Skip to end of metadata
Go to start of metadata

Introduction

The DAVinCI system consists of 2400 processor cores in 192 Westmere nodes (12 processor cores per node) at 2.83 GHz with 48 GB of RAM per node (4 GB per core) and 6 Sandy Bridge nodes (16 processor cores per node) at 2.2GHz with 128 GB of RAM per node (8 GB per core).  All of the nodes are connected via QDR InfiniBand (40 Gb/s ) both to each other and to the GPFS fast scratch storage system.  16 of the Westmere nodes are equipped with NVIDIA Fermi GPGPUs.  This system is a hybrid system accommodating HTC (high throughput computing) serial jobs and tightly-coupled parallel (MPI) jobs along with General-Purpose Computing on Graphics Processing Units (GPGPU) and 3D stereo visualization.  It will accommodate serial jobs that need one node (12 cores or less) and multinode parallel jobs (24 cores up to 1152 cores).

Prerequisite for Using This System

All of the clusters that make up our shared computing resources run the Linux operating system (not Windows or MacOS operating systems).  In order to effectively utilize the clusters you must have knowledge of Linux, how to navigate the filesystem, how to create, edit, rename, and delete files, and how to run basic commands and write small scripts.  If you need assistance in this area please review the tutorials that are available on our web site.

Logging Into the Cluster

The cluster login nodes can be accessed through Secure Shell from any machine on the Rice campus network.  You will need an active NetID and password in order to login (unless otherwise instructed).

You must apply for an account.

If you do not have an account on the Shared Computing Resources then you should apply for one. You will need a faculty sponsor for your account who is willing to pay the access fee.

If you need off-campus access, please visit our Off-Campus Access Guide.

To login to the system from a Linux or Unix machine, use the ssh command:

Substitute the actual host name of the cluster in place of hostname.rice.edu above.  For example, stic.rice.edu or davinci.rice.edu, depending on which cluster you want to access.

To transfer files into the cluster from a Linux or Unix machine, use the scp command:

Substitute the actual host name of the cluster in place of  hostname.rice.edu  above.  For example, stic.rice.edu or davinci.rice.edu, depending on which cluster you want to access.

For more information about using Secure Shell, please see our Using SSH to Login and Copy Files Guide.

Login Nodes

Once you are logged in to the system, you are logged into one of several login nodes as shown in the diagram below. These nodes are intended for users to compile software, prepare data files, and submit jobs to the job queue. They are not intended for running compute jobs. Please run all compute jobs in one of the job queues described later in this document.

login-node-diagram

Diagram courtesy of Chris Hunter, Rice University

Do not run compute jobs on login nodes

Cluster login nodes are multi-user access points intended for users to compile software, copy and prepare data files, and submit jobs to the job queue. Any user running intensive computational tasks directly on the login node risks disciplinary action up to and including the loss of their access privileges.

Data and Quotas

A summary of all filesystems available to all users is presented in the following table:

Filesystem

Accessed via environment variable

Physical Path

Size

Quota

Type

Purge Policy

Home directories

$HOME

/home

2 TB

4 GB

NFS

none

Group Project directories

$PROJECTS

/projects

4 TB

50 GB per group

NFS

none

Shared Scratch high performance I/O

$SHARED_SCRATCH

/scratch

129 TB

None

GPFS

14 days

Local Scratch on each node

$LOCAL_SCRATCH

/tmp

200 GB

None

Local

at the end of each job

Work storage space

$WORK

/work

214 TB

2 TB per group

GPFS

none

 

 

$SHARED_SCRATCH is not permanent storage

$SHARED_SCRATCH is to be used only for job I/O.  Delete everything you do not need for another run at the end of the job or move to $WORK for analysis. Staff may periodically delete files from the $SHARED_SCRATCH file system even if files are less than 14 days old. A full file system inhibits use of the system for everyone. Using programs or scripts to actively circumvent the file purge policy will not be tolerated.

 

Data Backups Are the Responsibility of Each User

Backing up and archiving data remains the sole responsibility of the end user. At this point in time the shared computing enterprise does not offer these services in any automated way. We strongly encourage all users to take full advantage of Central IT's storage services, departmental servers, and/or individual group resources to prevent accidental loss or deletion of critical data. Please feel free to contact the CRC for advice on best practices in data management (e.g. IT's Subversion (SVN) services provide the safest environment for maintaining and developing source code). We welcome any suggestions for offering a higher level of data security as we move forward with shared computing at Rice.


