Thread parallelism¶

Why thread parallelism is important¶

Because it is one way to speedup (pun intended) the calculation.

Goal¶

In this session, we are going to benchmark thread parallelism.

Benchmark script¶

benchmark_2d_integration.sh is the script that starts a benchmark, by submitting multiple jobs to the Slurm queue, using the Slurm script below.

The goal of the benchmark script is to do a fixed unit of work with increasingly more cores.

As the script itself only does light calculations, you can run it directly. Here is how to call the script:

bash benchmark_2d_integration.sh [account] [language]

For example:

bash benchmark_2d_integration.sh staff r

If you use the incorrect spelling, the script will help you.

Slurm script¶

This is the script that schedules a job with thread parallelism.

The goal of the script is to submit a calculation that uses thread parallelism, with a custom amount of cores.

This Slurm script is called by the benchmark script, i.e. not directly by a user. If the Slurm script is absent, the benchmark script will (try to) download it for you.

sbatch -A [account] -n [number_of_cores] do_[language]_2d_integration.sh

sbatch -A staff -n 1 do_r_2d_integration.sh

# On Dardel
sbatch -A staff -n 1 -p main do_r_2d_integration.sh

There are 3 Slurm scripts, 1 per language:

Each of these Slurm scripts are called by the benchmark script, where the benchmark script supplies the desired number of cores.

Language script¶

This is the code (in your favorite language) that performs a job with thread parallelism.

The goal of the language script is to have a fixed unit of work that can be done by a custom amount of cores.

This language script is called by the Slurm script, i.e. not directly by a user. If the calculation script is absent, the benchmark script will (try to) download it for you.

[interpreter] [script_name] [number_of_cores] [grid_size]

julia integration2d.jl 1 1
Rscript integration2d.R 1 1

Exercises¶

Exercise 1: start the benchmark on your HPC cluster¶

The goal of this exercise is to start the benchmark script on your HPC cluster, as well as some troubleshooting.

On your HPC cluster:

• Download the benchmark script

wget https://raw.githubusercontent.com/UPPMAX/R-matlab-julia-HPC/refs/heads/main/docs/advanced/thread_parallelism/benchmark_2d_integration.sh

• Run the benchmark script

bash benchmark_2d_integration.sh staff r

• Check the Slurm output files for problems. If there are problems: fix these, then run the benchmark script again

cat *.out

Exercise 2: read the benchmark script¶

Now that the benchmark script is running, we have the time to figure out what it is doing.

• What is the most important single line in this script, i.e. the line it is all about?

sbatch -A "${slurm_job_account}" -N "${n_nodes}" -n "${n_cores}" "${script_name}"

sbatch -A "${slurm_job_account}" -N "${n_nodes}" -n "${n_cores}" -p main "${script_name}"

• In English, describe what the line does in general terms

• This line of code is part of a for loop. In English, what does the for loop achieve?

Exercise 3: read the Slurm script¶

Exercise 4: read the calculation script¶

Exercise 5: analyse the results¶

grep -EoRh "^[jmlr].*,.*" --include=*.out | sort | uniq

You will see the collected results.

Exercise 6: compare to others¶

Exercise X1¶

What went wrong here? Why is this a problem?

[richel@pelle1 thread_parallelism]$ squeue --me
             JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
             54197     pelle do_r_2d_   richel  R       0:14      1 p66
             54200     pelle do_r_2d_   richel  R       0:14      4 p[64-67]
             54216     pelle do_r_2d_   richel  R       0:14      3 p[104-106]
             54217     pelle do_r_2d_   richel  R       0:14      6 p[106-111]
             54169     pelle do_r_2d_   richel  R       0:15      1 p70

Exercise X2¶

What went wrong here? Why is this a problem?

Julia single-thread tuns

Exercise X3: always program in Assembly?¶

Figure from paper

Where to go next?¶

Distributed parallelism

Troubleshooting¶

T1. Invalid account or account/partition combination specified¶

sbatch: error: Batch job submission failed: Invalid account or account/partition combination specified

You’ve specified the wrong account.

Run projinfo.

T2. There is no package called ‘doParallel’¶

This is an R error.

You can find it by checking the log files:

cat *.out

When you see, for example, the text below, it is clearly stated that there is no package called doParallel.

HPC cluster: tetralith
Slurm job account used: naiss2025-22-934
Number of cores booked in Slurm: 32
Error in library(doParallel, quietly = TRUE) : 
  there is no package called ‘doParallel’
Execution halted

To fix this:

• load the correct module
• install that package from the terminal.

To load the correct module, load the R module(s) as loaded by the do_r_2d_integration.sh script, for example:

module load R/4.4.0-hpc1-gcc-11.3.0-bare

if [ ${hpc_cluster} == "rackham" ]
then
  module load R_packages/4.1.1 >/dev/null 2>&1
fi

module load R_packages/4.1.1

To install that package from the terminal, check this course’s material on how to do so.

T3. ‘namespace ‘rlang’ 0.4.12 is already loaded, but >= 1.1.0 is required’¶

Error in loadNamespace(i, c(lib.loc, .libPaths()), versionCheck = vI[[i]]) : 
  namespace ‘rlang’ 0.4.12 is already loaded, but >= 1.1.0 is required
Calls: <Anonymous> ... waldo_compare -> loadNamespace -> namespaceImport -> loadNamespace
Execution halted

This only happens on Rackham, since 2025-09-25.

¶

Warning: Executing startup failed in matlabrc.
This indicates a potentially serious problem in your MATLAB setup, which should
be resolved as soon as possible.  Error detected was:
MATLAB:undefinedVarOrClass
Unable to resolve the name 'java.net.InetAddress.getLocalHost.getHostAddress'. 
Error using run
RUN cannot execute the file 'do_2d_integration.m 48'. RUN requires a valid
MATLAB script