Added energy consumption to Skyline

protocol/innpv.proto

@@ -77,6 +77,7 @@ message FromServer {
     BreakdownResponse breakdown = 8;
 
     HabitatResponse habitat = 9;
+    EnergyResponse energy = 10;
   }
 
   // Deprecated messages
@@ -97,6 +98,33 @@ message HabitatResponse {
   repeated HabitatDevicePrediction predictions = 1;
 }
 
+// Energy messages
+// =======================================
+
+message EnergyResponse {
+  float total_consumption = 1;
+  repeated EnergyConsumptionComponent components = 2;
+
+  // A list of past energy measurements
+  repeated EnergyResponse past_measurements = 3;
+}
+
+// Reports the energy consumption of one system component (e.g. CPU+DRAM or GPU)
+enum EnergyConsumptionComponentType {
+  ENERGY_UNSPECIFIED = 0;
+  ENERGY_CPU_DRAM = 1;
+  ENERGY_NVIDIA = 2;
+}
+
+message EnergyConsumptionComponent {
+  EnergyConsumptionComponentType component_type = 1;
+  float consumption_joules = 2;
+}
+
+// Records past experiments
+
+
+
 // =======================================
 
 message InitializeResponse {

pyproject.toml

@@ -32,6 +32,7 @@ numpy = "^1.22"
 torch = "*"
 nvidia-ml-py3 = "*"
 toml = "^0.10.2"
+pyRAPL = "^0.2.3"
 
 [tool.poetry.dev-dependencies]
 

skyline/analysis/request_manager.py

@@ -107,6 +107,23 @@ def _handle_analysis_request(self, analysis_request, context):
                 context,
             )
 
+            # send energy response
+            if not context.state.connected:
+                logger.debug(
+                    'Aborting request %d from (%s:%d) early '
+                    'because the client has disconnected.',
+                    context.sequence_number,
+                    *(context.address),
+                )
+                return
+
+            energy_resp = next(analyzer)
+            self._enqueue_response(
+                self._send_energy_response,
+                energy_resp,
+                context,
+            )
+
             elapsed_time = time.perf_counter() - start_time
             logger.debug(
                 'Processed analysis request %d from (%s:%d) in %.4f seconds.',
@@ -174,3 +191,11 @@ def _send_habitat_response(self, habitat_resp, context):
         except:
             logger.exception(
                 'Exception occurred when sending a habitat response.')
+
+    def _send_energy_response(self, energy_resp, context):
+        # Called from the main executor. Do not call directly!
+        try:
+            self._message_sender.send_energy_response(energy_resp, context)
+        except:
+            logger.exception(
+                'Exception occurred when sending an energy response.')

skyline/analysis/runner.py

@@ -14,6 +14,9 @@ def analyze_project(project_root, entry_point, nvml):
     print("analyze_project: running habitat_predict()")
     yield session.habitat_predict()
 
+    print("analyze_project: running energy_compute()")
+    yield session.energy_compute()
+
 
 def main():
     # This is used for development and debugging purposes

skyline/analysis/session.py

@@ -11,6 +11,7 @@
 
 import skyline.protocol_gen.innpv_pb2 as pm
 from skyline.analysis.static import StaticAnalyzer
+from skyline.energy.measurer import EnergyMeasurer
 from skyline.exceptions import AnalysisError, exceptions_as_analysis_errors
 from skyline.profiler.iteration import IterationProfiler
 from skyline.tracking.tracker import Tracker
@@ -131,6 +132,38 @@ def new_from(cls, project_root, entry_point):
             StaticAnalyzer(entry_point_code, entry_point_ast),
         )
 
+    def energy_compute(self) -> pm.EnergyResponse:
+        energy_measurer = EnergyMeasurer()
+
+        model = self._model_provider()
+        inputs = self._input_provider()
+        iteration = self._iteration_provider(model)
+        resp = pm.EnergyResponse()
+        try: 
+            energy_measurer.begin_measurement()
+            iterations = 20
+            for _ in range(iterations):
+                iteration(*inputs)
+            energy_measurer.end_measurement()
+
+            resp.total_consumption = energy_measurer.total_energy()/float(iterations)
+
+            cpu_component = pm.EnergyConsumptionComponent()
+            cpu_component.component_type = pm.ENERGY_CPU_DRAM
+            cpu_component.consumption_joules = energy_measurer.cpu_energy()/float(iterations)
+
+            gpu_component = pm.EnergyConsumptionComponent()
+            gpu_component.component_type = pm.ENERGY_NVIDIA
+            gpu_component.consumption_joules = energy_measurer.gpu_energy()/float(iterations)
+
+            resp.components.extend([cpu_component, gpu_component])
+
+            #TODO save each response to a database and get the past energy measurements
+        except PermissionError as err:
+            # Remind user to set their CPU permissions
+            print(err)
+        return resp
+
     def habitat_compute_threshold(self, runnable, context):
         tracker = habitat.OperationTracker(context.origin_device)
         with tracker.track():