Research Data Compliance

Due to recent changes in NSF, NIH, DOD, and other government granting agencies, Research Data Management has become an important area of growth for Rice and is a critical factor in both conducting and funding research. The onus of maintaining and preserving research data generated by funded research is placed squarely upon the research faculty, post docs, and graduate students conducting the research. It is imperative that you are aware of your compliance responsibilities so as not to jeopardize the ability of Rice University to receive federal funding. We will help in any way possible to provide you the information and assistance you need, but the best place to start is the campus research data management website.

To see your current quota and your disk usage for your home directory, run this command:

To see the quota and usage for the $PROJECTS directories for all groups that you belong to, run this command:

To see the quota and usage for the $WORK directories on DAVinCI, BIOU, and PO for the primary group to which you belong, run this command:

 

The clustered file system $SHARED_SCRATCH provides fast, high-bandwidth I/O for running jobs. Though not limited by quotas, $SHARED_SCRATCH is intended for in-flight data being used as input and output for running jobs, and may be periodically cleaned through voluntary and involuntary means as use and abuse dictate.

Volatility of $SHARED_SCRATCH

The $SHARED_SCRATCH filesystem is designed for speed rather than data integrity and therefore may be subject to catastrophic data loss! It is designed for input and output files of running jobs, not persistent storage of data and software.

When dealing with $SHARED_SCRATCH always copy your data in. A "cp" will update the access time on files whereas a move "mv" will preserve the access time. This is important as our periodic cleaning mechanism may purge files where the access time is maintained via the "mv" command.

Avoid I/O over NFS

$HOME and $PROJECTS should not be used for job I/O. Jobs found to be using $HOME and $PROJECTS for job I/O are subject to termination without notice.

Use Variables Everywhere!

NOTE: The physical paths for the above file systems are subject to change. You should always access the filesystems using environment variables, especially in job scripts.

For information on how to use $PROJECTS, please see our FAQ.

 

Environment and Shells

The default shell on all the CRC clusters is bash. Other popular shells are available. To have your account's default shell changed from bash to one of these, please file a help request and specify the cluster, username, and desired shell in the ticket. Once your shell is changed this is reflective on all clusters with which you have access. Any active login sessions when your shell is changed will need to be terminated to effect change.

Due to the nature of high performance applications and the batch scheduling system used on CRC clusters, managing your shell environment variables properly is vital.

Customizing Your Environment With the Module Command

Each user can customize their environment using the module command. This command lets you select software and will source the appropriate paths and libraries. All the requested user applications are located under the /opt/apps directory.

To list what applications are available, use the spider sub command:

To see a description of a specific package, use the spider sub command again:

To load the module for OpenMPI built with the GCC compilers, for example, use the load sub command:

To see a list of modules that you have loaded, use this command:

To change to the Intel compiler build of OpenMPI use the swap sub command:

To unload all of your modules, use this command:

To make sure a set of modules are loaded automatically at login, use the module save sub command:

Sometimes a module will not load without explicit dependencies. The following outlines this "error" and what to do.

The Job Scheduler

The batch job scheduling system implemented on this system uses SLURM. SLURM is responsible for resource management, job scheduling, and monitoring. 

Fairshare Scheduling Policy

We implement the SLURM Fairshare feature to provide a fair utilization of the available resources.  This is accomplished by allowing historical resource utilization information to be incorporated into job feasibility and priority decisions. This is normally the most significant component of a job's priority, which ultimately defines the position of the job on a queue. We do not use a FIFO (First-In-First-Out) scheduler.  Your jobs' priority will be determined by your utilization over the past seven days (sliding window), with high utilization resulting in lower priority for new jobs.

Backfill Scheduling Policy

This is a scheduling optimization which allows SLURM to make better use of available resources by running jobs out of order. Using job data such as walltime and resources requested, the scheduler can start other, lower-priority jobs so long as they do not delay the highest priority jobs.  Because of the way it works, essentially filling in holes in node space, backfill tends to favor smaller and shorter running jobs more than larger and longer running ones.

Accurate Walltime Improves Scheduling of Jobs

It is important to specify an accurate walltime for your job in your SLURM submission script.  Selecting the default walltime for jobs that are known to run for less time may result in the job being delayed by the scheduler due to an overestimation of the time the job needs to run.

Automatic Queue Routing

Each of our compute resources has a pre-defined default queue.  If you submit your job without specifying a queue, your job will be automatically routed to the default queue.  Therefore, be aware of which queue you intend for the job to run in and specify this queue in your SLURM batch script.

 

Available Partitions and System Load

Partition Name

Maximum nodes
Per Job

Minimum threads
Per Job

Maximum threads
Per Job

Maximum jobs
running per user

Maximum
run time

commons

48

1

576500

8:00:00

serial_long11125003-00:00:00
interactive414850030:00

The definition of the queues are as follows:

commons - default partition, intended for all jobs.

serial_long - partition intended for serial jobs needing longer than 24 hours but less than 72 hours (3 days).

interactive - partition intended for interactive jobs or short running testing jobs.

 

Use the following command to determine the partitions with which you have access. Please note in the output the Account column information needs to be provided to your batch script in addition to the partition information.

 

Determining Partition Status

A good way to obtain the status of all partitions and their current usage is to run the following SLURM command:

Here is a brief description of the relevant fields:

PARTITION: Name of a partition. Node that the suffix "*" identifies the default partition.
AVAIL: Partition state: up or down.
TIMELIMIT:  Maximum time limit for an user job in days-hours:minutes:seconds.
NODES: Count of nodes with this particular configuration by node state in the form "[A]vailable/[I]dle/[O]ther/[T]otal
STATE: State of the nodes.
NODELIST: Names of nodes associated with this configuration/partition.

See the manpage for sinfo for more information

Submitting Jobs with SLURM

Once you have an executable program and are ready to run it on the compute nodes, you must create a job script that performs the following functions:

  • Use job batch options to request the resources that will be needed (i.e. number of processors, run time, etc.), and
  • Use commands to prepare for execution of the executable (i.e. cd to working directory, source shell environment files, copy input data to a scratch location, copy needed output off of scratch location, clean up scratch files, etc).

After the job script has been constructed you must submit it to the job scheduler for execution. The remainder of this section will describe the anatomy of a job script and how to submit and monitor jobs.

Please note script options are being provided using the long options and not the short options for readability and consistency e.g. --nodes versus -N.

Per Cluster restrictions

Warning per cluster restrictions may require you to customize the following generic instructions. e.g. NOTS has a maximum of 1 node per job, but DAVinCI does not that restriction. Please refer to the introduction of each cluster for requirements.


SLURM Batch Script Options

All jobs must be submitted via a SLURM batch script or invoking sbatch at the command line . See the table below for SLURM submission options.

Option

Description

#SBATCH --account=AccountName

#SBATCH --partition=PartitionName

Required: You need to specify both the name of the account and partition to schedule jobs especially to use a condo on the cluster.

Use the command sacctmgr show assoc user=netID to show which accounts and partitions with which you have access.

#SBATCH --job-name=YourJobName

Required:  Assigns a job name.  The default is the name of SLURM job script.

#SBATCH --ntasks=2 Recommended: The number of tasks per job. Usually used for MPI jobs.

You can get further explanation here .

#SBATCH --nodes=2

Recommended: The number of nodes requested.
You can get further explanation
here .

#SBATCH --ntasks-per-node=2

Recommended: The number of tasks per node. Usually used in combination with --nodes for MPI jobs.
You can get further explanation
here .

#SBATCH --cpus-per-task=4Recommended:  The number processes per task. Usually used for OpenMP or multi-threaded jobs.

#SBATCH --partition=PartitionName

Recommended:  Specify the name of the Partition (queue) to use. Use this to specify the default partition or a special partition i.e. non-condo partiton with which you have access.

#SBATCH --time=08:00:00

Recommended:  The maximum run time needed for this job to run, in days-hh:mm:ss.  If not specified, the default run time will be chosen automatically.

#SBATCH --mem-per-cpu=1024M

Optional: The maximum amount of physical memory used by any single process of the job ([M]ega|[G]iga|[T]era)Bytes.

The value of "mem-per-cpu" multiplied by "tasks" (mem X tasks) should not exceed the amount of memory on a node.

 See our FAQ for more details. 

