hpcg: add weakscaling over some nblocks to check which axis is better

This commit is contained in:
Raúl Peñacoba 2021-03-26 16:20:04 +01:00 committed by Rodrigo Arias Mallo
parent 1a6075a2b1
commit b60a46b683
5 changed files with 240 additions and 9 deletions

View File

@ -0,0 +1,104 @@
{
stdenv
, stdexp
, bsc
, targetMachine
, stages
, genInput
}:
with stdenv.lib;
let
# Initial variable configuration
varConf = {
n = [ { x = 192; y = 192; z = 192; } ];
nprocs = [
{ x = 2; y = 1; z = 1; }
{ x = 4; y = 1; z = 1; }
{ x = 8; y = 1; z = 1; }
{ x = 16; y = 1; z = 1; }
{ x = 32; y = 1; z = 1; }
{ x = 1; y = 2; z = 1; }
{ x = 1; y = 4; z = 1; }
{ x = 1; y = 8; z = 1; }
{ x = 1; y = 16; z = 1; }
{ x = 1; y = 32; z = 1; }
{ x = 1; y = 1; z = 2; }
{ x = 1; y = 1; z = 4; }
{ x = 1; y = 1; z = 8; }
{ x = 1; y = 1; z = 16; }
{ x = 1; y = 1; z = 32; }
];
# nblocks = [ 12 24 48 96 192 384 768 1536 ];
nblocks = [ 384 768 1536 ];
ncommblocks = [ 1 ];
# nodes = [ 1 ];
# nodes = [ 1 2 4 8 16 ];
};
# Generate the complete configuration for each unit
genConf = c: targetMachine.config // rec {
expName = "hpcg.oss";
unitName = "${expName}.nb${toString nblocks}";
inherit (targetMachine.config) hw;
# hpcg options
inherit (c) n nprocs nblocks ncommblocks;
gitBranch = "garlic/tampi+isend+oss+task";
# Repeat the execution of each unit 30 times
loops = 10;
disableAspectRatio = true;
# Resources
qos = "debug";
ntasksPerNode = hw.socketsPerNode;
time = "02:00:00";
# task in one socket
cpusPerTask = hw.cpusPerSocket;
nodes = (nprocs.x * nprocs.y * nprocs.z) / ntasksPerNode;
jobName = "hpcg-${toString n.x}-${toString n.y}-${toString n.z}-${gitBranch}";
};
# Compute the array of configurations
configs = stdexp.buildConfigs {
inherit varConf genConf;
};
input = genInput configs;
exec = {nextStage, conf, ...}: stages.exec {
inherit nextStage;
argv = [
"--nx=${toString conf.n.x}"
"--ny=${toString conf.n.y}"
"--nz=${toString conf.n.z}"
# Distribute all processes in X axis
"--npx=${toString conf.nprocs.x}"
"--npy=${toString conf.nprocs.y}"
"--npz=${toString conf.nprocs.z}"
"--nblocks=${toString conf.nblocks}"
"--ncomms=${toString conf.ncommblocks}"
# The input symlink is generated by the input stage, which is generated by
# the genInput function.
"--load=input"
# Disable HPCG Aspect Ratio to run any mpi layout
] ++ optional (conf.disableAspectRatio) "--no-ar=1";
};
program = {nextStage, conf, ...}: bsc.apps.hpcg.override {
inherit (conf) gitBranch;
};
pipeline = stdexp.stdPipeline ++ [ input exec program ];
in
stdexp.genExperiment { inherit configs pipeline; }

View File

@ -19,9 +19,22 @@ let
{ x = 8; y = 1; z = 1; }
{ x = 16; y = 1; z = 1; }
{ x = 32; y = 1; z = 1; }
{ x = 1; y = 2; z = 1; }
{ x = 1; y = 4; z = 1; }
{ x = 1; y = 8; z = 1; }
{ x = 1; y = 16; z = 1; }
{ x = 1; y = 32; z = 1; }
{ x = 1; y = 1; z = 2; }
{ x = 1; y = 1; z = 4; }
{ x = 1; y = 1; z = 8; }
{ x = 1; y = 1; z = 16; }
{ x = 1; y = 1; z = 32; }
];
# nblocks = [ 12 24 48 96 192 384 768 1536 ];
nblocks = [ 384 ];
nblocks = [ 384 768 1536 ];
ncommblocks = [ 1 ];
# nodes = [ 1 ];
# nodes = [ 1 2 4 8 16 ];
@ -40,7 +53,7 @@ let
gitBranch = "garlic/tampi+isend+oss+task";
# Repeat the execution of each unit 30 times
loops = 3;
loops = 10;
disableAspectRatio = true;

