpu_num": 16,
"cpu_speed": 2894
},
"bench_results": [
{
"thread_cnt": 1,
"bench_results": [
110484492
],
"pcm": [],
"perf": null
},
{
"thread_cnt": 2,
"bench_results": [
222437918
],
"pcm": [],
"perf": null
},
{
"thread_cnt": 3,
"bench_results": [
315043334
],
"pcm": [],
"perf": null
}
]
}
The flamegraph feature generates the flamegraph of the benchmark function (instead of the whole program) with zero config.
The pcm feature collects pcm related data, such as l3 cache hit/miss, memory bandwidth (including DRAM and PM), UPI bandwidth etc. It requires a pcm-server running on the target host.
The perf feature collects common perf stats, such as CPU_CYCLES, INSTRUCTIONS, BRANCH_MISSES etc.
Note that the above features may be mutually exclusive, i.e. you may enable one feature at a time.
Shumai has two environment variables to control how the benchmark is executed:
- SHUMAI_THREAD: only run the benchmark with the specified number of threads, it must be specified in the benchmark config.
- SHUMAI_FILTER: filters the config, it must be a valid regex string.
k