summaryrefslogtreecommitdiff
path: root/src/energy
diff options
context:
space:
mode:
Diffstat (limited to 'src/energy')
-rw-r--r--src/energy/estimator.rs2
-rw-r--r--src/energy/trackers/kernel.rs2
-rw-r--r--src/energy/trackers/mock.rs2
-rw-r--r--src/energy/trackers/perf.rs87
4 files changed, 53 insertions, 40 deletions
diff --git a/src/energy/estimator.rs b/src/energy/estimator.rs
index 85cb0eb..7964b00 100644
--- a/src/energy/estimator.rs
+++ b/src/energy/estimator.rs
@@ -1,5 +1,5 @@
pub trait Estimator: Send + 'static {
- fn start_trace(&mut self, pid: u64);
+ fn start_trace(&mut self, pid: u64, running_on_e_core: bool);
fn stop_trace(&mut self, pid: u64);
fn read_consumption(&mut self, pid: u64, cpu: i32) -> Option<f64>;
}
diff --git a/src/energy/trackers/kernel.rs b/src/energy/trackers/kernel.rs
index 89a0d89..f7e0811 100644
--- a/src/energy/trackers/kernel.rs
+++ b/src/energy/trackers/kernel.rs
@@ -30,7 +30,7 @@ const STOP_TRACE: Ioctl<Write, &u64> = unsafe { PERF_MON.write(0x81) };
const READ_POWER: Ioctl<WriteRead, &u64> = unsafe { PERF_MON.write_read(0x82) };
impl Estimator for KernelDriver {
- fn start_trace(&mut self, pid: u64) {
+ fn start_trace(&mut self, pid: u64, running_on_e_core: bool) {
let _ = START_TRACE.ioctl(&mut self.file, &pid);
}
diff --git a/src/energy/trackers/mock.rs b/src/energy/trackers/mock.rs
index 597cab0..b2a504e 100644
--- a/src/energy/trackers/mock.rs
+++ b/src/energy/trackers/mock.rs
@@ -3,7 +3,7 @@ use crate::energy::estimator::Estimator;
pub struct MockEstimator;
impl Estimator for MockEstimator {
- fn start_trace(&mut self, _pid: u64) {}
+ fn start_trace(&mut self, _pid: u64, _running_on_e_core: bool) {}
fn stop_trace(&mut self, _pid: u64) {}
diff --git a/src/energy/trackers/perf.rs b/src/energy/trackers/perf.rs
index 71e87d8..8da2bf8 100644
--- a/src/energy/trackers/perf.rs
+++ b/src/energy/trackers/perf.rs
@@ -9,22 +9,25 @@ use perf_event::{
use crate::freq::FrequencyKHZ;
use std::sync::{Arc, RwLock};
-use crate::energy::{rapl, Estimator};
+use crate::energy::Estimator;
use crate::model::ArrayBackend;
pub struct PerfEstimator {
registry: HashMap<u64, Counters>,
- model: crate::model::Net<ArrayBackend>,
+ model_p: crate::model::Net<ArrayBackend>,
+ model_e: crate::model::Net<ArrayBackend>,
device: <ArrayBackend as burn::prelude::Backend>::Device,
shared_cpu_current_frequencies: Arc<RwLock<Vec<FrequencyKHZ>>>,
}
impl PerfEstimator {
pub fn new(shared_cpu_current_frequencies: Arc<RwLock<Vec<FrequencyKHZ>>>) -> Self {
- let model = crate::model::load_model();
+ let model_p = crate::model::load_model_p();
+ let model_e = crate::model::load_model_e();
Self {
registry: Default::default(),
- model,
+ model_p,
+ model_e,
device: Default::default(),
shared_cpu_current_frequencies,
}
@@ -36,9 +39,21 @@ struct Counters {
counters: Vec<Counter>,
old_time: u64,
old_total_energy: f64,
+ running_on_e_core: bool,
}
-static EVENT_TYPES: &[Event] = &[
+static EVENT_TYPES_P: &[Event] = &[
+ Event::Hardware(Hardware::BRANCH_INSTRUCTIONS),
+ Event::Hardware(Hardware::BRANCH_MISSES),
+ Event::Hardware(Hardware::CACHE_MISSES),
+ Event::Hardware(Hardware::CACHE_REFERENCES),
+ Event::Hardware(Hardware::CPU_CYCLES),
+ Event::Hardware(Hardware::INSTRUCTIONS),
+ Event::Hardware(Hardware::REF_CPU_CYCLES),
+ Event::Software(Software::TASK_CLOCK),