#SBATCH --export=ALLRequired:  Exports all environment variables to the job.  See our FAQ for details.
#SBATCH --mail-user=YourEmailAddressRecommended:  Email address for job status messages.
#SBATCH --mail-type=ALLRecommended:  SLURM will notify the user via email when the job reaches the following states BEGIN, END, FAIL or REQUEUE.
#SBATCH --nodes=1 --exclusiveOptional:  Using both of these options will give your job exclusive access to a node such that no other jobs can share the node. 
This combination of arguments will assign eight tasks to your job and will give it exclusive access to all of the resources
(i.e. memory) of the entire node without interference from other jobs.  Please see our FAQ for more details on exclusive access.

#SBATCH --output=mypath

Optional:  The full path for the standard output (stdout) and standard error (stderr) "slurm-%j.out" file, where the "%j" is replaced by the job ID.  Current working directory is the default.

#SBATCH --error=mypath

Optional:  The full path for the standard error (stderr) "slurm-%j.out" files. Use this only when you want to separate (stderr) from (stdout). Current working directory is the default.

Serial Job Script

A job script may consist of SLURM directives, comments and executable statements. A SLURM directive provides a way of specifying job attributes in addition to the command line options. For example, we could create a myjob.slurm script this way:

myjob.slurm

This example script will submit a job to the default partition using 1 processor and 1GB of memory per processor, with a maximum run time of 30 minutes.

Definition of --ntasks-per-node

For the clusters the  --ntasks-per-node  option means  tasks per node.

Accurate run time value is strongly recommended

It is important to specify an accurate run time for your job in your SLURM submission script.  Selecting eight hours for jobs that are known to run for much less time may result in the job being delayed by the scheduler due to an overestimation of the time the job needs to run.

How to specify mem

The --mem value represents memory per processor core.  If your --mem value multiplied by the number of tasks (--ntasks-per-node) exceeds the amount of memory per node, your job will not run.  If your job is going to use the entire node, then you should use the --exclusive option instead of the --mem or --ntasks-per-node options (See Here).  It is good practice to specify the --mem option if you are going to be using less than an entire node and thus sharing the node with other jobs.

If you need to debug your program and want to run in interactive mode, the same request above could be constructed like this (via the srun command):

For more details on interactive jobs, please see our FAQ on this topic.

SLURM Environment Variables in Job Scripts

When you submit a job, it will inherit several environment variables that are automatically set by SLURM. These environment variables can be useful in your job submission scripts as seen in the examples above. A summary of the most important variables are presented in the table below.

Variable Name

Description

$SHARED_SCRATCH

Location of shared scratch space.  See our FAQ for more details.

$LOCAL_SCRATCHLocation of local scratch space on each node.

$SLURM_JOB_NODELIST

Environment variable containing a list of all nodes assigned to the job.

$SLURM_SUBMIT_DIR

Path from where the job was submitted.

Job Launcher (srun)

For jobs that need two or more processors and are compiled with MPI libraries, you must use srun to launch your job.  The job launcher's purpose is to spawn copies of your executable across the resources allocated to your job. We currently support srun for this task and do not support the mpirun or mpiexec launchers. By default srun only needs your executable, the rest of the information will be extracted from SLURM.

The following is an example of how to use srun inside your SLURM batch script. This example will run myMPIprogram as a parallel MPI code on all of the processors allocated to your job by SLURM:

myMPIjob.slurm

This example script will submit a job to the default partition using 24 processor cores and 1GB of memory per processor core, with a maximum run time of 30 minutes.

Your Program must use MPI

The above example assumes that myMPIprogram is a program designed to be parallel (using MPI). If your program has not been parallelized then running on more than one processor will not improve performance and will result in wasted processor time and could result in multiple copies of your program being executed.

The following example will run myMPIprogram on only four processors even if your batch script requested more than four.

To ensure that your job will be able to access an mpi runtime, you must load an mpi module before submitting your job as follows:

Submitting and Monitoring Jobs

Once your job script is ready, use sbatch to submit it as follows:

This will return a jobID number while the output and error stream of the job will be saved to one file inside the directory where the job was submitted, unless you specified otherwise.

The status of the job can be obtained using SLURM commands.  See the table below for a list of commands:

Command

Description

squeue

Show a detailed list of all submitted jobs.

squeue -j jobID

Show a detailed description of the job given by jobID.

squeue -- start -j jobID

Gives an estimate of the expected start time of the job given by jobID.

