From c14b948595dedb4653a934a7f0b554846627c872 Mon Sep 17 00:00:00 2001
From: Michael Zikeli <michael.zikeli@fau.de>
Date: Mon, 10 Feb 2025 11:18:34 +0100
Subject: [PATCH] Refactor benchmark scenario creation to support multiple
inner_outer_split configurations for weak and strong scaling benchmarks.
---
.../simulation_setup/benchmark_configs_RDM.py | 45 ++++++-----
.../simulation_setup/benchmark_configs_RDM.py | 78 ++++++++++---------
2 files changed, 66 insertions(+), 57 deletions(-)
diff --git a/apps/benchmarks/UniformGridCPU/simulation_setup/benchmark_configs_RDM.py b/apps/benchmarks/UniformGridCPU/simulation_setup/benchmark_configs_RDM.py
index 071d1b4e0..251f5a28f 100644
--- a/apps/benchmarks/UniformGridCPU/simulation_setup/benchmark_configs_RDM.py
+++ b/apps/benchmarks/UniformGridCPU/simulation_setup/benchmark_configs_RDM.py
@@ -170,13 +170,14 @@ def weak_scaling_benchmark():
scenarios = wlb.ScenarioManager()
for t in ["simpleOverlap"]:
- scenarios.add(Scenario(benchmark_name="weakScalingUniformGrid",
- time_step_strategy=t,
- inner_outer_split=(1, 1, 1),
- cells_per_block=(WeakX, WeakY, WeakZ),
- boundary_setup=True,
- outer_iterations=1,
- db_file_name="weakScalingUniformGridOneBlock.sqlite3"))
+ for split in [(1, 1, 1), (16, 1, 1)]:
+ scenarios.add(Scenario(benchmark_name="weakScalingUniformGrid",
+ time_step_strategy=t,
+ inner_outer_split=split,
+ cells_per_block=(WeakX, WeakY, WeakZ),
+ boundary_setup=True,
+ outer_iterations=1,
+ db_file_name="weakScalingUniformGridOneBlock.sqlite3"))
def strong_scaling_benchmark():
@@ -190,13 +191,15 @@ def strong_scaling_benchmark():
cells_per_block = tuple([d // b for d, b in zip(domain_size, reversed(blocks))])
for t in ["simpleOverlap"]:
- scenarios.add(Scenario(benchmark_name="strongScalingUniformGridOneBlock",
- cells_per_block=cells_per_block,
- time_step_strategy=t,
- outer_iterations=1,
- timesteps=num_time_steps(cells_per_block),
- boundary_setup=True,
- db_file_name="strongScalingUniformGridOneBlock.sqlite3"))
+ for split in [(1, 1, 1), (16, 1, 1)]:
+ scenarios.add(Scenario(benchmark_name="strongScalingUniformGridOneBlock",
+ cells_per_block=cells_per_block,
+ time_step_strategy=t,
+ inner_outer_split=split,
+ outer_iterations=1,
+ timesteps=num_time_steps(cells_per_block),
+ boundary_setup=True,
+ db_file_name="strongScalingUniformGridOneBlock.sqlite3"))
def single_node_benchmark():
@@ -205,12 +208,14 @@ def single_node_benchmark():
wlb.log_info_on_root("")
scenarios = wlb.ScenarioManager()
- scenario = Scenario(benchmark_name="singleNodeUniformGridOneBlock",
- cells_per_block=(128, 128, 128),
- time_step_strategy='kernelOnly',
- outer_iterations=1,
- timesteps=10)
- scenarios.add(scenario)
+ for split in [(1, 1, 1), (16, 1, 1)]:
+ scenario = Scenario(benchmark_name="singleNodeUniformGridOneBlock",
+ cells_per_block=(128, 128, 128),
+ time_step_strategy='kernelOnly',
+ inner_outer_split=split,
+ outer_iterations=1,
+ timesteps=10)
+ scenarios.add(scenario)
def validation_run():
diff --git a/apps/benchmarks/UniformGridGPU/simulation_setup/benchmark_configs_RDM.py b/apps/benchmarks/UniformGridGPU/simulation_setup/benchmark_configs_RDM.py
index 74269c0bc..bb9da2611 100755
--- a/apps/benchmarks/UniformGridGPU/simulation_setup/benchmark_configs_RDM.py
+++ b/apps/benchmarks/UniformGridGPU/simulation_setup/benchmark_configs_RDM.py
@@ -213,15 +213,16 @@ def weak_scaling_overlap(cuda_enabled_mpi=False):
# overlap
for t in ["simpleOverlap"]:
- scenarios.add(Scenario(benchmark_name="weakScalingUniformGrid",
- cells_per_block=(WeakX, WeakY, WeakZ),
- cuda_blocks=(128, 1, 1),
- time_step_strategy=t,
- inner_outer_split=(8, 8, 8),
- cuda_enabled_mpi=cuda_enabled_mpi,
- outer_iterations=1,
- boundary_setup=True,
- db_file_name="weakScalingUniformGrid.sqlite3"))
+ for split in [(1, 1, 1), (8, 8, 8), (16, 1, 1)]:
+ scenarios.add(Scenario(benchmark_name="weakScalingUniformGrid",
+ cells_per_block=(WeakX, WeakY, WeakZ),
+ cuda_blocks=(128, 1, 1),
+ time_step_strategy=t,
+ inner_outer_split=split,
+ cuda_enabled_mpi=cuda_enabled_mpi,
+ outer_iterations=1,
+ boundary_setup=True,
+ db_file_name="weakScalingUniformGrid.sqlite3"))
def strong_scaling_overlap(cuda_enabled_mpi=False):
@@ -236,17 +237,18 @@ def strong_scaling_overlap(cuda_enabled_mpi=False):
# overlap
for t in ["simpleOverlap"]:
- scenarios.add(Scenario(benchmark_name="strongScalingUniformGridOneBlock",
- cells_per_block=cells_per_block,
- cuda_blocks=(128, 1, 1),
- time_step_strategy=t,
- inner_outer_split=(1, 1, 1),
- cuda_enabled_mpi=cuda_enabled_mpi,
- outer_iterations=1,
- timesteps=num_time_steps(cells_per_block),
- blocks=blocks,
- boundary_setup=True,
- db_file_name="strongScalingUniformGridOneBlock.sqlite3"))
+ for split in [(1, 1, 1), (8, 8, 8), (16, 1, 1)]:
+ scenarios.add(Scenario(benchmark_name="strongScalingUniformGridOneBlock",
+ cells_per_block=cells_per_block,
+ cuda_blocks=(128, 1, 1),
+ time_step_strategy=t,
+ inner_outer_split=split,
+ cuda_enabled_mpi=cuda_enabled_mpi,
+ outer_iterations=1,
+ timesteps=num_time_steps(cells_per_block),
+ blocks=blocks,
+ boundary_setup=True,
+ db_file_name="strongScalingUniformGridOneBlock.sqlite3"))
def single_gpu_benchmark():
@@ -267,23 +269,25 @@ def single_gpu_benchmark():
cuda_blocks = [(128, 1, 1), ]
for block_size in block_sizes:
for cuda_block_size in cuda_blocks:
- # cuda_block_size = (256, 1, 1) and block_size = (64, 64, 64) would be cut to cuda_block_size = (64, 1, 1)
- if cuda_block_size > block_size:
- continue
- if not cuda_block_size_ok(cuda_block_size):
- wlb.log_info_on_root(f"Cuda block size {cuda_block_size} would exceed register limit. Skipping.")
- continue
- if not domain_block_size_ok(block_size, gpu_mem):
- wlb.log_info_on_root(f"Block size {block_size} would exceed GPU memory. Skipping.")
- continue
- scenario = Scenario(benchmark_name="singleNodeUniformGridOneBlock",
- cells_per_block=block_size,
- cuda_blocks=cuda_block_size,
- time_step_strategy='kernelOnly',
- timesteps=num_time_steps(block_size, 2000),
- outer_iterations=1,
- additional_info=additional_info)
- scenarios.add(scenario)
+ for split in [(1, 1, 1), (8, 8, 8), (16, 1, 1)]:
+ # cuda_block_size = (256, 1, 1) and block_size = (64, 64, 64) would be cut to cuda_block_size = (64, 1, 1)
+ if cuda_block_size > block_size:
+ continue
+ if not cuda_block_size_ok(cuda_block_size):
+ wlb.log_info_on_root(f"Cuda block size {cuda_block_size} would exceed register limit. Skipping.")
+ continue
+ if not domain_block_size_ok(block_size, gpu_mem):
+ wlb.log_info_on_root(f"Block size {block_size} would exceed GPU memory. Skipping.")
+ continue
+ scenario = Scenario(benchmark_name="singleNodeUniformGridOneBlock",
+ cells_per_block=block_size,
+ cuda_blocks=cuda_block_size,
+ time_step_strategy='kernelOnly',
+ inner_outer_split=split,
+ timesteps=num_time_steps(block_size, 2000),
+ outer_iterations=1,
+ additional_info=additional_info)
+ scenarios.add(scenario)
def validation_run():
--
GitLab