saiph: simplify granularity and ss experiments

This commit is contained in:
Rodrigo Arias 2021-04-01 19:01:24 +02:00
parent 0e0f1b265f
commit 10b1ff8f7a
3 changed files with 100 additions and 70 deletions

View File

@ -32,8 +32,8 @@
}; };
saiph = { saiph = {
granularity-saiph = callPackage ./saiph/granularity-saiph.nix { }; granularity = callPackage ./saiph/granularity.nix { };
scalability-saiph = callPackage ./saiph/scalability-saiph.nix { }; ss = callPackage ./saiph/ss.nix { };
}; };
creams = rec { creams = rec {

View File

@ -1,4 +1,4 @@
###################################################################################### #######################################################################
# Saiph, granularity experiment: # Saiph, granularity experiment:
# #
# App:Heat 3D - garlic/tampi+isend+oss+task+simd branch # App:Heat 3D - garlic/tampi+isend+oss+task+simd branch
@ -15,12 +15,13 @@
# Granularity experiment configuration: # Granularity experiment configuration:
# Single-core run # Single-core run
# MPI binded to sockets: MPI procs = 2 # MPI binded to sockets: MPI procs = 2
# Mesh distributed across third dimension to ensure contiguous communications # Mesh distributed across third dimension to ensure contiguous
# communications
# --> nbgx = 1, nbgy = 1 # --> nbgx = 1, nbgy = 1
# First dimension cannot be locally blocked (simd reasons) # First dimension cannot be locally blocked (simd reasons)
# Second and third dimension local blocking limited by local mesh size # Second and third dimension local blocking limited by local mesh size
# #
###################################################################################### #######################################################################
# Common packages, tools and options # Common packages, tools and options
{ {
@ -29,43 +30,54 @@
, bsc , bsc
, targetMachine , targetMachine
, stages , stages
, garlicTools
}: }:
with stdenv.lib; with stdenv.lib;
with garlicTools;
let let
#*** Variable configurations *** # Variable configurations
varConf = with bsc; { varConf = with targetMachine.config; {
# Local blocks per dimension # Local blocks per dimension
nbl1 = [ 1 2 3 4 6 12 24 48 96 ]; nblx = [ 1 ]; # SIMD
nbl2 = [ 1 2 3 4 6 12 24 48 96 ]; nbly = range2 1 (hw.cpusPerNode * 8);
nblz = [ 8 ];
sizex = [ 3 ];
gitBranch = [ "garlic/tampi+isend+oss+task+simd" ];
}; };
#*** Generate the complete configuration for each unit *** # Generate the complete configuration for each unit
genConf = with bsc; c: targetMachine.config // rec { genConf = c: targetMachine.config // rec {
# Experiment, units and job names # Experiment, units and job names
expName = "saiph-granularity"; expName = "saiph-granularity";
unitName = "${expName}-N${toString nodes}" + "nbg_${toString nbgx}-${toString nbgy}-${toString nbgz}" + "nbl_1-${toString nbly}-${toString nblz}"; unitName = "${expName}"
jobName = "${unitName}"; + "-N${toString nodes}"
+ "-nbg.x${toString nbgx}.y${toString nbgy}.z${toString nbgz}"
+ "-nbl.x${toString nblx}.y${toString nbly}.z${toString nblz}";
jobName = unitName;
