1. INTRODUCTION TO VIPBG LINUX CLUSTERBy
Helen Wang
October 12, 2016

2. What is HPC
High-Performance Computing (HPC) clusters are characterized by many cores and processors, lots of memory, high-speed networking, and large data stores – all shared across many rack-mounted servers. User programs that run on a cluster are called jobs, and they are typically managed through a queueing system for optimal utilization of all available resources.
An HPC cluster is made of many separate servers, called nodes, filling into racks. HPC typically involves simulation of numerical models or analysis of data from scientific instrumentation.
At the core of HPC is manageable hardware and systems software either developed by systems programmers or provided by software company, which allow researchers to devote their energies to their code.

3. Basic Beowulf Cluster Structure

4. A brief look of our cluster

5. What our cluster can do Intensive computation
Leveraging a total of 600+ cores, and large memory available, the Linux cluster system is configured to ensure multiple job types run efficiently.
Parallel computation
Linear processing - Each section of computation does not need node communication much, such as image processing – near-linear result with adding more nodes
Explicit parallel computation
programs that have explicitly coded parallel sections of their code that can run independently such as OpenMx
Fully parallelizable programs in that they contain one or more interdependent serial or atomic components that must be synchronized across all instances of the parallel application. This synchronization is usually done through libraries such as R library snow, or MPI

6. why we use cluster and what to expect
Use it when you have lots of independent programs to run to finish a project – think about parallel
Use it when your program is taking very long time to finish –give your laptop a break
Use it when your program is using up lot of memory – to avoid crash on your pc
What to expect
It is not going to be super fast to run a single command line even with hundreds of cpu cores
It is not set up for one single user even it looks like one super computer to you
You are sharing resources with lots of hardcore users

7. What is our cluster configuration
a newly updated Linux Beowulf cluster, installed with applications supporting computationally intensive processing. The following description can be used for the "Computing Resources" section of grant applications.
- 32 Dell PE R630/R620 servers with CentOS/Redhat bits Linux OS
cores using Intel Xeon 56XX processors (2.67GHz to 3.4GHz)
- Total 4TB RAM ( 80GB-128GB per node)
450TB network attached storage with 500TB backup storage –
10TB internal disk storage  ( 120GB-900GB per node)
- 40GB InfiniBand network connections to all nodes and storage.
- Fail-over redundant master servers

8. Software available on cluster
R with CRAN packages and BIOCONDUCTOR packages
C++/G++ and Fortran compilers
JAVA compiler
Perl compiler
Python 3.4 and Biopython compiler
SAS-9.4 for Linux 64bit
PLINK
MPI and OpenMPI
PBS Pro 12.2 (a portable batch system for cluster)
Additional open source software upon user request

9. Who can use it and when?
The cluster supports department faculty, research staff, and students working on intensive computation in both serial and parallel processing. All users are able to access cluster through VCU VPN for a secured connection.
You can use it at anytime and anywhere, mostly when you start your graduate thesis. Please consider start your computation work on cluster for at least 6-8 months in advanced, since the system can be busy and you won’t get the sufficient resources for estimated time.
Come to see me to prepare the code and parallel processing.

10. VIPBG Cluster Login Info
Server Name: light.vipbg.vcu.edu
IP:
2nd server as failover: group.vipbg.vcu.edu
IP: (invisible on mission)
Software recommended to access servers:
PC USERS:
1. MobaXterm
2. ssh /open ssh / putty / winscp
MAC USERS: Mac Terminal, Fetch

11. Access Cluster Server and nodes
Master node (master1 / master2): light.vipbg.vcu.edu
Running CentOS ( redhat kenrnel)Version 6.7, x86-64
Open source or Software download – choose 64 bits CentOS or RHEL 6 if possible
Purposes:
front-end user interface;
slow; - not for running any jobs. Jobs running on master will be terminated without notice
accessible from outside through VPN;
Slave nodes (nodes):
node1–node29 : dual quart core or 6 core Intel Xeon processors with GB GB RAM
computation;
not prefer to access user interface, accessible via master and managed by portable batch management ( PBS );
fast;
internal network; X, not accessible directly from outside

