Last active
April 29, 2024 14:39
-
-
Save glwagner/1c271e80ceba48b7325207716a496af3 to your computer and use it in GitHub Desktop.
The comptuational cost of ocean large eddy simulation on CPU and GPU
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
using Oceananigans.Units | |
using GLMakie | |
# Wall time per time step in a large eddy simulation with 256^3 resolution. | |
# | |
# Numbers provided by Qing Li (HKUST Guangzhou, qingli411.github.io) | |
# | |
# Details for the "LESMIP" contained in | |
# https://github.com/qingli411/A2022_LESMIP/blob/main/notebook/Compare_LES-BF5hWD05WV00.ipynb | |
# | |
# "_a5000" refers to the Nvidia RTX A5000. | |
# "_128" refers to a simulation on 128 Intel Xeon Gold 6348 CPUs | |
oceananigans_a100 = 1.074hour / 16650 | |
palm_a5000 = 13.37hour / 14238 # note this is not used in the plots below | |
ncarles_128 = 3.47hour / 28880 | |
# Cost estimates for other computational configurations | |
ϵ_palm = 0.83 # estimated cost of PALM vs NCAR LES on CPU | |
ncarles_256 = ncarles_128 / 2 | |
ncarles_512 = ncarles_256 / 2 | |
palm_128 = ϵ_palm * ncarles_128 | |
palm_256 = ϵ_palm * ncarles_256 | |
palm_512 = ϵ_palm * ncarles_512 | |
oceananigans_4a100 = oceananigans_a100 / 4 | |
data = [ncarles_128, | |
ncarles_256, | |
ncarles_512, | |
palm_128, | |
palm_256, | |
palm_512, | |
oceananigans_a100, | |
oceananigans_4a100] | |
colors = Makie.wong_colors() | |
x = [1, 1, 1, 2, 2, 2, 3, 3] | |
dodge = [1, 2, 3, 1, 2, 3, 1, 2] | |
color_id = [1, 2, 3, 1, 2, 3, 4, 5] | |
color = [colors[id] for id in color_id] | |
xticks = (1:3, ["NCAR LES", "PALM", "Oceananigans"]) | |
fig = Figure(size=(800, 300)) | |
ax = Axis(fig[1, 1]; xticks, ylabel = "Wall time per time-step (s)") | |
barplot!(ax, x, data; dodge, color) | |
labels = ["128 CPUs (2 nodes at 64 CPU/node)", | |
"256 CPUs (4 nodes, est)", | |
"512 CPUs (8 nodes, est)", | |
"1 A100 GPU", | |
"4 A100 GPUs (1 node at 4 GPU/node, est)"] | |
elements = [PolyElement(polycolor = colors[i]) for i in 1:length(labels)] | |
title = "Computational cost" | |
Legend(fig[1, 2], elements, labels, title) | |
Label(fig[0, 1:2], "Computational cost of large eddy simulation with 256³ resolution") | |
# Single node estimate | |
ncarles_64 = 2ncarles_128 | |
relative_cost_1node = ncarles_64 / oceananigans_4a100 # ~147x | |
@info "Oceananigans is $relative_cost_1node times faster than NCAR LES" | |
save("cost_analysis.png", fig) | |
display(fig) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
This script produces this plot: