Add nx-cugraph Benchmarking Scripts

benchmarks/nx-cugraph/pytest-based/README.md

@@ -0,0 +1,45 @@
+## `nx-cugraph` Benchmarks
+
+### Overview
+
+This directory contains a set of scripts designed to benchmark NetworkX with the `nx-cugraph` backend and deliver a report that summarizes the speed-up and runtime deltas over default NetworkX.
+
+Our current benchmarks provide the following datasets:
+
+| Dataset     | Nodes | Edges | Directed |
+| --------    | ------- | ------- | ------- |
+| netscience  | 1,461    | 5,484 | Yes |
+| email-Eu-core  | 1,005    | 25,571 | Yes |
+| cit-Patents  | 3,774,768    | 16,518,948 | Yes |
+| hollywood  | 1,139,905    | 57,515,616 | No |
+| soc-LiveJournal1  | 4,847,571    | 68,993,773 | Yes |
+
+
+
+### Scripts
+
+#### 1. `run-gap-benchmarks.sh`
+This script allows users to run selected algorithms across multiple datasets and backends. All results are stored inside a sub-directory (`logs/`) and output files are named based on the combination of parameters for that benchmark.
+
+NOTE: If running with all algorithms, datasets, and backends, this script may take a few hours to finish running.
+
+**Usage:**
+  ```bash
+  bash run-gap-benchmarks.sh  # edit this script directly
+  ```
+
+#### 2. `get_graph_bench_dataset.py`
+This script downloads the specified dataset using `cugraph.datasets`.
+
+**Usage:**
+  ```bash
+  python get_graph_bench_dataset.py [dataset]
+  ```
+
+#### 3. `create_results_summary_page.py`
+This script is designed to be run after `run-gap-benchmarks.sh` in order to generate an HTML page displaying a results table comparing default NetworkX to nx-cugraph. The script also provides information about the current system.
+
+**Usage:**
+  ```bash
+  python create_results_summary_page.py > report.html
+  ```

benchmarks/nx-cugraph/pytest-based/bench_algos.py

@@ -271,9 +271,8 @@ def bench_from_networkx(benchmark, graph_obj):
 
 
 # normalized_param_values = [True, False]
-# k_param_values = [10, 100]
 normalized_param_values = [True]
-k_param_values = [10]
+k_param_values = [10, 100, 1000]
 
 
 @pytest.mark.parametrize(
@@ -282,6 +281,10 @@ def bench_from_networkx(benchmark, graph_obj):
 @pytest.mark.parametrize("k", k_param_values, ids=lambda k: f"{k=}")
 def bench_betweenness_centrality(benchmark, graph_obj, backend_wrapper, normalized, k):
     G = get_graph_obj_for_benchmark(graph_obj, backend_wrapper)
+
+    if k > G.number_of_nodes():
+        pytest.skip(reason=f"{k=} > {G.number_of_nodes()=}")
+
     result = benchmark.pedantic(
         target=backend_wrapper(nx.betweenness_centrality),
         args=(G,),
@@ -305,6 +308,10 @@ def bench_edge_betweenness_centrality(
     benchmark, graph_obj, backend_wrapper, normalized, k
 ):
     G = get_graph_obj_for_benchmark(graph_obj, backend_wrapper)
+
+    if k > G.number_of_nodes():
+        pytest.skip(reason=f"{k=} > {G.number_of_nodes()=}")
+
     result = benchmark.pedantic(
         target=backend_wrapper(nx.edge_betweenness_centrality),
         args=(G,),
@@ -473,6 +480,26 @@ def bench_pagerank_personalized(benchmark, graph_obj, backend_wrapper):
     assert type(result) is dict
 
 
+def bench_shortest_path(benchmark, graph_obj, backend_wrapper):
+    """
+    This passes in the source node with the highest degree, but no target.
+    """
+    G = get_graph_obj_for_benchmark(graph_obj, backend_wrapper)
+    node = get_highest_degree_node(graph_obj)
+
+    result = benchmark.pedantic(
+        target=backend_wrapper(nx.shortest_path),
+        args=(G,),
+        kwargs=dict(
+            source=node,
+        ),
+        rounds=rounds,
+        iterations=iterations,
+        warmup_rounds=warmup_rounds,
+    )
+    assert type(result) is dict
+
+
 def bench_single_source_shortest_path_length(benchmark, graph_obj, backend_wrapper):
     G = get_graph_obj_for_benchmark(graph_obj, backend_wrapper)
     node = get_highest_degree_node(graph_obj)

