Merge pull request #57 from mlcommons/HolisticTraceAnalysis

Support GPU-to-CPU synchronization dependency with HolisticTraceAnalysis

.github/workflows/end_to_end_tests.yml

@@ -26,6 +26,15 @@ jobs:
         git checkout 7b19f586dd8b267333114992833a0d7e0d601630
         pip install .
 
+    - name: Install HTA
+      run: |
+        git clone https://github.com/facebookresearch/HolisticTraceAnalysis.git
+        cd HolisticTraceAnalysis
+        git checkout d731cc2e2249976c97129d409a83bd53d93051f6
+        git submodule update --init
+        pip install -r requirements.txt
+        pip install -e .
+
     - name: Test chakra_trace_link Without Arguments
       run: |
         chakra_trace_link || [ $? -eq 2 ]

.github/workflows/python_tests.yml

@@ -26,6 +26,15 @@ jobs:
         git checkout 7b19f586dd8b267333114992833a0d7e0d601630
         pip install .
 
+    - name: Install HTA
+      run: |
+        git clone https://github.com/facebookresearch/HolisticTraceAnalysis.git
+        cd HolisticTraceAnalysis
+        git checkout d731cc2e2249976c97129d409a83bd53d93051f6
+        git submodule update --init
+        pip install -r requirements.txt
+        pip install -e .
+
     - name: Install Dependencies
       run: |
         pip install -r requirements-dev.txt

USER_GUIDE.md

@@ -35,7 +35,19 @@ $ git checkout 7b19f586dd8b267333114992833a0d7e0d601630
 $ pip install .
 ```
 
-### Step 4: Uninstalling Chakra
+### Step 4: Install Holistic Trace Analysis
+Installing Holistic Trace Analysis is necessary for Trace link.
+
+```bash
+$ git clone https://github.com/facebookresearch/HolisticTraceAnalysis.git
+$ cd HolisticTraceAnalysis
+$ git checkout d731cc2e2249976c97129d409a83bd53d93051f6
+$ git submodule update --init
+$ pip install -r requirements.txt
+$ pip install -e .
+```
+
+### Step 5: Uninstalling Chakra
 To uninstall Chakra, use the following command within the virtual environment.
 
