User Guide

Kabré Usage Tutorial

Kabré is a word from Ngäbe Language which means a bunch. This fits the current cluster composition, which features multiple parallel architectures. Through this tutorial you will learn how Kabré is composed, how to connect to the cluster, submit jobs, retrieve results and about environment modules.


To complete this tutorial you will require a ssh client. In Unix and Linux platforms, there is commonly a terminal emulator capable of establishing ssh sessions. In Windows plataforms you should download a ssh client programm, like Putty or simmilar.

Also, you will requiere an active account in Kabré and valid credentials. If you don’t have one, please contact us at, explaining your situation.

Understanding Kabré’s composition

The following image shows a network diagram of Kabré. We will discuss about the mayor components in the diagram.

Meta node

This is a very special node, it supports many services of the cluster. Its existence should be transparent for all users. If, for some reason, you find yourself in Meta-node, please leave and inform that situation, it could mean a problem. Running programs in Meta-node is considered bad behavior.


These nodes are a kind of shared-working area. When you log into Kabré, you will be assigned to one login node. Some common tasks you execute here are:

  • Creating and editing files
  • Creating directories and moving files
  • Coping files to and from your computer
  • Compiling code
  • Submitting jobs
  • Managing your active jobs

Running parallel code or heavy tasks on login nodes is considered bad behavior.

Machine Learning Nodes (Nukwa)

Each of these 4 nodes features a Nvidia Tesla K40 GPU. The host has an Intel Xeon with 4 cores @ 3.2 GHz without hyperthreading and 16 GB, therefore, only applications with an intensive use of GPUs would get a relevant speed-up.

Simulation Nodes (Nu)

The war horse of Kabré. 20 Intel Xeon Phi KNL nodes, each one with 64 cores @ 1.3 GHz and 96 GB; each core has two AVX units. If your application can be splitted in many smalls pieces and use vectorization, this architecture can provide impressive speed ups.

Big Data Nodes (Andalan)

Two of these three nodes feature an Intel Xeon,  each one with 24 cores @ 2.20 GHz, 2 threads per core and 64 GB RAM. The thrid node features an Intel Xeon with 16 cores @ 2.10 GHz, 2 threads per core, and 64 GB RAM. These nodes work quite well for sequential tools that need processing power.

Bioinformatics Nodes (Dribe)

This node features an Intel Xeon with 36 cores @ 3.00 GHz, 2 threads per core and 1024 GB RAM. This node works quite well for tools that require a high memory demand.

Interacting with Kabré

In this section we will cover ssh connections, ssh keys, and how to copy files.

SSH connections and SSH Keys

To start, open a terminal emulator and open a ssh session by typing

$ ssh

Remember to change user with your user name. Type your password when prompted. You will be logged to some login node. This is a typical linux terminal, so, try out some known commands, like lscdmkdir and so on.

Commonly, you will be using the same computer to interact with Kabré, your laptop or workstation, for example. A SSH Key is a file that will keep your connection secure while avoiding typing your password everytime you log in. To generate one, in your local computer (laptop, workstation…) open a terminal, go to your home directory and type

$ ssh-keygen -t rsa -C ""
 and follow the instructions. If you chose default options, the new key will be in ~/.ssh/, now you have to copy the public key to Kabré, to do so type
$ scp ~/.ssh/

Now, within a ssh session in Kabré, type

$ cat >> .ssh/authorized_keys
$ rm

Alternatively, you may execute this procedure in a single command if it is available in your working station.

$ ssh-copy-id

Now, if you open a new terminal and type ssh you will be logged without prompting for the password. This is because Kabré has the public key of your computer. It is also convenient for your computer to have the public key of Kabré, simply append it to autorized_keys in your local computer.

$ scp . 
$ cat >> ~/.ssh/autorized_keys
$ rm

Kabré’s file system

We count with two directories: /home and /work.