benchmarks/nx-cugraph/pytest-based/create_results_summary_page.py

@@ -0,0 +1,245 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import re
+import pathlib
+import json
+import platform
+import psutil
+import socket
+import subprocess
+
+
+def compute_perf_vals(cugraph_runtime, networkx_runtime):
+    speedup_string = f"{networkx_runtime / cugraph_runtime:.3f}X"
+    delta = networkx_runtime - cugraph_runtime
+    if abs(delta) < 1:
+        if abs(delta) < 0.001:
+            units = "us"
+            delta *= 1e6
+        else:
+            units = "ms"
+            delta *= 1e3
+    else:
+        units = "s"
+    delta_string = f"{delta:.3f}{units}"
+
+    return (speedup_string, delta_string)
+
+
+def get_mem_info():
+    return round(psutil.virtual_memory().total / (1024**3), 2)
+
+
+def get_cuda_version():
+    output = subprocess.check_output("nvidia-smi", shell=True).decode()
+    try:
+        return next(
+            line.split("CUDA Version: ")[1].split()[0]
+            for line in output.splitlines()
+            if "CUDA Version" in line
+        )
+    except subprocess.CalledProcessError:
+        return "Failed to get CUDA version."
+
+
+def get_first_gpu_info():
+    try:
+        gpu_info = (
+            subprocess.check_output(
+                "nvidia-smi --query-gpu=name,memory.total,memory.free,memory.used --format=csv,noheader",
+                shell=True,
+            )
+            .decode()
+            .strip()
+        )
+        if gpu_info:
+            gpus = gpu_info.split("\n")
+            num_gpus = len(gpus)
+            first_gpu = gpus[0]  # Get the information for the first GPU
+            gpu_name, mem_total, _, _ = first_gpu.split(",")
+            return f"{num_gpus} x {gpu_name.strip()} ({round(int(mem_total.strip().split()[0]) / (1024), 2)} GB)"
+        else:
+            print("No GPU found or unable to query GPU details.")
+    except subprocess.CalledProcessError:
+        print("Failed to execute nvidia-smi. No GPU information available.")
+
+
+def get_system_info():
+    print(f"<p>Hostname: {socket.gethostname()}</p>")
+    print(
+        f'<p class="text-indent"">Operating System: {platform.system()} {platform.release()}</p>'
+    )
+    print(f'<p class="text-indent">Kernel Version  : {platform.version()}</p>')
+    with open("/proc/cpuinfo") as f:
+        print(
+            f'<p>CPU: {next(line.strip().split(": ")[1] for line in f if "model name" in line)} ({psutil.cpu_count(logical=False)} cores)</p>'
+        )
+    print(f'<p class="text-indent">Memory: {get_mem_info()} GB</p>')
+    print(f"<p>GPU: {get_first_gpu_info()}</p>")
+    print(f"<p>CUDA Version: {get_cuda_version()}</p>")
+
+
+if __name__ == "__main__":
+    logs_dir = pathlib.Path("logs")
+
+    dataset_patt = re.compile(".*ds=([\w-]+).*")
+    backend_patt = re.compile(".*backend=(\w+).*")
+    k_patt = re.compile(".*k=(10*).*")
+
+    # Organize all benchmark runs by the following hierarchy: algo -> backend -> dataset
+    benchmarks = {}
+
+    # Populate benchmarks dir from .json files
+    for json_file in logs_dir.glob("*.json"):
+        # print(f"READING {json_file}")
+        try:
+            data = json.loads(open(json_file).read())
+        except json.decoder.JSONDecodeError:
+            # print(f"PROBLEM READING {json_file}, skipping.")
+            continue
+
+        for benchmark_run in data["benchmarks"]:
+            # example name: "bench_triangles[ds=netscience-backend=cugraph-preconverted]"
+            name = benchmark_run["name"]
+
+            algo_name = name.split("[")[0]
+            if algo_name.startswith("bench_"):
+                algo_name = algo_name[6:]
+            # special case for betweenness_centrality
+            match = k_patt.match(name)
+            if match is not None:
+                algo_name += f", k={match.group(1)}"
+
+            match = dataset_patt.match(name)
+            if match is None:
+                raise RuntimeError(
+                    f"benchmark name {name} in file {json_file} has an unexpected format"
+                )
+            dataset = match.group(1)
+            if dataset.endswith("-backend"):
+                dataset = dataset[:-8]
+
+            match = backend_patt.match(name)