There are variations to these commands that can also be useful.  They are described below:

Command

Description

squeue -l

Show a list of all running jobs.

squeue -u username

Show a list of all jobs in queue owned by the user specified by username.

scontrol show job jobID

To get a verbose description of the job given by jobID. The output can be used as a template when you are attempting to modify a job.

There are many different states that a job can be after submission: BOOT_FAIL (BF), CANCELLED (CA), COMPLETED (CD), CONFIGURING (CF), COMPLETING (CG), FAILED (F), NODE_FAIL (NF), PENDING (PD), PREEMPTED (PR), RUNNING (R), SUSPENDED (S), TIMEOUT (TO), or SPECIAL_EXIT (SE). The squeue command with no arguments will list all jobs in their current state.  The most common states are described below.

Running (R): These are jobs that are running.

Pending (PD): These jobs are eligible to run but there is simply not enough resources to allocate to them at this time.

Deleting Jobs

A job can be deleted by using the scancel command as follows:

Compiling and Optimizing

Several programming models are supported on this system. Programs that are sequential, parallel or distributed can be run. Sequential programs require one processor to run. Parallel and distributed programs utilize multiple processors concurrently. Parallel programs are a subset of distributed programs. Generally speaking, distributed computing involve parametric sweeps, task farming, etc. Message passing, threaded applications generally fit under the scope of parallel computing.  SPMD (single process, multiple data) is one of the most popular method of parallelism, where a single executable works on its own data.

The supported compilers on this system are Intel, PGI, and GCC. OpenMPI implementations of Intel and GCC are available and can be loaded upon demand using the module command. The preferred compiler for this system is Intel.

Compiling Serial Code

To compile serial code you must first load the appropriate compiler environment module .  To load the Intel compiler, execute this command:

Once the environment is set, you can compile your program with one of the following (using Intel compiler as an example):

When invoked as described above, the compiler will perform the preprocessing, compilation, assembly and linking stages in a single step. The output file (or executable) is specified by executablename and the source code file is specificed by sourcecode.f77, for example. Omitting the -o executablename option will result in the executable being named a.out by default. For additional instructions and advanced options please view the online manual pages for each compiler (i.e. execute the command man ifort ).

Compiling Parallel Code

To compile a parallel version of your code that has OpenMPI library calls, use the appropriate OpenMPI library. Again, use module command to load the appropriate compiler environment as follows (Intel versions highly recommended):

module command

Description

module load GCC OpenMPI

For gcc compiled OpenMPI.

module load iccifort OpenMPI

For Intel compiled OpenMPI.

module load GCC impi

For gcc compiled Intel MPI.

module load iccifort impiFor Intel compiled Intel MPI.

 

To compile your code you will have use the OpenMPI scripts that are currently in your default path. The OpenMPI scripts are responsible for invoking the compiler, linking your program with the OpenMPI library and setting the OpenMPI include files.

Once the environment is set, you can compile your program with one of the following (assuming the Intel compiler as above):

When invoked as described above, the compiler will perform the preprocessing, compilation, assembly and linking stages in a single step. The output file (or executable) is specified by executablename and the source code file is specificed by mpi_sourcecode.f77, for example. Omitting the -o executablename option will result in the executable being named a.out by default. For additional instructions and advanced options please view the online manual pages for each compiler (i.e. execute the command man mpif77 ).

GNU Compiler

The GNU compiler is installed as part of the Red Hat Enterprise Linux distribution. Use man gcc to view the online manual for the C and C++ compiler, and man gfortran to view the online manual for the Fortran compiler.

Using the GPGPUs

If you want to build software using CUDA, you will need to load the CUDA toolkit into your default environment by using this command:

The GPUs are requested using the --gres=gpu flag.

For example, requesting a job with two nodes and two GPUs per node would look like this:

Two GPGPUs per node

There are two GPGPUs per compute node in the commons partition.  There are no GPGPUs in the serial partition.

To be able to use the GPUs you need to use either CUDA or OpenCL programming.  The following NVIDIA link has a lot of information and examples that you can download and try on the nodes:

http://developer.nvidia.com/gpu-computing-sdk

 

Getting Help

Request Help with the Center for Research Computing Resources

 

  • No labels