Writing good code

Overview

Teaching: 20 min
Exercises: 10 min

Questions

How do we write a good job script.

Objectives

Write a script that can be run serial or parallel.

Write a script that using SLURM environment variables.

Understand the limitations of random number generation.

When talking about ‘a script’ we could be referring to multiple things.

Slurm/Bash script - Almost everyone will be using one of these to submit their Slurm jobs.
Work script - If your work involves running another script (usually in a language other than Bash like Python, R or MATLAB) that will have to be invoked in your bash script.

This section will cover best practice for both types of script.

Use environment variables

In this lesson we will take a look at a few of the things to watch out for when writing scripts for use on the cluster. This will be most relevant to people writing their own code, but covers general practices applicable to everyone.

There is a lot of useful information contained within environment variable.

Slurm Environment

For a small demo of the sort of useful info contained within env variables, run the command.

sbatch --output "slurm_env.out" --wrap "env | grep"

once the job has finished check the results with,

cat slurm_env.out

SLURM_JOB_START_TIME=1695513911
SLURM_NODELIST=wbn098
SLURM_JOB_NAME=wrap
SLURMD_NODENAME=wbn098
SLURM_TOPOLOGY_ADDR=top.s13.s7.wbn098
SLURM_PRIO_PROCESS=0
SLURM_NODE_ALIASES=(null)
SLURM_JOB_QOS=staff
SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.switch.node
SLURM_JOB_END_TIME=1695514811
SLURM_MEM_PER_CPU=512
SLURM_NNODES=1
SLURM_JOBID=39572365
SLURM_TASKS_PER_NODE=2
SLURM_WORKING_CLUSTER=mahuika:hpcwslurmctrl01:6817:9984:109
SLURM_CONF=/etc/opt/slurm/slurm.conf
SLURM_JOB_ID=39572365
SLURM_JOB_USER=cwal219
__LMOD_STACK_SLURM_I_MPI_PMI_LIBRARY=L29wdC9zbHVybS9saWI2NC9saWJwbWkyLnNv
SLURM_JOB_UID=201333
SLURM_NODEID=0
SLURM_SUBMIT_DIR=/scale_wlg_persistent/filesets/home/cwal219
SLURM_TASK_PID=8747
SLURM_CPUS_ON_NODE=2
SLURM_PROCID=0
SLURM_JOB_NODELIST=wbn098
SLURM_LOCALID=0
SLURM_JOB_GID=201333
SLURM_JOB_CPUS_PER_NODE=2
SLURM_CLUSTER_NAME=mahuika
SLURM_GTIDS=0
SLURM_SUBMIT_HOST=wbn003
SLURM_JOB_PARTITION=large
SLURM_JOB_ACCOUNT=nesi99999
SLURM_JOB_NUM_NODES=1
SLURM_SCRIPT_CONTEXT=prolog_task

Can you think of some examples as to how these variables could be used in your script?

Solution

SLURM_JOB_CPUS_PER_NODE could be used to pass CPU numbers directly to any programs being used.

Some other things.

Variables in Slurm Header

Environment variables set by Slurm cannot be referenced in the Slurm header.

Default values

It is good practice to set default values when using environment variables when there is a chance they will be run in an environment where they may not be present.

FOO="${VARIABLE:-default}"

FOO will be to to the value of VARIABLE if is set, otherwise it will be set to default.

As a slight variation on the above example. (:= as opposed to :-).

FOO="${VARIABLE:=default}"

FOO will be to to the value of VARIABLE if is set, otherwise it will be set to default, VARIABLE will also be set to default.

num_cpus <- 2

The number of CPU’s being used is fixed in the script. We can save time and reduce chances for making mistakes by replacing this static value with an environment variable. We can use the environment variable SLURM_CPUS_PER_TASK.

num_cpus <- strtoi(Sys.getenv('SLURM_CPUS_PER_TASK')) 

Slurm sets many environment variables when starting a job, see Slurm Documentation for the full list.

The problem with this approach however, is our code will throw an error if we run it on the login node, or on our local machine or anywhere else that SLURM_CPUS_PER_TASK is not set.

