Karst at Indiana University

On this page:

Note:
For help determining which of IU's research supercomputing systems is best suited to meet your needs, use the UITS Supercomputing Pathfinder.

System overview

Karst at Indiana University is a high-throughput computing cluster designed to deliver large amounts of processing capacity over long periods of time. Karst provides batch processing and node-level co-location services that make it well suited for running high-throughput and data-intensive parallel computing jobs.

Besides being available to IU students, faculty, and staff for standard, cluster-based, high-throughput computing, Karst offers two alternative service models to the IU community:

  • Condominium computing: The condominium computing service model provides a way for IU schools, departments, and researchers to fund computational nodes for their own research purposes without shouldering the cost, overhead, and management requirements of purchasing individual systems. Condominium nodes are housed in the IU Bloomington Data Center, and are managed, backed up, and secured by UITS Research Technologies staff. Condominium nodes are available to "members" whenever they are needed, but when they are not in use, idle condominium nodes become available to other researchers and students on Karst. In this way, condominium computing promotes cost-effective expansion of IU's high-performance computing capabilities, enables efficient provisioning of computing resources to the entire IU research community, and helps conserve natural resources and energy.
  • Dedicated computing: The dedicated computing service model lets schools and departments host nodes that are dedicated solely to their use within Karst's physical, electrical, and network framework. This provides 24/7 access for school or departmental use, while leveraging the network and physical components of Karst, and the security and energy efficiency benefits provided by location within the IU Data Center.

Karst's system architecture provides the advanced performance needed to accommodate high-end, data-intensive applications critical to scientific discovery and innovation. As configured upon entering production, Karst comprised 228 general-access compute nodes, 28 condominium nodes, and 16 dedicated data nodes (for separate handling of data-intensive operations). Each node is an IBM NeXtScale nx360 M4 server equipped with two Intel Xeon E5-2650 v2 8-core processors. Compute nodes have 32 GB of RAM and 250 GB of local disk storage. Data nodes have 64 GB of RAM and 24 TB of local storage. All nodes are housed in the IU Bloomington Data Center, run Red Hat Enterprise Linux (RHEL) 6, and are connected via 10-gigabit Ethernet to the IU Science DMZ.

System access

Access is available to IU students, faculty, staff, and sponsored affiliates. For details, see the "Research system accounts (all campuses)" section of What computing accounts are available at IU, and for whom?

Once your account is created, you can user your IU username and passphrase to log into Karst (karst.uits.iu.edu) with any SSH2 client. Public key authentication also is permitted; see How do I set up SSH public-key authentication to connect to a remote system?

Alternatively, if you are unaccustomed to working in Linux command-line environments, you can use the Karst Desktop graphical remote desktop application, which lets you work on Karst from a desktop window running on your personal computer; for more, see At IU, what is Karst Desktop?

Note:
The scheduled monthly maintenance window for Karst is the first Tuesday of each month, 7am-7pm.

Available software

For a list of packages available on Karst, see Karst Modules in the IU Cyberinfrastructure Gateway.

Karst users can request software using the HPC software request form.

Setting up your user environment

On the research computing resources at Indiana University, the Modules environment management system provides a convenient method for dynamically customizing your software environment.

For more about using Modules to configure your user environment, see On the research computing systems at IU, how do I use Modules to manage my software environment?

File storage options

For file storage information, see On IU's research systems, how much allocated and short-term storage capacity is available to me?

Note:
Before storing data on this system, make sure you understand the information in the Working with data containing PHI section (below).

Working with data containing PHI

The Health Insurance Portability and Accountability Act of 1996 (HIPAA) established rules protecting the privacy and security of individually identifiable health information. The HIPAA Privacy Rule and Security Rule set national standards requiring organizations and individuals to implement certain administrative, physical, and technical safeguards to maintain the confidentiality, integrity, and availability of protected health information (PHI).

This system meets certain requirements established in the HIPAA Security Rule that enable its use for research involving data that contain protected health information (PHI). You may use this resource for research involving data that contain PHI only if you institute additional administrative, physical, and technical safeguards that complement those UITS already has in place. For more, see When using UITS Research Technologies systems and services, what are my legal responsibilities for protecting the privacy and security of data containing protected health information (PHI)? If you need help or have questions, contact UITS HIPAA Consulting.

Note:
Although PHI is one type of institutional data classified as Critical at IU, other types of institutional data classified as Critical are not permitted on Research Technologies systems. Except for PHI, the most sensitive classification of institutional data allowed on Research Technologies resources is Restricted. For help determining which institutional data elements classified as Critical are considered PHI, see Which data elements in the classifications of institutional data are considered protected health information (PHI)?

Running jobs on Karst

IU's research computing clusters use the TORQUE resource manager (based on OpenPBS) and the Moab Workload Manager to manage and schedule batch jobs. Moab uses fairshare scheduling to track usage and prioritize jobs.