Directory Quota Purpose Backup
/home/userid 10GB scripts and important data monthly
/work/userid 100GB temporary data NO

To know the amount of available space left on each directory, you can use the following command:

$ df -h


  • The user is responsible for managing both home and work directories. The former is meant for programs and sensitive data, while the latter is meant for massive and temporary data.
  • Capacity of the work directory can be extended upon request. A clear justification for the extra space required must be provided. Send your request to
  • Although the home directory is backed up monthly, we strongly recommend users to work with a version control system (v.g. git) for their scripts and source codes.

Copy files between your computer and Kabré

The command scp is similar to cp command, it copies files from origin to destiny through a SSH session. It has the following syntax:

$ scp [user@][host:][path]origin_file [user@][host:][path]destiny_file

Default values are

  • user: your local user
  • host: local host
  • path: current working directory

scp must be executed on your local machine, not in Kabré. Maybe your application generates a lot of visualization files and you want to download those files to your computer, remember that ~ means home directory and * matches any sequence of characters:

$ scp*.viz ~/app/results/visualization

Or maybe you want to upload a parameters file to use in a simulation:

$ scp ~/research/app/parameters.dat

Understanding Kabré’s queues system

Login nodes are suitable for light task, as mentioned before: editing files, compiling, copying files, and similars. Heavy tasks are expected to run on Nu or Nukwa nodes. To enforce a fair sharing of resources among users, your task should be submitted to a queues system. It is like forming up at the bank, once your task makes its way to the head of the line, it will be granted all requested resources an will run until complete or until it consumes its time slot.

Currently, there are different queues for every component in Kabré, that means you cannot mix Nukwa nodes and Nu nodes in a single job, for example. The following table shows all available queues:

Partition (Queue) Platform Number of nodes Time slot
nu Xeon Phi KNL 1 72 h
nu-debug Xeon Phi KNL 1 8 h
nu-wide Xeon Phi KNL 12 24 h
nu-long Xeon Phi KNL 1 744 h
nukwa GPU 1 72 h
nukwa-debug GPU 1 8 h
nukwa-wide GPU 2 24 h
nukwa-long GPU 1 168 h
andalan Xeon 1 72 h
dribe Xeon 1 72 h

The process of submitting a job in Kabré could be divided in four steps: writting a SLURM file, queueing your job, monitoring jobs and retrieving results.

Writing a SLURM file

This configuration file tells the queue system all it needs to know about your job, so it can be placed in the right queue and executed. Lets try it out with a minimun working example. Below is a C code that approximates the value of pi using a Montecarlo method. Log into Kabré, copy the text to a file and save it with name pi_threads.c .

#include <pthread.h>
#include <math.h>
#include <stdlib.h>
#include <time.h>
#include <stdio.h>

typedef struct {
 int num_of_points;
 int partial_result;
} pi_params;

void * calc_partial_pi(void * p){
 pi_params * params = p;
 int count = 0;
 double r1, r2;
 unsigned int seed = time(NULL);

for(int i = 0; i < params->num_of_points; ++i){
 r1 = (double)rand_r(&seed)/RAND_MAX;
 r2 = (double)rand_r(&seed)/RAND_MAX;
 if(hypot(r1, r2) < 1)
 params->partial_result = count;

int main(int argc, char * argv[]){

if(argc != 3){
 printf("Usage: $ %s num_thread num_points\n", argv[0]);

int num_threads = atoi(argv[1]);
 int num_points = atoi(argv[2]);
 int num_points_per_thread = num_points / num_threads;

pthread_t threads[num_threads];
 pi_params parameters[num_threads];

for(int i = 0; i < num_threads; i++){
 parameters[i].num_of_points = num_points_per_thread;
 pthread_create(threads+i, NULL, calc_partial_pi, parameters+i);

for(int i = 0; i < num_threads; i++)
 pthread_join(threads[i], NULL);

double approx_pi = 0;
 for(int i = 0; i < num_threads; i++)
 approx_pi += parameters[i].partial_result;
 approx_pi /= (num_threads * num_points_per_thread) / 4;

printf("Result is %f, error %f\n", approx_pi, fabs(M_PI-approx_pi));


Currently, you are in a login node, so it is OK to compile the code there, do so by typing:

$ gcc -std=gnu99 pi_threads.c -lm -lpthread -o pi_threads

The following is an example SLURM file. All lines starting with #SBATCH are configuration commands for the queues system. Options shown are the most common, and possibly the only ones you will need.

