diff options
Diffstat (limited to 'src/energy/trackers/perf.rs')
| -rw-r--r-- | src/energy/trackers/perf.rs | 106 |
1 files changed, 26 insertions, 80 deletions
diff --git a/src/energy/trackers/perf.rs b/src/energy/trackers/perf.rs index 96c73d9..8862e91 100644 --- a/src/energy/trackers/perf.rs +++ b/src/energy/trackers/perf.rs @@ -2,34 +2,30 @@ use std::collections::HashMap; use burn::tensor::Tensor; use perf_event::{ - events::{Event, Hardware, Software}, - Builder, Counter, Group, + events::{Event, Hardware}, + Builder, Counter, }; use crate::freq::FrequencyKHZ; use std::sync::{Arc, RwLock}; use crate::energy::Estimator; -use crate::model::ArrayBackend; +use crate::model::BurnBackend; pub struct PerfEstimator { registry: HashMap<u64, Counters>, - model_p: crate::model::Net<ArrayBackend>, - model_e: crate::model::Net<ArrayBackend>, - device: <ArrayBackend as burn::prelude::Backend>::Device, + model: crate::model::Net<BurnBackend>, + device: <BurnBackend 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_p = crate::model::load_model("perf_pcore.pt"); - let model_p = crate::model::load_model("perf.pt"); - let model_e = crate::model::load_model("perf.pt"); + let model = crate::model::load_model("perf.pt"); // let model_e = crate::model::load_model("perf_ecore.pt"); Self { registry: Default::default(), - model_p, - model_e, + model, device: Default::default(), shared_cpu_current_frequencies, } @@ -37,67 +33,41 @@ impl PerfEstimator { } struct Counters { - group: Group, counters: Vec<Counter>, old_time: u64, old_total_energy: f64, cpu: i32, - running_on_e_core: bool, } 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), + #[cfg(not(feature = "amd"))] Event::Hardware(Hardware::REF_CPU_CYCLES), ]; -static EVENT_TYPES_E: &[Event] = &[ - 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), -]; impl Estimator for PerfEstimator { - fn start_trace(&mut self, pid: u64, cpu: i32, running_on_e_core: bool) -> Result<(), ()> { - let mut group = match Group::new_with_pid_and_cpu(pid as i32, -1) { - Ok(counters) => counters, - Err(e) => { - eprintln!( - "Failed to create performance counter group for PID {}: {}", - pid, e - ); - return Err(()); - } - }; - - let counters: Result<Vec<_>, _> = if running_on_e_core || true { - // println!("starting e core counter"); - EVENT_TYPES_E - } else { - EVENT_TYPES_P - } - .iter() - .map(|kind| { - Builder::new() - // .group(&mut group) - .kind(kind.clone()) - .observe_pid(pid as i32) - .inherit_thread(true) - .build() - }) - .collect(); + fn start_trace(&mut self, pid: u64, cpu: i32) -> Result<(), ()> { + let counters: Result<Vec<_>, _> = EVENT_TYPES_P + .iter() + .map(|kind| { + Builder::new() + // .group(&mut group) + .kind(kind.clone()) + .observe_pid(pid as i32) + .inherit_thread(true) + .build() + }) + .collect(); let mut counters = match counters { Ok(counters) => counters, Err(e) => { eprintln!( - "Failed to create performance counter group for PID {}: {}", + "Failed to create performance counter for PID {}: {}", pid, e ); return Err(()); @@ -118,11 +88,9 @@ impl Estimator for PerfEstimator { let old_time = counters[0].read_count_and_time().unwrap().time_running; let counters = Counters { counters, - group, old_time, old_total_energy: 0., cpu, - running_on_e_core, }; self.registry.insert(pid, counters); Ok(()) @@ -132,20 +100,12 @@ impl Estimator for PerfEstimator { self.registry.remove(&pid); } - fn update_information(&mut self, pid: u64, cpu: i32, is_ecore: bool) { - let mut core_type_changed = false; + fn update_information(&mut self, pid: u64, cpu: i32) { if let Some(info) = self.registry.get_mut(&pid) { info.cpu = cpu; - info.running_on_e_core = is_ecore; - core_type_changed = is_ecore != info.running_on_e_core; } else { eprintln!("Tried to update an unknown task") } - if core_type_changed { - // println!("migrating task to {}", cpu); - // self.stop_trace(pid); - // self.start_trace(pid, cpu, is_ecore); - } } fn read_consumption(&mut self, pid: u64) -> Option<f64> { @@ -154,13 +114,6 @@ impl Estimator for PerfEstimator { return None; }; - // let counts = match counters.group.read() { - // Ok(counts) => counts, - // Err(e) => { - // println!("failed to read group: {e}"); - // return None; - // } - // }; let time_running_ns = counters.counters[0] .read_count_and_time() .unwrap() @@ -172,7 +125,6 @@ impl Estimator for PerfEstimator { counters.old_time = time_running_ns; let mut values = vec![ - //if counters.running_on_e_core { 1. } else { 0. }, (self.shared_cpu_current_frequencies.read().unwrap()[counters.cpu as usize] / 1000) as f64, ]; @@ -181,21 +133,15 @@ impl Estimator for PerfEstimator { let count: u64 = ty.read().unwrap(); values.push((count as f64) * correction_factor); } + #[cfg(feature = "amd")] values.push(values[4]); - 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 result = self + .model + .forward(Tensor::from_floats(&values.as_slice()[0..], &self.device)); let energy = result.into_scalar() as f64; - if counters.running_on_e_core { - // dbg!(energy); - } counters.old_total_energy += energy / correction_factor; - counters.group.reset().unwrap(); for counter in counters.counters.iter_mut() { counter.reset().unwrap(); } |