Add weak scaling config generation script

2024-03-11 00:14:03 +01:00
parent d41d8b564b
commit 79227e93d9
4 changed files with 389 additions and 12 deletions
@@ -20,21 +20,33 @@ def print_table(data, spec):
 if __name__ == "__main__":
    p = argparse.ArgumentParser(description="Turn files generated by timing.py into pgf datafiles")
    p.add_argument("timing_file")
+    p.add_argument("--weak", action="store_true")
    args = p.parse_args()

    with open(args.timing_file, "r", encoding="utf8") as f:
        jobs = json.load(f)

-    scaling_spec = {
-        "label": lambda job: job["accounting"][0]["nodes"]["count"],
-        "nodes": lambda job: job["accounting"][0]["nodes"]["count"],
-        "tasks": lambda job: job["accounting"][0]["tasks"]["count"],
-        "mean_time": lambda job: job["means"]["TimeStep"],
-        "std_time": lambda job: job["stds"]["TimeStep"],
-        "speedup": lambda job: jobs[0]["means"]["TimeStep"] / job["means"]["TimeStep"],
-        # Standard deviation scaled to speedup
-        "speedup_std": lambda job: (jobs[0]["means"]["TimeStep"] / job["means"]["TimeStep"]) * (job["stds"]["TimeStep"] / job["means"]["TimeStep"]),
-        # 95% confidence interval
-        "speedup_error": lambda job: (jobs[0]["means"]["TimeStep"] / job["means"]["TimeStep"]) * (job["stds"]["TimeStep"] / job["means"]["TimeStep"]) / math.sqrt(len(jobs)) * 1.96,
-    }
+    if not args.weak:
+        scaling_spec = {
+            "label": lambda job: job["accounting"][0]["nodes"]["count"],
+            "nodes": lambda job: job["accounting"][0]["nodes"]["count"],
+            "tasks": lambda job: job["accounting"][0]["tasks"]["count"],
+            "mean_time": lambda job: job["means"]["TimeStep"],
+            "std_time": lambda job: job["stds"]["TimeStep"],
+            "speedup": lambda job: jobs[0]["means"]["TimeStep"] / job["means"]["TimeStep"],
+            # Standard deviation scaled to speedup
+            "speedup_std": lambda job: (jobs[0]["means"]["TimeStep"] / job["means"]["TimeStep"]) * (job["stds"]["TimeStep"] / job["means"]["TimeStep"]),
+            # 95% confidence interval
+            "speedup_error": lambda job: (jobs[0]["means"]["TimeStep"] / job["means"]["TimeStep"]) * (job["stds"]["TimeStep"] / job["means"]["TimeStep"]) / math.sqrt(len(jobs)) * 1.96,
+        }
+    else:
+        scaling_spec = {
+            "nodes": lambda job: job["accounting"][0]["nodes"]["count"],
+            "tasks": lambda job: job["accounting"][0]["tasks"]["count"],
+            "mean_time": lambda job: job["means"]["TimeStep"],
+            "std_time": lambda job: job["stds"]["TimeStep"],
+            "efficiency": lambda job: jobs[0]["means"]["TimeStep"] / job["means"]["TimeStep"],
+            "efficiency_error": lambda job: (jobs[0]["means"]["TimeStep"] / job["means"]["TimeStep"]) * (job["stds"]["TimeStep"] / job["means"]["TimeStep"]) / math.sqrt(len(jobs)) * 1.96,
+        }
+
    print_table(jobs, scaling_spec)
@@ -0,0 +1,152 @@
+#!/usr/bin/env python
+
+
+import jinja2
+import json
+import sys
+
+
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Tuple
+
+
+SIZE = (192, 192, 192)
+
+
+templates_env = jinja2.Environment(
+    loader=jinja2.FileSystemLoader(Path(__file__).parent.parent / "templates"),
+    autoescape=jinja2.select_autoescape()
+)
+        
+
+@dataclass
+class Experiment:
+    job_name: str
+    account: str
+    partition: str
+    nastja_binary_path: str
+    nodes: int
+    tasks: int
+    num_blocks: Tuple[int, int, int]
+    domain_scale: Tuple[int, int, int]
+    time: str = "00:15:00"
+    extra_sbatch_line: str = ""
+    logfile_path: str = "/p/project/cellsinsilico/paulslustigebude/ma/experiments/eval/logs/%x-%A.%a"
+    config_path: str = "/p/project/cellsinsilico/paulslustigebude/ma/experiments/eval/generated/config/${SLURM_JOB_NAME}.json"
+    output_dir_path: str = "/p/scratch/cellsinsilico/paul/nastja-out/${SLURM_JOB_NAME}-${SLURM_ARRAY_JOB_ID}.${SLURM_ARRAY_TASK_ID}"
+
+    def get_config(self):
+        with (Path(__file__).parent.parent / "templates" / "weak.json").open(encoding="utf8") as f:
+            config = json.load(f)
+
+        size = (
+            SIZE[0] * self.domain_scale[0],