The default wall time for jobs running on Karst compute nodes is 30 minutes; the default virtual memory per job is 8 MB.

User processes on the login nodes are limited to 20 minutes of CPU time. Processes on the login nodes that run longer than 20 minutes are terminated automatically (without warning). If your application requires more than 20 minutes of CPU time, submit a batch job or an interactive session using the TORQUE qsub command.

Because of this limit:

  • When running Java programs, add the -Xmx parameter (values must be multiples of 1,024 greater than 2 MB) on the command line to specify the Java Virtual Machine (JVM) maximum heap size. For example, to run a Java program (e.g., Hello_DeathStar) with a maximum heap size of 640 MB , on the command line, enter:
  •   java -Xmx640m Hello_DeathStar
    
  • You should perform any debugging or testing by running interactive jobs on the compute nodes. Specify the DEBUG queue for short, quick-turnaround test jobs (maximum wall time: 30 minutes) or the INTERACTIVE queue for longer jobs (maximum wall time: 8 hours).

Submitting jobs

Use the TORQUE qsub command to submit non-interactive or interactive batch jobs for execution on Karst's compute nodes:

  • Non-interactive jobs: To run a job in batch mode on Karst, first prepare a TORQUE job script that specifies the application you want to run and the resources required to run it, and then submit it to TORQUE with the qsub command.
  • Do not specify a destination queue in your job script or your qsub command. TORQUE passes your job to the system's default routing queue (BATCH) for placement, based on its resource requirements, in the SERIAL, NORMAL, or LONG execution queue. From there, your job dispatches whenever the required resources become available.

    If your job has resource requirements that are different from the defaults (but not exceeding the maximums allowed), specify them either with TORQUE directives in your job script, or with the -l (a lower-case "L"; short for resource_list) option in your qsub command.

    Note:
    Command-line options override PBS directives in your job script.

    On the command line, you can specify multiple attributes with either one -l switch followed by multiple comma-separated attributes, or multiple -l switches, one for each attribute. For example, to submit a job (e.g., death_star.script) that requires 24 hours of wall time (instead of the default 30 minutes) and 100 GB of virtual memory (instead of the default 8 MB) to run on four cores on one node, you may enter either of the following commands (they are equivalent):

      qsub -l nodes=1:ppn=4,vmem=10gb,walltime=24:00:00 death_star.script
      qsub -l nodes=1:ppn=4 -l vmem=10gb -l walltime=24:00:00 death_star.script
    
    Note:
    Each Karst compute node has 16 CPU cores available, allowing you to "stack" up to 16 jobs on one compute node (provided none of the jobs has unique resource constraints, such as special memory requirements). To run up to 16 jobs "stacked" on a single compute node, specify -l nodes=1:ppn=1 in the script or qsub command for each job.
  • Interactive jobs: To submit an interactive job, use qsub with the -I (to specify an interactive job) and -q interactive (to specify submission to the INTERACTIVE batch queue) switches; for example:
  •   qsub -I -q interactive -l nodes=1:ppn=4,vmem=10gb,walltime=4:00:00
    

    Submitting your job to the INTERACTIVE queue directs it to a specific set of nodes that are configured for shared access (versus single-user in the general batch queues). Consequently, your interactive job most likely will dispatch faster in the INTERACTIVE queue than in the general execution queues.

Useful qsub options include:

Option Function
-a <date_time> Execute the job only after specified date and time (<date_time>).
-I Run the job interactively. (Interactive jobs are forced to be not re-runnable.)
-m e Mail a job summary report when the job terminates.
-q <queue_name> Specify the destination queue (<queue_name>) for the job. (On Karst, use this only for submitting jobs to the DEBUG, INTERACTIVE, or PREEMPT queues.)
-r [y|n] Declare whether the job is re-runnable. Use the argument n if the job is not re-runnable. The default value is y (re-runnable).
-V Export all environment variables in your current environment to the job.

For more, see the qsub manual page.

Monitoring jobs

To monitor the status of a queued or running job, use the TORQUE qstat command.

Useful qstat options include:

Option Action
-a Display all jobs.
-f Write a full status display to standard output.
-n List the nodes allocated to a job.
-r Display jobs that are running.
-u user1,user2 Display jobs owned by specified users.

For more, see the qstat manual page.

Deleting jobs

To delete a queued or running job, use the qdel command.

Note:
Occasionally, a node will become unresponsive and unable to respond to the TORQUE server's requests to kill a job. In such cases, try using qdel -W <delay> to override the delay between SIGTERM and SIGKILL signals; for <delay>, specify a value in seconds.

For more, see the qdel manual page.

Queue information

Karst employs a default routing queue that funnels jobs, according to their resource requirements, into three execution queues configured to maximize job throughput and minimize wait times (i.e., the amount of time a job remains queued, waiting for required resources to become available). Depending on the resource requirements specified in either your batch job script or your qsub command, the routing queue (BATCH) automatically places your job into the SERIAL, NORMAL, or LONG queue.