View File

@ -65,13 +65,13 @@
inherit genInput;
};
ossScalability = callPackage ./hpcg/oss.scalability.192.nix {
inherit genInput;
};
# slices = callPackage ./hpcg/slices.nix {
# ossScalability = callPackage ./hpcg/oss.scalability.192.nix {
# inherit genInput;
# };
ossSlicesWeakscaling = callPackage ./hpcg/oss.slices.weakscaling.nix {
inherit genInput;
};
};
heat = rec {

View File

@ -0,0 +1,110 @@
# This R program takes as argument the dataset that contains the results of the
# execution of the heat example experiment and produces some plots. All the
# knowledge to understand how this script works is covered by this nice R book:
#
# Winston Chang, R Graphics Cookbook: Practical Recipes for Visualizing Data,
# OReilly Media (2020). 2nd edition
#
# Which can be freely read it online here: https://r-graphics.org/
#
# Please, search in this book before copying some random (and probably oudated)
# reply on stack overflow.
# We load some R packages to import the required functions. We mainly use the
# tidyverse packages, which are very good for ploting data,
library(ggplot2)
library(dplyr, warn.conflicts = FALSE)
library(scales)
library(jsonlite)
library(viridis, warn.conflicts = FALSE)
# Here we simply load the arguments to find the input dataset. If nothing is
# specified we use the file named `input` in the current directory.
# We can run this script directly using:
# Rscript <path-to-this-script> <input-dataset>
# Load the arguments (argv)
args = commandArgs(trailingOnly=TRUE)
# Set the input dataset if given in argv[1], or use "input" as default
if (length(args)>0) { input_file = args[1] } else { input_file = "input" }
df = jsonlite::stream_in(file(input_file), verbose=FALSE) %>%
# Then we flatten it, as it may contain dictionaries inside the columns
jsonlite::flatten() %>%
# Now the dataframe contains all the configuration of the units inside the
# columns named `config.*`, for example `config.cbs`. We first select only
# the columns that we need:
select(config.nblocks,
config.ncommblocks,
config.hw.cpusPerSocket,
config.nodes,
config.nprocs.x,
config.nprocs.y,
config.nprocs.z,
unit,
time
) %>%
# And then we rename those columns to something shorter:
rename(nblocks=config.nblocks,
ncommblocks=config.ncommblocks,
cpusPerSocket=config.hw.cpusPerSocket,
nodes=config.nodes,
npx=config.nprocs.x,
npy=config.nprocs.y,
npz=config.nprocs.z
) %>%
mutate(axisColor=as.factor(ifelse(npx != 1, "X", ifelse(npy != 1, "Y", "Z")))) %>%
mutate(blocksPerCpu = nblocks / cpusPerSocket) %>%
mutate(nblocks = as.factor(nblocks)) %>%
mutate(blocksPerCpu = as.factor(blocksPerCpu)) %>%
mutate(nodes = as.factor(nodes)) %>%
mutate(unit = as.factor(unit)) %>%
group_by(unit) %>%
# And compute some metrics which are applied to each group. For example we
# compute the median time within the runs of a unit:
mutate(median.time = median(time)) %>%
mutate(normalized.time = time / median.time - 1) %>%
mutate(log.median.time = log(median.time)) %>%
# Then, we remove the grouping. This step is very important, otherwise the
# plotting functions get confused:
ungroup()
dpi=300
h=5
w=5
w=3*w
# We plot the time of each run as we vary the block size
p = ggplot(df, aes(x=blocksPerCpu, y=time, color=axisColor)) +
# We add a points (scatter plot) using circles (shape=21) a bit larger
# than the default (size=3)
geom_point(shape=21, size=3) +
facet_wrap(~ nodes, labeller="label_both") +
# The bw theme is recommended for publications
theme_bw() +
# Here we add the title and the labels of the axes
labs(x="Blocks Per CPU", y="Time (s)", title="HPCG weak scalability: time",
color="Axis",
subtitle=input_file) +
# And set the subtitle font size a bit smaller, so it fits nicely
theme(plot.subtitle=element_text(size=8))
# Then, we save the plot both in png and pdf
ggsave("time.png", plot=p, width=w, height=h, dpi=dpi)
ggsave("time.pdf", plot=p, width=w, height=h, dpi=dpi)

View File

@ -38,9 +38,13 @@ in
};
hpcg = with exp.hpcg; {
# /nix/store/8dr191vch1nw7vfz8nj36d5nhwnbdnf3-plot
ossGranularity = stdPlot ./hpcg/oss.granularity.R [ ossGranularity ];
ossScalability = stdPlot ./hpcg/oss.scalability.R [ ossScalability ];
# slices = stdPlot ./hpcg/oss.R [ slices ];
# ossScalability = stdPlot ./hpcg/oss.scalability.R [ ossScalability ];
# /nix/store/a3x76fbnfbacn2xhz3q65fklfp0qbb6p-plot
ossWeakscalingPerAxisPerBlock = stdPlot ./hpcg/oss.slices.weakscaling.R [ ossSlicesWeakscaling ];
};
saiph = with exp.saiph; {