Configuration Description
–job-name=<job_name> Specify the job’s name
–output=<result_name> Name of output file
–partition=<partition_name> In which queue should it run
–ntasks=<number> Number of process to run
–time=<HH:MM:SS> Approximate job’s duration

The body of a SLURM file is bash code. Copy the example in a file named pi_threads.slurm

#SBATCH --job-name=pi_threads
#SBATCH --output=result.txt
#SBATCH --partition=nu
#SBATCH --ntasks=1
#SBATCH --time=00:10:00

module load gcc/7.2.0

srun ./pi_threads 64 100000000000

Note: The command line arguments 64 and 100000000000 are the specific parametrs needed to run pi_threads.c

Now, from the command line, invoke the queues submitter:

$ sbatch pi_threads.slurm

And that’s all! Your job will be queued and executed, in this case, on a Xeon Phi node.

Monitoring your active jobs

A pasive way of monitoring your jobs is to indicate SLURM to send an email when done. This can be configured in the SLURM file using the following options:

Configuration Description
–mail-user=<email> Where to send email alerts
–mail-type= <BEGIN|END|FAIL|REQUEUE|ALL> When to send email alerts
#SBATCH --mail-type=END,FAIL

After submitting your job you can check its status with these commands:

Configuration Description
squeue -u <username> Check jobs for a specific user
sinfo Display all nodes (with attribute)
scontrol show job <job_id> Status of particular job
scancel <job_id> Delete job

To show details on the state of a job:

$ squeue -j <jobid>

To show details every <num_seconds> on the state of a job:

$ watch -n <num_seconds> squeue -j <jobid>

Valid Job States

To understand the Job State Codes that you may encounter check the following:

Code State
CA Canceled
CD Completed
CF Configuring
CG Completing
F Failed
NF Node Fail
PD Pending
R Running
TO Timeout

Retrieving results

By default, every job will generate an output files as in the example above:


You can copy this file to your local computer or run another script for post-processing the output.

Interactive jobs

Sometimes you want to have direct access to some node. Using ssh directly is a bad practice, because the queue system could send someone else’s job to execute on the node you are currently using. The polite way to ask for direct access is through an interactive job that will put you in an interactive sell on a compute node. This allows you to experiment with different options and variables that will provide immediate feedback. To access:

$ salloc

This command starts an interactive job in Nu-debug node. Interactive jobs are allowed in debug queues only. The next command opens an interactive job on a Nukwa node and reserves a GPU for your use:

$ srun --partition=nukwa-debug --pty --preserve-env $SHELL

Environment modules

Different users have different needs, sometimes those needs could be conflincting, for example, multiple versions of the same library. These situations are solved with environment modules. A typical case is different versions of Python. To exemplify, ask for the queue in which it should be run by typing:

$ SBATCH --partition=nu

Go ahead and type $ python, you should get into the default pyhton interpreter, the header should be like this:

Python 2.7.5 (default, Nov 6 2016, 00:28:07) 
[GCC 4.8.5 20150623 (Red Hat 4.8.5-11)] on linux2
Type "help", "copyright", "credits" or "license" for more information.

Besides the default interpreter, you can execute Intel Distribution for Python, a specifically tunned compilation with packages commonly used in scientific computing. To get intel python, type

module load intelpython/3.5

Now, type again $ python, you will get a different header:

Python 3.5.2 |Intel Corporation| (default, Oct 20 2016, 03:10:33) 
[GCC 4.8.2 20140120 (Red Hat 4.8.2-15)] on linux
Type "help", "copyright", "credits" or "license" for more information.
Intel(R) Distribution for Python is brought to you by Intel Corporation.
Please check out:

To check which modules are already loaded, type

$ module list

To get a list of all available modules, type

$ module avail

Behind scenes, module command is just configuring paths, aliases and other environment variables, so, modules are loaded only for the current shell session. You can request specific modules in you jobs, just add «module load module_name» lines to SLURM file body, below all #SBATCH lines and before runing your program.