# saiph options # saiph options
totalTasks = ntasksPerNode * nodes;
nodes = 1; nodes = 1;
enableManualDist = true; # allows to manually set nbg{x-y-z} enableManualDist = true; # allows to manually set nbg{x-y-z}
nbgx = 1; nbgx = 1;
nbgy = 1; nbgy = 1;
nbgz = nodes*2; # forcing distribution by last dim nbgz = totalTasks; # forcing distribution by last dim
nblx = 1; # simd reasons
nbly = c.nbl1; # takes values from varConf
nblz = c.nbl2; # takes values from varConf
mpi = impi;
gitBranch = "garlic/tampi+isend+oss+task+simd";
gitCommit = "8052494d7dc62bef95ebaca9938e82fb029686f6"; # fix a specific commit
rev = "0";
# Repeat the execution of each unit 30 times inherit (c) nblx nbly nblz gitBranch sizex;
loops = 30;
blocksPerTask = nblx * nbly * nblz * 1.0;
blocksPerCpu = blocksPerTask / cpusPerTask;
# fix a specific commit
gitCommit = "8052494d7dc62bef95ebaca9938e82fb029686f6";
# Repeat the execution of each unit 10 times
loops = 10;
# Resources # Resources
inherit (targetMachine.config) hw; inherit (targetMachine.config) hw;
@ -75,35 +87,45 @@ let
}; };
#*** Compute the final set of configurations *** #*** Compute the final set of configurations ***
# Compute the array of configurations: cartesian product of all factors # Compute the array of configurations: cartesian product of all
# factors
allConfigs = stdexp.buildConfigs { allConfigs = stdexp.buildConfigs {
inherit varConf genConf; inherit varConf genConf;
}; };
# Filter to remove non-desired configurations: # Filter to remove non-desired configurations:
# --> tasks/proc < 0.5 # --> tasks/proc < 0.5
# --> nblz > 50 # --> nblz > 50
configs = filter (el: if ((builtins.mul el.nbly el.nblz) < (builtins.mul 0.5 el.cpusPerTask) || el.nblz > 50) then false else true) allConfigs; isGoodConfig = c:
let
maxNblz = c.cpusPerTask * 2;
in
! (c.blocksPerCpu < 0.5 || c.nblz > maxNblz);
configs = filter (isGoodConfig) allConfigs;
#*** Sets the env/argv of the program *** #*** Sets the env/argv of the program ***
exec = {nextStage, conf, ...}: with conf; stages.exec { exec = {nextStage, conf, ...}: stages.exec {
inherit nextStage; inherit nextStage;
env = '' env = ''
export OMP_NUM_THREADS=${toString hw.cpusPerSocket} export OMP_NUM_THREADS=${toString conf.cpusPerTask}
''; '';
}; };
#*** Configure the program according to the app *** #*** Configure the program according to the app ***
program = {nextStage, conf, ...}: program = {nextStage, conf, ...}: bsc.apps.saiph.override {
let inherit (conf) enableManualDist
customPkgs = stdexp.replaceMpi conf.mpi; nbgx nbgy nbgz nblx nbly nblz
in sizex
customPkgs.apps.saiph.override { gitBranch gitCommit;
inherit (conf) enableManualDist nbgx nbgy nbgz nblx nbly nblz mpi gitBranch gitCommit;
L3SizeKB = conf.hw.cacheSizeKB.L3;
cachelineBytes = conf.hw.cachelineBytes;
}; };
#*** Add stages to the pipeline *** #*** Add stages to the pipeline ***
pipeline = stdexp.stdPipeline ++ [ exec program ]; pipeline = stdexp.stdPipeline ++ [ exec program ];
in in
stdexp.genExperiment { inherit configs pipeline; }
stdexp.genExperiment { inherit configs pipeline; }