Generally it is best not to diverge your codebase especially if you don’t have it under version control, so lets add some compatibility for those use cases.

num_cpus <- strtoi(Sys.getenv('SLURM_CPUS_PER_TASK', unset = "1")) 

Now if SLURM_CPUS_PER_TASK variable is not set, 1 CPU will be used. You could also use some other method of detecting CPUs, like detectCores().

Interoperability

windows + mac + linux headless + interactive

Verbose

Having a printout of job progress is fine for an interactive terminal, but when you aren’t seeing the updates in real time anyway, it’s just bloat for your output files.

Let’s add an option to mute the updates.

print_progress <- FALSE

if (print_progress && percent_complete%%1==0){

Reproduceability

As this script uses Pseudorandom number generation there are a few additional factors to consider. It is desirable that our output be reproducible so we can confirm that changes to the code have not affected it.

We can do this by setting the seed of the PRNG. That way we will get the same progression of ‘random’ numbers.

We are using the environment variable SLURM_ARRAY_TASK_ID for reasons we will get to later. We also need to make sure a default seed is set for the occasions when SLURM_ARRAY_TASK_ID is not set.

seed <- strtoi(Sys.getenv('SLURM_ARRAY_TASK_ID', unset = "0"))
set.seed(seed)

Now your script should look something like this;

#!/usr/bin/env Rscript


# Function for shared memory execution
doTask <- function(size_x, size_y, seed, print_progress){
    suppressPackageStartupMessages(library(doParallel))

    message(sprintf("Summing [ %e x %e ] matrix, seed = '%i'",size_x,size_y, seed))
    message(sprintf("Running on '%s' with %i CPU(s).", Sys.info()["nodename"], num_cpus))

    set.seed(seed)

    registerDoParallel((num_cpus/2))

    results_all <- foreach(z=0:size_x) %dopar% {
        percent_complete= z*100/size_x
        if (print_progress && percent_complete%%1==0){
            message(sprintf(" %i%% done...\r", percent_complete))
        }
        sum(rnorm(size_y))
    }
    Reduce("+",results_all)
}

# 50 calculations, store the result in 'x'

ntasks <- strtoi(Sys.getenv('SLURM_NTASKS', unset = "1")) 
seed <- strtoi(Sys.getenv('SLURM_ARRAY_TASK_ID', unset = "0"))
num_cpus <- as.integer(strtoi(Sys.getenv('SLURM_CPUS_PER_TASK', unset = "1")))

size_x <-60000 # This on makes memorier
size_y <-40000 # This one to make longer

# Time = (size_x/n) * size_y + c
# Mem  = (size_x * n) * c1 + size_y * c2

print_progress <- TRUE
# print_progress <- interactive() # Whether to print progress or not.

#If more than 1 task, use doMPI 
if (ntasks > 1){
    suppressPackageStartupMessages(library(doSNOW))
     cl <- makeMPIcluster(outfile="")

    results_all <- foreach(z=1:ntasks) %dopar% {
        doTask(size_x, ceiling(size_y/ntasks), z+seed, print_progress)
    }
    
    results = Reduce("+",results_all)
    stopCluster(cl)
    message(sprintf("Sums to %f", results))
}else{
    results = doTask(size_x, size_y, seed, print_progress)
    message(sprintf("Sums to %f", results))
}

Readability

Comments!

Debugging

#!/bin/bash -e

Exit bash script on error

#!/bin/bash -x

Print environment.

env

Print environment, if someone else has problems replicating the problem, it will likely come down to differences in your environment.

cat $0

Will print your input Slurm script to you output, this can help identify when changes in your submission script leads to errors.

Version control

Version control is when changes to a document are tracked over time.

In many cases you may be using the same piece of code across multiple environments, in these situations it can be difficult to keep track of changes made and your code can begin to diverge. Setting up version control like Git can save a lot of time.

Portability

Testing

More often than not, problems come in the form of typos, or other small errors that become apparent within the first few seconds/minutes of script.

Running on login node?

Control + c to kill.

Key Points

Write your script in a way that is independent of data or environment. (elaborate)

lesson home

Introduction to High-Performance Computing

lesson home

Writing good code

Overview

Use environment variables

Slurm Environment

Solution

Variables in Slurm Header

Default values

Interoperability

Verbose

Reproduceability

Readability

Debugging

Version control

Portability

Testing

Key Points

lesson home

lesson home