 ```bash
@@ -49,6 +61,7 @@ Merge Chakra host execution trace and Chakra device execution trace to encode GP
 $ chakra_trace_link \
     --chakra-host-trace /path/to/chakra_host_trace \
     --chakra-device-trace /path/to/chakra_device_trace \
+    --rank [RANK] \
     --output-file /path/to/chakra_host_device_trace.json
 ```
 

src/converter/pytorch_converter.py

@@ -244,15 +244,23 @@ def convert_json_to_protobuf_nodes(
                             [
                                 ChakraAttr(name="comm_type", int64_val=collective_comm_type),
                                 ChakraAttr(name="comm_size", int64_val=pytorch_gpu_node.comm_size),
-                                *( [ChakraAttr(name="pg_name", string_val=pytorch_gpu_node.pg_name)] if pytorch_gpu_node.pg_name != "" else [] ),
+                                *(
+                                    [ChakraAttr(name="pg_name", string_val=pytorch_gpu_node.pg_name)]
+                                    if pytorch_gpu_node.pg_name != ""
+                                    else []
+                                ),
                             ]
                         )
 
                     elif chakra_gpu_node.type in {COMM_SEND_NODE, COMM_RECV_NODE}:
                         chakra_gpu_node.attr.extend(
                             [
                                 ChakraAttr(name="comm_size", int64_val=pytorch_gpu_node.comm_size),
-                                *( [ChakraAttr(name="pg_name", string_val=pytorch_gpu_node.pg_name)] if pytorch_gpu_node.pg_name != "" else [] ),
+                                *(
+                                    [ChakraAttr(name="pg_name", string_val=pytorch_gpu_node.pg_name)]
+                                    if pytorch_gpu_node.pg_name != ""
+                                    else []
+                                ),
                             ]
                         )
 
@@ -449,6 +457,15 @@ def convert_ctrl_dep_to_data_dep(
                 last_visited_non_gpu = current_node
                 last_visited_any = current_node
 
+            if json_node.sync_dep:
+                for sync_dep in json_node.sync_dep:
+                    if sync_dep not in current_node.data_deps:
+                        current_node.data_deps.append(sync_dep)
+                        logging.info(
+                            f"Node ID {current_node.id} now has an synchonization dependency on Node ID {sync_dep}"
+                        )
+
+            # Add children to the stack
             children_chakra_ids = [child.id for child in json_node.children]
             for child_chakra_id in sorted(children_chakra_ids, reverse=True):
                 child_chakra_node = protobuf_node_map.get(child_chakra_id)

src/converter/pytorch_node.py

@@ -19,7 +19,7 @@ class PyTorchNodeType(Enum):
     CPU_OP = 1
     GPU_OP = 2
     LABEL = 3  # Non-operator nodes
-    METADATA = 4 # Metadata nodes
+    METADATA = 4  # Metadata nodes
 
 
 class PyTorchNode:
@@ -110,6 +110,7 @@ def _parse_data_1_0_3_chakra_0_0_4(self, node_data: Dict[str, Any]) -> None:
         self.exclusive_dur = node_data.get("exclusive_dur", 0)
         self.ts = node_data.get("ts")
         self.inter_thread_dep = node_data.get("inter_thread_dep")
+        self.sync_dep = node_data.get("sync_dep")
         self.cat = node_data.get("cat")
         self.stream = node_data.get("stream", 0)
         # In Colletive comms nodes, pg_name is in node_data if exists.

src/trace_link/chakra_device_trace_loader.py

@@ -16,6 +16,7 @@ def load(
     ) -> Tuple[
         List[KinetoOperator],
         Dict[int, List[KinetoOperator]],
+        Dict[int, List[KinetoOperator]],
         Dict[int, KinetoOperator],
         List[KinetoOperator],
         Dict[int, KinetoOperator],
@@ -26,6 +27,7 @@ def load(
         Dict[int, KinetoOperator],
         List[KinetoOperator],
         List[int],
+        Dict[int, KinetoOperator],
     ]:
         """
         Load and process the Chakra device trace.
@@ -57,6 +59,7 @@ def load(
         logging.debug("Chakra device trace has been loaded and processed successfully.")
         return (
             dev_data["kineto_cpu_ops"],
+            dev_data["kineto_tid_ops_map"],
             dev_data["kineto_tid_cpu_ops_map"],
             dev_data["kineto_correlation_cuda_runtime_map"],
             dev_data["kineto_gpu_ops"],
@@ -68,6 +71,7 @@ def load(
             dev_data["kineto_rf_id_to_kineto_op_map"],
             dev_data["sorted_kineto_cpu_ops"],
             dev_data["sorted_kineto_cpu_op_ts"],
+            dev_data["kineto_external_id_to_kineto_op_map"],
         )
 
     def construct_dev_data_structures(self, kineto_ops: List[KinetoOperator], trace_file: str) -> Dict:
@@ -90,13 +94,17 @@ def construct_dev_data_structures(self, kineto_ops: List[KinetoOperator], trace_
         thread_info = {}
 
         kineto_cpu_ops = []
+        kineto_tid_ops_map = {}
