[SD-113] Measure power consumption related to model in memory

jetson/power_logging/model/benchmark.py

@@ -29,6 +29,9 @@ class BenchmarkMetrics(BaseModel):
     latencies: list[float]  # in seconds
     avg_latency: float  # in seconds
     avg_throughput: float
+    warmup: tuple[str, str]
+    model_load: tuple[str, str]
+    pt_model_load_end: str
 
 
 def load_model(model_name: str) -> Any:
@@ -74,6 +77,7 @@ def benchmark(args: argparse.Namespace) -> None:
     input_data = torch.randn(args.input_shape, device=DEVICE)
     model = load_model(args.model)
     model.eval().to(DEVICE)
+    pt_model_load_end = datetime.now().strftime("%Y%m%d-%H%M%S")
 
     dtype = torch.float32
     if args.dtype == "float16":
@@ -85,6 +89,7 @@ def benchmark(args: argparse.Namespace) -> None:
     model = model.to(dtype)
     print(f"Using {DEVICE=} for benchmarking")
 
+    model_load_start = datetime.now().strftime("%Y%m%d-%H%M%S")
     exp_program = torch.export.export(model, tuple([input_data]))
     model = torch_tensorrt.dynamo.compile(
         exported_program=exp_program,
@@ -99,20 +104,30 @@ def benchmark(args: argparse.Namespace) -> None:
         # Setting it to True returns PythonTorchTensorRTModule which has different profiling approach
         use_python_runtime=True,
     )
+    model_load_end = datetime.now().strftime("%Y%m%d-%H%M%S")
+
+    print("Sleeping for 10 seconds...")
+    time.sleep(10)
 
     st = time.perf_counter()
     print("Warm up ...")
+    warmup_start = datetime.now().strftime("%Y%m%d-%H%M%S")
     with torch.no_grad():
         for _ in range(args.warmup):
             _ = model(input_data)
+    warmup_end = datetime.now().strftime("%Y%m%d-%H%M%S")
     print(f"Warm complete in {time.perf_counter() - st:.2f} sec ...")
 
+    print("Sleeping for 10 seconds...")
+    time.sleep(10)
+
     print("Start timing using tensorrt backend ...")
     torch.cuda.synchronize()
     # Recorded in milliseconds
     start_events = [torch.cuda.Event(enable_timing=True) for _ in range(args.runs)]
     end_events = [torch.cuda.Event(enable_timing=True) for _ in range(args.runs)]
 
+    print("Starting inference...")
     with torch.no_grad():
         for i in tqdm(range(args.runs)):
             # Hack for enabling profiling
@@ -129,7 +144,7 @@ def benchmark(args: argparse.Namespace) -> None:
             _ = model(input_data)
             end_events[i].record()
 
-            print("Sleeping for 1 second(s)...")
+            # print("Sleeping for 1 second(s)...")
             time.sleep(1)
 
         end.record()
@@ -153,6 +168,9 @@ def benchmark(args: argparse.Namespace) -> None:
         latencies=timings,  # in seconds
         avg_throughput=avg_throughput,
         avg_latency=np.mean(timings),  # in seconds
+        warmup=(warmup_start, warmup_end),
+        model_load=(model_load_start, model_load_end),
+        pt_model_load_end=pt_model_load_end,
     )
 
     model_dir = f"{args.result_dir}/{args.model}"

jetson/power_logging/power_research/__init__.py

@@ -0,0 +1 @@
+"""Init for power research."""

jetson/power_logging/power_research/inference_cycles/Alexnet_Inference_cycle_power_analysis.ipynb

@@ -10,6 +10,7 @@
     "from pathlib import Path\n",
     "import json\n",
     "import matplotlib.pyplot as plt\n",
+    "import sys; sys.path.insert(0, '..')\n",
     "\n",
     "from utils import parse_latencies, parse_power_log"
    ]

jetson/power_logging/power_research/inference_cycles/Resnet_Inference_cycle_power_analysis.ipynb

@@ -10,6 +10,7 @@
     "from pathlib import Path\n",
     "import json\n",
     "import matplotlib.pyplot as plt\n",
+    "import sys; sys.path.insert(0, '..')\n",
     "\n",
     "from utils import parse_latencies, parse_power_log"
    ]

jetson/power_logging/power_research/inference_cycles/VGG_Inference_cycle_power_analysis.ipynb

@@ -10,6 +10,7 @@
     "from pathlib import Path\n",
     "import json\n",
     "import matplotlib.pyplot as plt\n",
+    "import sys; sys.path.insert(0, '..')\n",
     "\n",
     "from utils import parse_latencies, parse_power_log"
    ]

jetson/power_logging/power_research/inference_cycles/utils.py → jetson/power_logging/power_research/utils.py

@@ -56,6 +56,15 @@ def parse_latencies(
     return latency_data
 
 
+def get_power_index(power_data, timestamp):
+    index = 0
+    while True:
+        power_timestamp = power_data[index][0]
+        if power_timestamp > timestamp:
+            return index
+        index += 1
+
+
 def plot_inference_cycle_power(
     cycle: int, power_data: list[tuple], latency_data: dict[list[tuple]], delay: int
 ) -> None:

jetson/power_logging/raw_data/.gitignore

@@ -1,2 +1,3 @@
 /prebuilt_models
 /power_profile_between_inference
+/power_profile_model_loading

jetson/power_logging/raw_data/power_profile_model_loading.dvc

@@ -0,0 +1,6 @@
+outs:
+- md5: e4ea48d9fa9ba9a249f750e3032ab84e.dir
+  size: 531338892
+  nfiles: 21
+  hash: md5
+  path: power_profile_model_loading

jetson/power_logging/run_experiment.sh

@@ -6,7 +6,7 @@ set -eou pipefail
 IDLE_DURATION=120
 
 # Directory to store results
-RESULT_DIR="raw_data/power_profile_between_inference"
+RESULT_DIR="raw_data/power_profile_model_loading"
 
 echo "Running idling power measurement..."
 python measure_idling_power.py \
@@ -25,7 +25,7 @@ sleep 120
 # models=("alexnet" "vgg11" "vgg13" "vgg16" "vgg19" "mobilenet_v2" "mobilenet_v3_small" "mobilenet_v3_large" "resnet18" "resnet34" "resnet50" "resnet101" "resnet152" "lenet" "resnext50_32x4d" "resnext101_32x8d" "resnext101_64x4d" "convnext_tiny" "convnext_small" "convnext_base")
 # # Number of inference cycles
 # RUNS=30000
-models=("alexnet" "vgg16" "vgg19")
+models=("resnet18" "resnet34" "alexnet" "vgg16" "vgg19")
 RUNS=3000
 
 # Iterate through models and run measure_inference_power.py script