12. Access “LIGHT” from your computer
How to use MobaXterm or ssh to access server
Outside of VCU please use VCU VPN Client through ramsvpn.vcu.edu
Open new session then SSH to add the server name
Open “SSH settings” to fill the information
remote hostname:
username: YOUR_ACCOUNT_NAME
port number: 22
Open session settings to put merlot in Session Name
ssh –X for graphical access
Select the server to test the connection and exchange keys by giving password
Create profile or bookmark for the easy access every time

13. UNIX Commands You Need to Know
pwd
clear
mkdir cp cd mv ls
head
more less wc
man rm
chmod
grep with head and tail
nano
sed
cut
top
scp
nproc –all
lscpu
cat /proc/cpuinfo
echo Cores = $(( $(lscpu | awk '/Socket/{ print $2 }') * $(lscpu | awk '/Core/{ print $4 }') ))
vmstat -s | grep memory
cat /proc/meminfo | grep MemTotal | awk '{ print $2 }’
awk '/MemTotal/ {print $2}' /proc/meminfo

14. How to use cluster to submit jobs
IMPORTANT
The logon machine (LIGHT.VIPBG.VCU.EDU) is only used for login.
Jobs running on merlot will be terminated without notice.
What do you need to submit a job via PBS
an executable script, the script can be a program code ( R, SAS or other language) , shell script or a collection of command lines.
test how much resources and time you may need before you submit multiple jobs.
which queue to use to submit a job
Node and queue configuration
Nodes: Nodes are the physical computer servers incorporated together to make the cluster.
Queues: Queues are being used by pbs scheduler to send jobs to different nodes, each node is assigned to a queue to handle different type of jobs.
Most queue limits can be checked by running the command
qstat -q.
Note that if you need more job permissions, please send request to system admin and your supervisor to get a temporary expansion on job submissions

15. Nodes and queues configuration
Queue Name
Node assigned
Job Limit (proc/user)
Comments
Express
Node4 2
Run small urgent jobs
serial
13-22 20
Run R and generic jobs RAM<2GB
Openmx
9-12, 23-25 15
Openmx and parallel jobs
workq
7-8, 26,27,28
Run large memory jobs RAM<30GB
Mxq
Node5 10
Run traditional mx jobs

16. Submitting a Job
Jobs are submitted to a PBS queue so that PBS can dispatch them to be run on one or more of a cluster's compute nodes. There are two main types of PBS jobs:
Non-interactive Batch Jobs: This is the most common PBS job. A job script is created that contains PBS resource requests and the commands necessary to execute the job. The job script is then submitted to PBS to be run non-interactively.
Interactive Batch Jobs: This is a way to get an interactive terminal on one or more of the compute nodes of a cluster. Commands can then be run interactively through that terminal directly on the compute nodes for the duration of the job. Interactive jobs are helpful for such things as program debugging and running many short jobs.

17. #PBS –I walltime=HH:MM:SS
A PBS script is a standard Unix/Linux shell script that contains a few extra comments at the beginning that specify directives to PBS. These comments all begin with #PBS.
The most important PBS directives are:
PBS Directives
Description
#PBS –I walltime=HH:MM:SS
This directive specifies the maximum walltime (real time, not CPU time) that a job should take. If this limit is exceeded, PBS will stop the job. Keeping this limit close to the actual expected time of a job can allow a job to start more quickly than if the maximum walltime is always requested.
#PBS -l pmem=SIZEgb
This directive specifies the maximum amount of physical memory used by any process in the job. For example, if the job would run four processes and each would use up to 2 GB (gigabytes) of memory,  then the directive would read #PBS -l pmem=2gb
#PBS -l nodes=N:ppn=M
This specifies the number of nodes (nodes=N) and the number of processors per node (ppn=M) that the job should use. PBS treats a processor core as a processor, so a system with eight cores per compute node can have ppn=8 as its maximum ppn request. Note that unless a job has some inherent parallelism of its own through something like MPI or OpenMPI, requesting more than a single processor on a single node is usually wasteful and can impact the job start time.
#PBS -q queuename
This specifies what PBS queue a job should be submitted to. This is only necessary if a user has access to a special queue. This option can and should be omitted for jobs being submitted to a system's default queue.
#PBS -j oe
Normally when a command runs it prints its output to the screen. This output is often normal output and error output. This directive tells PBS to put both normal output and error output into the same output file.

18. An example of PBS script
#This is a sample PBS script. It will request 1 processor on 1 node for 10 hours. # #Request 1 processors on 1 node # #PBS -l nodes=1:ppn=1 # #Request 10 hours of walltime # #PBS -l walltime=10:00:00 # #Request 1 gigabyte of memory per process # #PBS -l mem=1gb # #Request that regular output and terminal output go to the same file # #PBS -j oe # #The following is the body of the script. By default, PBS scripts execute in your home directory, not the #directory from which they were submitted. The following line places you in the directory from which the job #was submitted. # cd $PBS_O_WORKDIR # #Now we want to run the program "hello". "hello" is in the directory that this script is being submitted from, #$PBS_O_WORKDIR. # echo " " echo " " echo "Job started on `hostname` at `date`" ./hello echo " " echo "Job Ended at `date`" echo " "

19. Template used on cluster
Modify template to create your own pbs script for running programs
#!/bin/bash
#PBS -q serial
#PBS -N MYSCRIPT #
# cd to the directory from which I submitted the job. Otherwise it will execute in my home directory.
set WORKDIR = ~/YOURWORDIR
#PBS -V
#echo “PBS batch job id is $PBS_JOBID“
echo "Working directory of this job is: " $WORKDIR
echo "Beginning to run job“
Command line you need to execute the job ( /home/huan/bin/calculate - PARAMETEERS)

20. Job Submission Syntax qsub SCRIPTFILE Existing job submission scripts
/usr/local/bin/q*
R USERS qR YOUR_R_SCRIPT
Large memory R jobs qRL YOUR_R_SCRIPT
SAS USERS qsas YOUR_SAS_CODE
Generic or other resources qsub YOUR_OWN_SCRIPT
Interactive Batch Jobs
Interactive PBS jobs are similar to non-interactive PBS jobs in that they are submitted to PBS via the command qsub.
When submitting an interactive PBS job, PBS script is not necessary. All PBS directives can be specified on the command line. The syntax for qsub for submitting an interactive PBS job is:
qsub -I ... pbs directives..
The -I flag above tells qsub that this is an interactive job.
The following example shows using qsub to submit an interactive job using one processor on one node for four hours
merlot:~$ qsub -I -l nodes=1:ppn=1 -l walltime=4:00:00 qsub: waiting for job merlot.bis.vcu.edu start qsub: job merlot.bis.vcu.edu ready node12:~$
There are two things of note here. The first is that the qsub command doesn't exit when run with the interactive -I flag. Instead, it waits until the job is started and gives a prompt on the first compute node assigned to a job. The second thing of note is the prompt node12:~$ - this shows that commands are now being executed on the compute node node12.

21. Monitoring and Managing Jobs
Check Job Status using qstat
Command
Description
qstat
Shows the status of all PBS jobs. The time displayed is the CPU time used by the job.
qstat –s
qstat -a
Shows the status of all PBS jobs. The time displayed is the walltime used by the job.
qstat –u USERID
Shows the status all PBS jobs submitted by the user userid. The time displayed is the walltime used by the job.
qstat -n
Shows the status all PBS jobs along with a list of compute nodes that the job is running on.
qstat –f JOBID
Shows detailed information about the job jobid.

22. Job Running Status State meaning Q
The job is queued and is waiting to start. R
The job is currently running E
The job is currently ending. H
The job has a user or system hold on it and will not be eligible to run until the hold is removed.

23. - qdel JOBID delete a job by Job_ID
Managing jobs
Deleting jobs
- qdel JOBID delete a job by Job_ID
- qdel $(qselect –u USERNAME) delete all jobs owned by USERNAME
View job output
If the PBS directive #PBS -j oe is used in a PBS script, the non-error and the error output are both written to the Jobname.oJob_ID file.
JobName.oJobID : This file would contain the non-error output that would normally be written to the screen.
JobName.eJobID: This file would contain the error output that would normally be written to the screen.

24. More to monitor a node To check a node configuration
$pbsnodes NODE#
To check a node status
nodestatus NODE#
Limitation for the name of the SCRIPT
No more than 10 characters
no space in between
no special characters.
use a temporary name if necessary and change it back when the job is done.

25. At Last Edit file using nano or vi
use samba connection to map a network drive on PC, recommending to use “EditPad Lite”
Useful links
Wiki page for vipbg cluster – need vcu eID to login