Using the Verification Cluster

The Verification Cluster is a cluster of servers run by NCSA for LSST DM development work.

To get an account, see the Onboarding Checklist.

This page is designed help you get started on the Verification Cluster:

  1. Overview of the Verification Cluster
  2. Account Password
  3. GPFS Directory Spaces
  4. Shared Software Stack in GPFS
  5. SLURM Job Submission

Overview of the Verification Cluster

lsst-dev01 is a system with 24 cores, 256 GB RAM, running the latest CentOS 7.x that serves as the front end of the Verification Cluster. lsst-dev01 is described in further detail at Using the lsst-dev Server and on the page of available development servers .

The Verification Cluster consists of 48 Dell C6320 nodes with 24 physical cores (2 sockets, 12 cores per processor) and 128 GB RAM. As such, the system provides a total of 1152 cores.

The Verification Cluster runs the Simple Linux Utility for Resource Management (SLURM) cluster management and job scheduling system. lsst-dev01 runs the SLURM controller and serves as the login or head node , enabling LSST DM users to submit SLURM jobs to the Verification Cluster.

lsst-dev01 and the Verification Cluster utilize the General Parallel File System (GPFS) to provided shared-disk across all of the nodes. The GPFS will have spaces for archived datasets and scratch space per user to support computation/analysis.

The legacy NFS /home directories are available on the front end lsst-dev01 (serving as the current home directories), but are not mounted on the compute nodes of the Verification Cluster.

Report system issues to lsst-sysadm _at_ ncsa.illinois.edu

Account Password

You can log into LSST development servers at NCSA such as lsst-dev01 with your NCSA account and password. You can reset your NCSA password at the following URL:

GPFS Directory Spaces

GPFS is available on the login node lsst-dev01 and on all of the compute nodes of the Verification Cluster. For convenience the bind mounts /scratch , /datasets , and /software have been created to provide views into corresponding spaces in GPFS. Users will find directories

/scratch/<username>

ready and available for use. The per user /scratch space is volatile with a 180 day purge policy and is not backed up.

Project managed datasets will be stored within the /datasets space. The population of /datasets with reference data collections is still in the early stages; a first example is the SDSS DR7 Stripe82 data, which can be found at

/datasets/stripe82/dr7/runs

To add/change/delete datasets, see Common Dataset Organization and Policy.

Shared Software Stack in GPFS

A shared software stack on the GPFS file systems, suitable for computation on the Verification Cluster, has been provided and is maintained by Science Pipelines and is available under /software/lsstsw. This stack may be initialized via:

% .  /software/lsstsw/stack/loadLSST.bash

SLURM Job Submission

Documentation on using SLURM client commands and submitting jobs may be found at standard locations (e.g., a quickstart guide). In addition to the basic SLURM client commands, there are higher level tools that can serve to distribute jobs to a SLURM cluster, with one example being the combination of pipe_drivers and ctrl_pool within LSST DM. For exhaustive documentation and specific use cases, we refer the user to such resources. On this page we display some simple examples for getting started with submitting jobs to the Verification Cluster.

To examine the current state and availability of the nodes in the Verification Cluster, one can use the SLURM command sinfo:

% sinfo
PARTITION AVAIL  TIMELIMIT  NODES  STATE NODELIST
debug*       up   infinite      6  fail* lsst-verify-worker[05,10,17,23,27,47]
debug*       up   infinite     42   idle lsst-verify-worker[01-04,06-09,11-16,18-22,24-26,28-46,48]

% sinfo  -N -l --states="idle"
Thu Sep 15 08:28:52 2016
NODELIST              NODES PARTITION       STATE CPUS    S:C:T MEMORY TMP_DISK WEIGHT FEATURES REASON
lsst-verify-worker01      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker02      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker03      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker04      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker06      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker07      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker08      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker09      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker11      1    debug*        idle   48   48:1:1      1        0      1   (null) none
...
lsst-verify-worker40      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker41      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker42      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker43      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker44      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker45      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker46      1    debug*        idle   48   48:1:1      1        0      1   (null) none
lsst-verify-worker48      1    debug*        idle   48   48:1:1      1        0      1   (null) none

In this view sinfo shows the nodes to reside within a single partition debug, and the worker nodes show 48 possible hyperthreads on a node (in the future this may be reduced to reflect the actual 24 physical cores per node). At the time of this sinfo invocation there were 42 verification nodes available, shown by the “idle” state. The SLURM configuration currently does not perform accounting, and places no quotas on users’ total time usage.