Note:
The maximum wall time allowed for jobs running on Karst is 14 days. If your job requires more than 14 days of wall time, email the High Performance Systems group for assistance.

You do not have to specify a queue in your job script or in your qsub command to submit your job to one of the three batch execution queues; your job will run in the SERIAL, NORMAL, or LONG queue unless you specifically submit it to the DEBUG, PREEMPT, or INTERACTIVE queue, the properties of which are as follows:

  • DEBUG: The DEBUG queue is intended for short, quick-turnaround test jobs requiring less than 1 hour of wall time.
  • Maximum wall time: 1 hour
    Maximum nodes per job: 4
    Maximum cores per job: 64
    Maximum number of jobs per queue: None
    Maximum number of jobs per user: 2
    Direct submission: Yes

    To submit a batch job to the DEBUG queue, either add the #PBS -q debug directive to your job script, or enter qsub -q debug on the command line.

    Note:
    For longer debugging or testing sessions, submit an interactive job to the INTERACTIVE queue instead.
  • INTERACTIVE: Interactive jobs submitted to the INTERACTIVE queue should experience less wait time (i.e., start sooner) than interactive jobs submitted to the batch execution queues.
  • Maximum wall time: 8 hours
    Maximum nodes per job: None
    Maximum cores per job: 8
    Maximum number of jobs per queue: 128
    Maximum number of jobs per user: 16
    Direct submission: Yes

    To submit an interactive job to the INTERACTIVE queue, on the command line, enter qsub with the -I and -q interactive options added; for example:

      qsub -I -q interactive -l nodes=1:ppn=1,walltime=4:00:00
    
    Note:
    If you enter qsub without the -q interactive option, your interactive job will be placed in the routing queue for submission to the SERIAL, NORMAL, or LONG batch execution queue, which most likely will entail a longer wait time for your job.
  • PREEMPT: Jobs submitted to the PREEMPT queue run on "condominium nodes" owned by members of the "condominium computing" service; however, when a job submitted by a "condominium node" owner is ready to dispatch (and no other nodes are available), the non-condominium job with the lowest accrued wall time will be preempted. Consequently, non-condominium jobs in the PREEMPT queue may dispatch multiple times before running to completion.
  • Maximum wall time: 14 days
    Maximum nodes per job: None
    Maximum cores per job: None
    Maximum number of jobs per queue: 1,800
    Maximum number of jobs per user: 200
    Direct submission: Yes

    To submit a job to the PREEMPT queue, add the #PBS -q preempt directive to your job script, or enter qsub -q preempt on the command line.

To see current status information for the work queues on Karst, on the command line, enter:

  qstat -q
Note:
To best meet the needs of all research projects affiliated with Indiana University, the High Performance Systems (HPS) team administers the batch job queues on UITS Research Technologies supercomputers using resource management and job scheduling policies that optimize the overall efficiency and performance of workloads on those systems. If the structure or configuration of the batch queues on any of IU's supercomputing systems does not meet the needs of your research project, fill out and submit the Research Technologies Ask RT for Help form (for "Select a group to contact", select High Performance Systems).

Requesting single user time

Although UITS Research Technologies cannot provide dedicated access to an entire compute system during the course of normal operations, "single user time" is made available by request one day a month during each system's regularly scheduled maintenance window to accommodate IU researchers with tasks requiring dedicated access to an entire compute system. To request such single user time, complete and submit the Research Technologies Ask RT for Help form, requesting to run jobs in single user time on HPS systems. If you have questions, email the HPS team.

Acknowledging grant support

The Indiana University cyberinfrastructure, managed by the Research Technologies division of UITS, is supported by funding from several grants, each of which requires you to acknowledge its support in all presentations and published works stemming from research it has helped to fund. Conscientious acknowledgment of support from past grants also enhances the chances of IU's research community securing funding from grants in the future. For the acknowledgment statement(s) required for scholarly printed works, web pages, talks, online publications, and other presentations that make use of this and/or other grant-funded systems at IU, see If I use IU's research cyberinfrastructure, what sources of funding do I need to acknowledge in my published work?

Support

Support for IU research computing systems, software, and services is provided by various UITS Research Technologies units. For help, see Research computing support at IU.

This is document bezu in the Knowledge Base.
Last modified on 2017-08-07 17:00:11.

  • Fill out this form to submit your issue to the UITS Support Center.
  • Please note that you must be affiliated with Indiana University to receive support.
  • All fields are required.

Please provide your IU email address. If you currently have a problem receiving email at your IU account, enter an alternate email address.

  • Fill out this form to submit your comment to the IU Knowledge Base.
  • If you are affiliated with Indiana University and need help with a computing problem, please use the I need help with a computing problem section above, or contact your campus Support Center.

Please provide your IU email address. If you currently have a problem receiving email at your IU account, enter an alternate email address.