View File

@ -31,62 +31,68 @@
, bsc , bsc
, targetMachine , targetMachine
, stages , stages
, garlicTools
}: }:
with stdenv.lib; with stdenv.lib;
with garlicTools;
let let
#*** Variable configurations *** #*** Variable configurations ***
varConf = with bsc; { varConf = with targetMachine.config; {
# FIXME: None of those selected nbl* and problem size is able to give good
# efficiency when testing strong scaling. We should find better values.
# Local blocks per dimension # Local blocks per dimension
nbl1 = [ 1 2 3 4 6 12 24 48 96 ]; nblx = [ 1 ]; # SIMD
nbl2 = [ 1 2 3 4 6 12 24 48 96 ]; nbly = [ 32 ];
# Number of nodes nblz = [ 8 ];
nodes = [ 1 2 4 8 ]; sizex = [ 3 6 ];
gitBranch = [ "garlic/tampi+isend+oss+task+simd" ];
nodes = range2 1 8;
}; };
#*** Generate the complete configuration for each unit *** #*** Generate the complete configuration for each unit ***
genConf = with bsc; c: targetMachine.config // rec { genConf = c: targetMachine.config // rec {
# Experiment, units and job names # Experiment, units and job names
expName = "saiph-scalability"; expName = "saiph-ss";
unitName = "${expName}-N${toString nodes}" + "nbg_${toString nbgx}-${toString nbgy}-${toString nbgz}" + "nbl_1-${toString nbly}-${toString nblz}"; unitName = "${expName}"
jobName = "${unitName}"; + "-N${toString nodes}"
+ "-nbg.x${toString nbgx}.y${toString nbgy}.z${toString nbgz}"
+ "-nbl.x${toString nblx}.y${toString nbly}.z${toString nblz}";
jobName = unitName;
# saiph options # saiph options
nodes = c.nodes; # takes values from varConf
enableManualDist = true; # allows to manually set nbg{x-y-z} enableManualDist = true; # allows to manually set nbg{x-y-z}
nbgx = 1; nbgx = 1;
nbgy = 1; nbgy = 1;
nbgz = nodes*2; # forcing distribution by last dim nbgz = nodes * ntasksPerNode; # forcing distribution by last dim
nblx = 1; # simd reasons
nbly = c.nbl1; # takes values from varConf inherit (c) nblx nbly nblz nodes sizex;
nblz = c.nbl2; # takes values from varConf
mpi = impi;
gitBranch = "garlic/tampi+isend+oss+task+simd"; gitBranch = "garlic/tampi+isend+oss+task+simd";
gitCommit = "8052494d7dc62bef95ebaca9938e82fb029686f6"; # fix a specific commit gitCommit = "8052494d7dc62bef95ebaca9938e82fb029686f6"; # fix a specific commit
rev = "0";
# Repeat the execution of each unit 30 times blocksPerTask = nblx * nbly * nblz * 1.0;
loops = 30; blocksPerCpu = blocksPerTask / cpusPerTask;
# Repeat the execution of each unit 10 times
loops = 10;
# Resources # Resources
inherit (targetMachine.config) hw; inherit (targetMachine.config) hw;
qos = "bsc_cs";
qos = "debug";
ntasksPerNode = hw.socketsPerNode; # MPI binded to sockets ntasksPerNode = hw.socketsPerNode; # MPI binded to sockets
cpusPerTask = hw.cpusPerSocket; # Using the 24 CPUs of each socket cpusPerTask = hw.cpusPerSocket; # Using the 24 CPUs of each socket
}; };
#*** Compute the final set of configurations *** #*** Compute the final set of configurations ***
# Compute the array of configurations: cartesian product of all factors # Compute the array of configurations: cartesian product of all factors
allConfigs = stdexp.buildConfigs { configs = stdexp.buildConfigs {
inherit varConf genConf; inherit varConf genConf;
}; };
# Filter to remove non-desired configurations:
# --> tasks/proc < 3
# --> nblz > 25
configs = filter (el: if ((builtins.mul el.nbly el.nblz) < (builtins.mul 3 el.cpusPerTask) || el.nblz > 25) then false else true) allConfigs;
#*** Sets the env/argv of the program *** #*** Sets the env/argv of the program ***
exec = {nextStage, conf, ...}: with conf; stages.exec { exec = {nextStage, conf, ...}: with conf; stages.exec {
@ -97,17 +103,19 @@ let
}; };
#*** Configure the program according to the app *** #*** Configure the program according to the app ***
program = {nextStage, conf, ...}: program = {nextStage, conf, ...}: bsc.apps.saiph.override {
let inherit (conf) enableManualDist
customPkgs = stdexp.replaceMpi conf.mpi; nbgx nbgy nbgz nblx nbly nblz
in sizex
customPkgs.apps.saiph.override { gitBranch gitCommit;
inherit (conf) enableManualDist nbgx nbgy nbgz nblx nbly nblz mpi gitBranch gitCommit;
L3SizeKB = conf.hw.cacheSizeKB.L3;
cachelineBytes = conf.hw.cachelineBytes;
}; };
#*** Add stages to the pipeline *** #*** Add stages to the pipeline ***
pipeline = stdexp.stdPipeline ++ [ exec program ]; pipeline = stdexp.stdPipeline ++ [ exec program ];
in in
stdexp.genExperiment { inherit configs pipeline; }
stdexp.genExperiment { inherit configs pipeline; }