Simple SLURM jobs

In submitting SLURM jobs to the Verification Cluster it is advisable to have the software stack, data, and any utilities stored on the GPFS /scratch , /datasets , and/or /software spaces so that all are reachable from lsst-dev01 and each of the worker nodes. Some simple SLURM job description files that make use of the srun command are shown in this section. These are submitted to the queue from a standard login shell on the front end lsst-dev01 using the SLURM client command sbatch, and their status can be checked with the command squeue :

For a single task on a single node:

% cat test1.sl
#!/bin/bash -l
#SBATCH -p debug
#SBATCH -N 1
#SBATCH -n 1
#SBATCH -t 00:10:00
#SBATCH -J job1

srun sleep.sh


% cat sleep.sh
#!/bin/bash
hostname -f
echo "Sleeping for 30 ... "
sleep 30


Submit with :
% sbatch test1.sl

Check status :
% squeue
    JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
      109     debug     job1    daues  R       0:02      1 lsst-verify-worker11

This example job was assigned jobid 109 by the SLURM scheduler, and consequently the standard output and error of the job were written to a default file slurm-109.out in the current working directory.

% cat slurm-109.out
 lsst-verify-worker11.ncsa.illinois.edu
 Sleeping for 30 ...

To distribute this script for execution to 6 nodes by 24 tasks per node (total 144 tasks), the form of the job description is:

% cat test144.sl
#!/bin/bash -l
#SBATCH -p debug
#SBATCH -N 6
#SBATCH -n 144
#SBATCH -t 00:10:00
#SBATCH -J job2

srun sleep.sh


Submit with :
% sbatch test144.sl

For these test submissions a user might submit from a working directory in the /scratch/<username> space with the executable script sleep.sh and the job description file located in the current working directory.

Interactive SLURM jobs

A user can schedule and gain interactive access to Verification Cluster compute nodes using the SLURM salloc command. Example usage is:

For a single node:

% salloc  -N  1 -p debug -t 00:30:00  /bin/bash
salloc: Granted job allocation 108

% squeue
         JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
           108     debug     bash    daues  R       0:58      1 lsst-verify-worker01
% hostname -f
lsst-dev01.ncsa.illinois.edu

% srun hostname -f
lsst-verify-worker01.ncsa.illinois.edu

One can observe that after the job resources have been granted, the user shell is still on the login node lsst-dev01. The command srun can be utilized to run commands on the job’s allocated compute nodes. Commands issued without srun will still be executed locally on lsst-dev01.

SLURM Example Executing Tasks with Different Arguments

In order to submit multiple tasks that each have distinct command line arguments (e.g., data ids), one can utilize the srun command with the --multi-prog option. With this option, rather than specifying a single script or binary for srun to execute, a filename is provided as the argument of the --multi-prog option. In this scenario an example job description file is:

% cat test1_24.sl
#!/bin/bash -l

#SBATCH -p debug
#SBATCH -N 1
#SBATCH -n 24
#SBATCH -t 00:10:00
#SBATCH -J sdss24

srun --output job%j-%2t.out --ntasks=24 --multi-prog cmds.24.conf

This description specifies that 24 tasks will be executed on a single node, and the standard output/error from each of the tasks will be written to a unique filename with format specified by the argument to --output. The 24 tasks to be executed are specified in the file cmds.24.conf provided as the argument to the --multi-prog option. This commands file will have a format that maps SLURM process ids (SLURM_PROCID) to programs to execute and their commands line arguments. An example command file has the form :

% cat cmds.24.conf
0 /scratch/daues/exec_sdss_i.sh run=4192 filter=r camcol=1 field=300
1 /scratch/daues/exec_sdss_i.sh run=4192 filter=r camcol=4 field=300
2 /scratch/daues/exec_sdss_i.sh run=4192 filter=g camcol=4 field=297
3 /scratch/daues/exec_sdss_i.sh run=4192 filter=z camcol=4 field=299
4 /scratch/daues/exec_sdss_i.sh run=4192 filter=u camcol=4 field=300
...
22 /scratch/daues/exec_sdss_i.sh run=4192 filter=u camcol=4 field=303
23 /scratch/daues/exec_sdss_i.sh run=4192 filter=i camcol=4 field=298

The wrapper script exec_sdss_i.sh used in this example could serve to “set up the stack” and place the data ids on the command line of processCcd.py :

% cat exec_sdss_i.sh
#!/bin/bash
# Source an environment setup script that holds the resulting env vars from e.g.,
#  . ${STACK_PATH}/loadLSST.bash
#  setup lsst_distrib
source /software/daues/envDir/env_lsststack.sh

inputdir="/scratch/daues/data/stripe82/dr7/runs/"
outdir="/scratch/daues/output/"

processCcd.py  ${inputdir}  --id $1 $2 $3 $4 --output ${outdir}/${SLURM_JOB_ID}/${SLURM_PROCID}