+            if match is None:
+                raise RuntimeError(
+                    f"benchmark name {name} in file {json_file} has an unexpected format"
+                )
+            backend = match.group(1)
+            if backend == "None":
+                backend = "networkx"
+
+            runtime = benchmark_run["stats"]["mean"]
+            benchmarks.setdefault(algo_name, {}).setdefault(backend, {})[
+                dataset
+            ] = runtime
+
+    # dump HTML table
+    ordered_datasets = [
+        "netscience",
+        "email_Eu_core",
+        "cit-patents",
+        "hollywood",
+        "soc-livejournal1",
+    ]
+
+    print(
+        """
+    <html>
+    <head>
+    <style>
+        table {
+            table-layout: fixed;
+            width: 100%;
+            border-collapse: collapse;
+        }
+        tbody tr:nth-child(odd) {
+            background-color: #ffffff;
+        }
+        tbody tr:nth-child(even) {
+            background-color: #d3d3d3;
+        }
+        tbody td {
+            text-align: center;
+            color: black;
+        }
+        th,
+        td {
+            padding: 10px;
+        }
+        .footer {
+            background-color: #f1f1f1;
+            padding: 10px;
+            font-size: 12px;
+            color: black;
+            width: 100%;
+        }
+        .text-indent {
+            text-indent: 20px; /* Indents the first line of the text by 30px */
+        }
+    </style>
+    </head>
+    <table>
+    <thead>
+    <tr>
+        <th></th>"""
+    )
+    for ds in ordered_datasets:
+        print(f"      <th>{ds}</th>")
+    print(
+        """   </tr>
+    </thead>
+    <tbody>
+    """
+    )
+    for algo_name in sorted(benchmarks):
+        algo_runs = benchmarks[algo_name]
+        print("   <tr>")
+        print(f"      <td>{algo_name}</td>")
+        # Proceed only if any results are present for both cugraph and NX
+        if "cugraph" in algo_runs and "networkx" in algo_runs:
+            cugraph_algo_runs = algo_runs["cugraph"]
+            networkx_algo_runs = algo_runs["networkx"]
+            datasets_in_both = set(cugraph_algo_runs).intersection(networkx_algo_runs)
+
+            # populate the table with speedup results for each dataset in the order
+            # specified in ordered_datasets. If results for a run using a dataset
+            # are not present for both cugraph and NX, output an empty cell.
+            for dataset in ordered_datasets:
+                if dataset in datasets_in_both:
+                    cugraph_runtime = cugraph_algo_runs[dataset]
+                    networkx_runtime = networkx_algo_runs[dataset]
+                    (speedup, runtime_delta) = compute_perf_vals(
+                        cugraph_runtime=cugraph_runtime,
+                        networkx_runtime=networkx_runtime,
+                    )
+                    print(f"      <td>{speedup}<br>{runtime_delta}</td>")
+                else:
+                    print(f"      <td></td>")
+
+        # If a comparison between cugraph and NX cannot be made, output empty cells
+        # for each dataset
+        else:
+            for _ in range(len(ordered_datasets)):
+                print("      <td></td>")
+
+        print("   </tr>")
+
+    print(
+        """
+    </tbody>\n</table>
+    <div class="footer">"""
+    )
+    get_system_info()
+    print("""</div>\n</html>""")

benchmarks/nx-cugraph/pytest-based/get_graph_bench_dataset.py

@@ -0,0 +1,35 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Checks if a particular dataset has been downloaded inside the datasets dir
+(RAPIDS_DATAEST_ROOT_DIR). If not, the file will be downloaded using the
+datasets API.
+
+Positional Arguments:
+    1) dataset name (e.g. 'email_Eu_core', 'cit-patents')
+       available datasets can be found here: `python/cugraph/cugraph/datasets/__init__.py`
+"""
+
+import sys
+
+import cugraph.datasets as cgds
+
+
+if __name__ == "__main__":
+    # download and store dataset (csv) by using the Datasets API
+    dataset = sys.argv[1].replace("-", "_")
+    dataset_obj = getattr(cgds, dataset)
+
+    if not dataset_obj.get_path().exists():
+        dataset_obj.get_edgelist(download=True)