+];
+//TODO: use correct counter
+static EVENT_TYPES_E: &[Event] = &[
Event::Hardware(Hardware::BRANCH_INSTRUCTIONS),
Event::Hardware(Hardware::BRANCH_MISSES),
Event::Hardware(Hardware::CACHE_MISSES),
@@ -50,8 +65,7 @@ static EVENT_TYPES: &[Event] = &[
];
impl Estimator for PerfEstimator {
- fn start_trace(&mut self, pid: u64) {
- //let Ok(mut group) = Group::new_with_pid_and_cpu(-1, 0) else {
+ fn start_trace(&mut self, pid: u64, running_on_e_core: bool) {
let mut group = match Group::new_with_pid_and_cpu(pid as i32, -1) {
Ok(counters) => counters,
Err(e) => {
@@ -62,23 +76,21 @@ impl Estimator for PerfEstimator {
return;
}
};
- //let Ok(mut group) = Group::new_with_pid_and_cpu(pid as i32, -1) else {
- // eprintln!("Failed to create performance counter group for PID {}", pid);
- // return;
- //};
- let counters: Result<Vec<_>, _> = EVENT_TYPES
- .iter()
- .map(|kind| {
- Builder::new()
- .group(&mut group)
- .kind(kind.clone())
- .observe_pid(pid as i32)
- //.observe_pid(-1)
- //.one_cpu(0)
- .build()
- })
- .collect();
+ let counters: Result<Vec<_>, _> = if running_on_e_core {
+ EVENT_TYPES_E
+ } else {
+ EVENT_TYPES_P
+ }
+ .iter()
+ .map(|kind| {
+ Builder::new()
+ .group(&mut group)
+ .kind(kind.clone())
+ .observe_pid(pid as i32)
+ .build()
+ })
+ .collect();
let counters = match counters {
Ok(counters) => counters,
@@ -95,13 +107,17 @@ impl Estimator for PerfEstimator {
eprintln!("Failed to enable performance counters: {}", e);
return;
}
- group.reset();
+ if let Err(e) = group.reset() {
+ eprintln!("Failed to reset performance counters: {}", e);
+ return;
+ }
let counters = Counters {
counters,
group,
old_time: 0,
old_total_energy: 0.,
+ running_on_e_core,
};
self.registry.insert(pid, counters);
}
@@ -122,36 +138,33 @@ impl Estimator for PerfEstimator {
return None;
}
};
- //let task_clock = counts[&counters.counters[7]];
- let task_clock = counts[&counters.counters[7]];
+ let num_counter = counters.counters.len();
+ let task_clock = counts[&counters.counters[num_counter - 1]];
if task_clock == 0 {
return None;
}
- let current_time = std::time::Instant::now();
let time_running_ns = counts.time_running();
let correction_factor = 10_000_000. / (time_running_ns - counters.old_time) as f64;
counters.old_time = time_running_ns;
- //dbg!(time_running_ns);
- //dbg!(counters.old_time.elapsed().as_millis());
let mut values =
vec![(self.shared_cpu_current_frequencies.read().unwrap()[cpu as usize] / 1000) as f64];
- //let mut values = vec![crate::benchmark::read_cpu_frequency(0).unwrap()];
- for ty in counters.counters.iter().take(7) {
+ for ty in counters.counters.iter().take(num_counter - 1) {
let count: u64 = counts[&ty];
values.push((count as f64) * correction_factor);
}
- //dbg!(&values);
- let result = self
- .model
- .forward(Tensor::from_floats(&values.as_slice()[0..8], &self.device));
+
+ let result = if counters.running_on_e_core {
+ &self.model_e
+ } else {
+ &self.model_p
+ }
+ .forward(Tensor::from_floats(&values.as_slice()[0..], &self.device));
let energy = result.into_scalar() as f64;
- //dbg!(energy, correction_factor);
counters.old_total_energy += energy / correction_factor;
counters.group.reset().unwrap();
- //dbg!(counters.old_total_energy);
Some(energy / correction_factor)
}
}
generated by cgit v1.3.1 (git 2.54.0) at 2026-10-17 22:01:03 +0000