skyline/energy/measurer.py

@@ -0,0 +1,117 @@
+import time
+from threading import Thread
+import numpy as np
+
+import pynvml as N
+from pyRAPL import Sensor
+
+class CPUMeasurer:
+    def __init__(self, interval):
+        self.interval = interval
+        self.power = []
+        self.last_cpu = None
+        self.last_dram = None
+
+    def measurer_init(self):
+        self.sensor = Sensor()
+        energy = self.sensor.energy()
+        self.last_cpu = np.array(energy[0::2])
+        self.last_dram = np.array(energy[1::2])
+
+    def measurer_measure(self):
+        # Get energy consumed so far (since last CPU reset)
+        energy = self.sensor.energy()
+        cpu = np.array(energy[0::2])
+        dram = np.array(energy[1::2])
+
+        # Compare against last measurement to determine energy since last measure
+        diff_cpu = cpu - self.last_cpu
+        diff_dram = dram - self.last_dram
+
+        # 1J = 10^6 uJ
+        # The cpu used this much since the last measurement
+        # We have mW = 1000*J/s = 1000*(uJ/10^6)/s
+        cpu_total = np.sum(diff_cpu)
+        cpu_mW = 1000 * (cpu_total / 1e6) / self.interval
+        self.power.append(cpu_mW)
+
+        self.last_cpu = cpu
+        self.last_dram = dram
+
+    def measurer_deallocate(self):
+        pass
+
+    def total_energy(self):
+        # J = W * s,    1W = 1000 mW
+        energy = self.interval * sum(self.power) / 1000.0
+        return energy
+
+class GPUMeasurer:
+    def __init__(self, interval):
+        self.interval = interval
+        self.power = []
+
+    def measurer_init(self):
+        N.nvmlInit()
+        self.device_handle = N.nvmlDeviceGetHandleByIndex(0)
+
+    def measurer_measure(self):
+        power = N.nvmlDeviceGetPowerUsage(self.device_handle)
+        self.power.append(power)
+
+    def measurer_deallocate(self):
+        N.nvmlShutdown()
+
+    def total_energy(self):
+        # J = W * s,    1W = 1000 mW
+        energy = self.interval * sum(self.power) / 1000.0
+        return energy
+
+class EnergyMeasurer:
+    def __init__(self):
+        self.sleep_interval = 0.1
+        self.measuring = False
+        self.measure_thread = None
+
+        self.measurers = {
+            "cpu": CPUMeasurer(self.sleep_interval),
+            "gpu": GPUMeasurer(self.sleep_interval),
+        }
+
+    def run_measure(self):
+        # Initialize
+        for m in self.measurers:
+            self.measurers[m].measurer_init()
+
+        # Run measurement loop
+        while self.measuring:
+            for m in self.measurers:
+                self.measurers[m].measurer_measure()
+            time.sleep(self.sleep_interval)
+
+        # Cleanup
+        for m in self.measurers:
+            self.measurers[m].measurer_deallocate()
+
+    def begin_measurement(self):
+        assert(self.measure_thread is None)
+        self.measure_thread = Thread(target=self.run_measure)
+        self.measuring = True
+        self.measure_thread.start()
+
+    def end_measurement(self):
+        self.measuring = False
+        self.measure_thread.join()
+        self.measure_thread = None
+
+    def total_energy(self):
+        total_energy = 0.
+        for m in self.measurers:
+            total_energy += self.measurers[m].total_energy() 
+        return total_energy
+
+    def cpu_energy(self):
+        return self.measurers["cpu"].total_energy()
+
+    def gpu_energy(self):
+        return self.measurers["gpu"].total_energy()

skyline/protocol/message_sender.py

@@ -64,6 +64,9 @@ def send_throughput_response(self, throughput, context):
 
     def send_habitat_response(self, habitat_resp, context):
         self._send_message(habitat_resp, 'habitat', context)
+
+    def send_energy_response(self, energy_resp, context):
+        self._send_message(energy_resp, 'energy', context)
 
     def _send_message(self, message, payload_name, context):
         try: