def single_gpu_benchmark():
scenarios = wlb.ScenarioManager()
block_sizes = [(i, i, i) for i in (64, 128, 256, 384)] + [(512, 512, 128)]
cuda_blocks = [(32, 1, 1), (64, 1, 1), (128, 1, 1), (256, 1, 1), (512, 1,
1),
(32, 2, 1), (64, 2, 1), (128, 2, 1), (256, 2, 1),
(32, 4, 1), (64, 4, 1), (128, 4, 1), (32, 8, 1), (64, 8, 1),
(32, 16, 1)]
for block_size in block_sizes:
for cuda_block_size in cuda_blocks:
cells = block_size[0] * block_size[1] * block_size[2]
time_steps_for_128_cubed = 1000
time_steps = (128**3 / cells) * time_steps_for_128_cubed
scenario = Scenario(cells_per_block=block_size,
gpuBlockSize=cuda_block_size,
timeStepStrategy='kernelOnly',
timesteps=int(time_steps))
scenarios.add(scenario)
def overlap_benchmark():
scenarios = wlb.ScenarioManager()
inner_outer_splits = [(1, 1, 1), (4, 1, 1), (8, 1, 1), (16, 1, 1),
(32, 1, 1), (4, 4, 1), (8, 8, 1), (16, 16, 1),
(32, 32, 1), (4, 4, 4), (8, 8, 8), (16, 16, 16),
(32, 32, 32)]
for comm_strategy in [
'UniformGPUScheme_Baseline', 'UniformGPUScheme_Memcpy'
]: # 'GPUPackInfo_Baseline', 'GPUPackInfo_Streams'
# no overlap
scenarios.add(
Scenario(timeStepStrategy='noOverlap',
communicationScheme=comm_strategy,
innerOuterSplit=(1, 1, 1)))
# overlap
for overlap_strategy in ['simpleOverlap', 'complexOverlap']:
for inner_outer_split in inner_outer_splits:
scenario = Scenario(timeStepStrategy=overlap_strategy,
communicationScheme=comm_strategy,
innerOuterSplit=inner_outer_split)
scenarios.add(scenario)
# encoding: utf-8
import itertools
import waLBerla as wlb
from base import get_block_decomposition, communication_schemes, overlap_communication, \
cuda_enabled_mpi, num_processes, calculate_time_steps, side_length_to_fill_memory
from benchmark import BenchmarkScenario, CommunicationSchemeType
# Stores the scenarios for the current simulation
scenarios = wlb.ScenarioManager()
# Generates all block decompositions of xyz, 2 directions at a time
#block_decompositions = itertools.combinations_with_replacement('xyz', r=2)
block_decompositions = ['xyz', 'yzx', 'zyx', 'yxz']
# compute number of cells depending on GPU memory i.e. by specifying the percentage of GPU memory to fill
gpu_memory_gb = 16
cells_per_block = [
side_length_to_fill_memory(pc, gpu_memory_gb) for pc in (0.8, 0.5, 0.05)
]
expected_mlups = 200 # to compute how many time steps have to be done
time_per_scenarios = 5 # benchmark time in seconds
fully_periodic = [False, True]
if num_processes == 1:
scenario_generator = itertools.product(communication_schemes, [
False,
], [
False,