+            SIZE[1] * self.domain_scale[1],
+            SIZE[2] * self.domain_scale[2],
+        )
+        blocksize = (
+            size[0] // self.num_blocks[0],
+            size[1] // self.num_blocks[1],
+            size[2] // self.num_blocks[2],
+        )
+        config["Geometry"] = {
+            "blockcount": list(self.num_blocks),
+            "blocksize": list(blocksize),
+        }
+
+        cells_filling = [{
+            "box": [
+                [0, 0, 0],
+                list(size)
+            ],
+            "celltype": 0,
+            "component": 0,
+            "pattern": "const",
+            "seed": 0,
+            "shape": "cube",
+            "value": 0,
+        }]
+        for z in range(self.domain_scale[2]):
+            for y in range(self.domain_scale[1]):
+                for x in range(self.domain_scale[0]):
+                    cx = x * SIZE[0] + SIZE[0] // 2
+                    cy = y * SIZE[1] + SIZE[1] // 2
+                    cz = z * SIZE[2] + SIZE[2] // 2
+                    cells_filling.append({
+                        "shape": "sphere",
+                        "pattern": "voronoi",
+                        "count": 715,
+                        "radius": 38,
+                        "center": [cx, cy, cz],
+                        "box": [
+                            [cx - 38, cy - 38, cz - 38],
+                            [cx + 38, cy + 38, cz + 38]
+                        ],
+                        "celltype": 9,
+                        "seed": 758960,
+                    })
+
+        config["Filling"]["cells"] = cells_filling
+
+        return config
+    
+    def write_batch_file(self, out_path: Path):
+        t = templates_env.get_template("strong-batch.j2")
+        t.stream(
+            name=self.job_name,
+            account=self.account,
+            partition=self.partition,
+            nodes=self.nodes,
+            tasks=self.tasks,
+            extra_sbatch_line=self.extra_sbatch_line,
+            time=self.time,
+            logfile_path=self.logfile_path,
+            nastja_binary_path=self.nastja_binary_path,
+            config_path=self.config_path,
+            output_dir_path=self.output_dir_path,
+        ).dump(str(out_path))
+
+
+def make_cpu_ex(x: int, y: int, z: int) -> Experiment:
+    num_blocks = x * y * z
+    assert num_blocks % 48 == 0
+    num_nodes = num_blocks // 48
+
+    assert x % 4 == 0
+    assert y % 4 == 0
+    assert z % 3 == 0
+
+    return Experiment(
+        job_name=f"weak-cpu-{x:02}-{y:02}-{z:02}",
+        account="cellsinsilico",
+        partition="batch",
+        nastja_binary_path="/p/project/cellsinsilico/paulslustigebude/nastja/build-nocuda/nastja",
+        nodes=num_nodes,
+        tasks=num_blocks,
+        num_blocks=(x, y, z),
+        domain_scale=(x // 4, y // 4, z // 3),
+    )
+
+
+experiments = [
+    make_cpu_ex(4, 4, 3),
+    make_cpu_ex(4, 4, 6),
+    make_cpu_ex(4, 4, 12),
+    make_cpu_ex(4, 8, 12),
+    make_cpu_ex(8, 8, 12),
+    make_cpu_ex(8, 8, 24),
+    make_cpu_ex(8, 16, 24),
+    make_cpu_ex(16, 16, 24),
+]
+
+if __name__ == "__main__":
+    outdir = Path(__file__).parent.parent / "generated"
+
+    for e in experiments:
+        print(f"Generating config for {e.job_name}", file=sys.stderr)
+        config_path = (outdir / "config" / e.job_name).with_suffix(".json")
+        with config_path.open("w", encoding="utf8") as f:
+            json.dump(e.get_config(), f, indent=2)
+        print(f"Generating batch file for {e.job_name}", file=sys.stderr)
+        e.write_batch_file(outdir / "batch" / e.job_name)
@@ -0,0 +1,20 @@
+#!/usr/bin/env python
+
+
+import argparse
+import pandas
+
+
+def show_seconds(s: float) -> str:
+    return f"{s:.2f}s"
+
+
+if __name__ == '__main__':
+    p = argparse.ArgumentParser(description="Make a latex table from a timings tsv")
+    p.add_argument("timingfile")
+    args = p.parse_args()
+
+    df = pandas.read_csv(args.timingfile, sep="\t")
+
+    for i in range(len(df)):
+        print(f"{df['nodes'][i]} & {df['tasks'][i]} & {show_seconds(df['mean_time'][i])} & {show_seconds(df['std_time'][i])} & {df['speedup'][i]:.02f} & {df['speedup_error'][i]:.02f} \\\\")