         kineto_tid_cpu_ops_map = {}
         kineto_correlation_cuda_runtime_map = {}
         kineto_gpu_ops = []
         kineto_id_arrow_op_map = {}
         kineto_id_cuda_launch_op_map = {}
+        kineto_external_id_to_kineto_op_map = {}
 
         for op in kineto_ops:
+            kineto_tid_ops_map.setdefault(op.tid, []).append(op)
+
             if op.is_cpu_op():
                 kineto_cpu_ops.append(op)
                 kineto_tid_cpu_ops_map.setdefault(op.tid, []).append(op)
@@ -144,10 +152,14 @@ def construct_dev_data_structures(self, kineto_ops: List[KinetoOperator], trace_
                 thread_start_end[0] = min(thread_start_end[0], op.timestamp)
                 thread_start_end[1] = max(thread_start_end[1], op.timestamp + op.inclusive_dur)
 
+            if op.external_id is not None:
+                kineto_external_id_to_kineto_op_map[op.external_id] = op
+
         kineto_rf_id_to_kineto_op_map = {op.rf_id: op for op in kineto_cpu_ops if op.rf_id is not None}
 
         return {
             "kineto_cpu_ops": kineto_cpu_ops,
+            "kineto_tid_ops_map": kineto_tid_ops_map,
             "kineto_tid_cpu_ops_map": kineto_tid_cpu_ops_map,
             "kineto_correlation_cuda_runtime_map": kineto_correlation_cuda_runtime_map,
             "kineto_gpu_ops": kineto_gpu_ops,
@@ -159,6 +171,7 @@ def construct_dev_data_structures(self, kineto_ops: List[KinetoOperator], trace_
             "kineto_rf_id_to_kineto_op_map": kineto_rf_id_to_kineto_op_map,
             "sorted_kineto_cpu_ops": [],
             "sorted_kineto_cpu_op_ts": [],
+            "kineto_external_id_to_kineto_op_map": kineto_external_id_to_kineto_op_map,
         }
 
     def calculate_exclusive_dur(self, kineto_tid_cpu_ops_map: Dict[int, List[KinetoOperator]]) -> None:

src/trace_link/kineto_operator.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Optional
+from typing import Any, Dict, List, Optional
 
 from et_replay.execution_trace import Node as PyTorchOperator
 
@@ -22,6 +22,7 @@ class KinetoOperator:
         host_op (Optional[PyTorchOperator]): Corresponding PyTorch operator object.
         parent_host_op_id (Optional[int]): ID of the parent PyTorch operator.
         inter_thread_dep (Optional[int]): Identifier for inter-thread dependencies.
+        sync_dep (List[KinetoOperator]): List of KinetoOperator objects that have dependencies on this operator.
         stream (Optional[int]): CUDA stream identifier associated with the operator.
         rf_id (Optional[int]): Record function identifier.
         correlation (int): Identifier used to correlate CUDA runtime and GPU operations.
@@ -49,6 +50,7 @@ def __init__(self, kineto_op: Dict[str, Any]) -> None:
         self.host_op: Optional[PyTorchOperator] = None
         self.parent_host_op_id: Optional[int] = None
         self.inter_thread_dep: Optional[int] = None
+        self.sync_dep: List[KinetoOperator] = []
         self.stream: Optional[int] = kineto_op.get("args", {}).get("stream", None)
         self.rf_id: Optional[int] = kineto_op.get("args", {}).get("Record function id", None)
         self.correlation: int = kineto_op.get("args", {}).get("correlation", -1)
@@ -61,13 +63,14 @@ def __repr__(self) -> str:
         Returns
             str: A string representation of the KinetoOperator.
         """
+        sync_dep_ids = [op.id for op in self.sync_dep]
         return (
             f"KinetoOperator(id={self.id}, category={self.category}, name={self.name}, "
             f"phase={self.phase}, inclusive_dur={self.inclusive_dur}, "
             f"exclusive_dur={self.exclusive_dur}, timestamp={self.timestamp}, "
             f"external_id={self.external_id}, ev_idx={self.ev_idx}, tid={self.tid}, "
             f"parent_host_op_id={self.parent_host_op_id}, inter_thread_dep={self.inter_thread_dep}, "
-            f"stream={self.stream}, rf_id={self.rf_id}, correlation={self.correlation})"
+            f"sync_dep={sync_dep_ids}, stream={self.stream}, rf_id={self.rf_id}, correlation={self.correlation})"
         )
 
     def is_cpu_op(self) -> bool:

src/trace_link/trace_link.py

@@ -18,6 +18,7 @@ def main() -> None:
             "Merging-PyTorch-and-Kineto-Traces"
         )
     )
+    parser.add_argument("--rank", type=int, required=True, help="Rank for the input traces")
     parser.add_argument(
         "--chakra-host-trace",
         type=str,
@@ -43,10 +44,11 @@ def main() -> None:
     logging.basicConfig(level=args.log_level.upper())
 
     linker = TraceLinker()
-    linker.link(args.chakra_host_trace, args.chakra_device_trace, args.output_file)
+    linker.link(args.rank, args.chakra_host_trace, args.chakra_device_trace, args.output_file)
 
     logging.info(f"Linking process successful. Output file is available at {args.output_file}.")
     logging.info("Please run the chakra_converter for further postprocessing.")
 
+
 if __name__ == "__main__":
     main()