Page Comparison

Short answer: Faculty can purchase priority access for 5 years if they pay the upfront cost for the nodes.

Long answer: High-priority access is available under a “condo model,” where faculty are able to purchase semi-dedicated nodes which get made available to all users when there are unused compute cycles. Under the condo model, faculty researchers fund the capital equipment costs of individual compute nodes, while the university funds the operating costs of running these nodes for five years. Faculty who purchase compute nodes receive access to equivalent resources at a higher priority than other researchers. The faculty can designate others to receive access at the same priority level, such as their graduate students, postdoctoral researchers, etc. With priority access, computational jobs are moved higher in the queuing system, and in most cases begin execution within twelve hours, depending upon other FairShare factors. A priority user can utilize their resources indefinitely. All access to resources is managed through the cluster’s job scheduler. When a priority user is not using their assigned resources, the nodes are made available to all UConn researchers for general use.

Please note that priority users will not be given separate partitions. Instead, they will be given a custom QoS because the QoS governs access to to priority resources (a.k.a. Trackable RESources, or TRes). 

If you are interested in investing in HPC resources, please fill out the HPC Condo Request form.

Short answer: First, you need to connect to UConn’s VPN. Then, you should be able to access the HPC.

Long Answer: The HPC Cluster only allows the connection of SSH from the campus-wide computers, for example:

• computers in the UConn libraries

• computers in campus offices/labs

• computers connected to UCONN-SECURE WiFi, etc.).

In order to To connect to the HPC when you are off campus, you will first need to connect to the UConn Virtual Private Network (VPN). After connecting to the VPN, you will be able to log in to the HPC as you normally do.

For instructions on how to install, set up, and connect your personal device(s) to UConn’s VPN, please go to this webpage.

The node you are on will normally be shown next to your netID when you log in to the Storrs HPC. For instance, if Jonathan the Husky’s netID was jtk10001jth10001, his terminal might look like this.

[jth10001@login6 ~]$

This would tell us that Jonathan is on the node called “login6.” Another way to check what node you are on is to use the hostname command. See below.

[jth10001@login6 ~]$ hostname
login6

Long wait times for your jobs? Errors about unavailable resources? We’ve been there and understand how frustrating it can be for jobs to take a long time to run. It’s an unfortunate consequence of having such a strong computational research community at UConn. LOTS of incredible research happens here, but it also means that there are LOTS of people competing for resources. There’s no getting around that problem, but there are a couple of steps we can take to increase the odds that our jobs get on ASAP.

1. Check what resources are available before we submit a job. (And then)

2. Target our submission to those available resources.

This FAQ will offer guidance on how to do both of those things.

Checking for Available Resources

The below sinfo nodeinfo command will give you a high-level view of what nodes are fully available (“idle”), in partial use (“mix”), not available (“alloc”. for allocated), or otherwise in need of maintenance (all other states besides idle, mix, or alloc). The nodes will be broken down in order by partition.

nodeinfo

The output for that command will look like this, but it will be much longer and provide info on every partition (not just class and debug).

PARTITION      NODES  STATE  TIMELIMIT   CPUS    GRES MEMORY   ACTIVE_FEATURES   NODELIST
class              1  inval    4:00:00     36   gpu:3 191954   gpu,v100,skylake  gpu05
class              1 maint*    4:00:00    128  (null) 515404   epyc128,cpuonly   cn563
class              3 drain$    4:00:00     36  (null) 191845+  skylake,cpuonly   cn[360-361,363]
class              8  maint    4:00:00     36  (null) 191845+  skylake,cpuonly   cn[362,365-366,368-372]
class              1  maint    4:00:00    128  (null) 515404   epyc128,cpuonly   cn561
class              1  comp*    4:00:00    128  (null) 515404   epyc128,cpuonly   cn564
class              3   drng    4:00:00    128  (null) 515404   epyc128,cpuonly   cn[560,565,569]
class              6  alloc    4:00:00    128  (null) 515404   epyc128,cpuonly   cn[559,562,566-568,570]
debug              1 maint*      30:00    128  (null) 515404   epyc128,cpuonly   cn563

... (etc.)

The above command gives us an overarching picture of usage on the cluster, and from there, we can use a more targeted command to get more information on individual nodes within a partition, like how many cores or GPUs are in use and how many are available. The base sinfo command is below and it targets the priority-gpu partition but we could amend it to target any other partition, like hi-core for example.

sinfo -p priority-gpu -t idle,mix -o%10n%20C%10G%10t%20R%b
HOSTNAMES CPUS(A/I/O/T)       GRES      STATE     PARTITION           ACTIVE_FEATURES
gpu06     13/23/0/36          gpu:1     mix       priority-gpu        location=local,gpu,v100,skylake
gpu20     36/28/0/64          gpu:3     mix       priority-gpu        location=local,epyc64,a100,gpu
gpu21     25/39/0/64          gpu:3     mix       priority-gpu        location=local,epyc64,a100,gpu
gpu22     3/61/0/64           gpu:3     mix       priority-gpu        location=local,epyc64,a100,gpu
gpu23     33/31/0/64          gpu:1     mix       priority-gpu        location=local,epyc64,a100,gpu
gtx02     2/18/0/20           gpu:3     mix       priority-gpu        location=local,gpu,gtx,skylake
gtx03     12/8/0/20           gpu:3     mix       priority-gpu        location=local,gpu,gtx,skylake
gtx08     1/19/0/20           gpu:2     mix       priority-gpu        location=local,gpu,gtx,skylake
gtx11     2/18/0/20           gpu:2     mix       priority-gpu        location=local,gpu,gtx,skylake
gtx15     28/4/0/32           gpu:8     mix       priority-gpu        location=local,gpu,rtx,skylake
gtx09     0/20/0/20           gpu:2     idle      priority-gpu        location=local,gpu,gtx,skylake

The column titled “CPUS (A/I/O/T)” tells us how many cores are available. “A” stands for Allocated, “I” stands for Idle, and “T” stands for Total. (“O” stands for Other but you can ignore that column) Since there are 39 cores in the “Idle” column for GPU21, that means 39 cores are available to use. But all 3 of the GPUs on GPU21 are in use so we can’t use any GPUs on that node. So, that gives us an idea of the resources. If I only needed cores and no GPUs, I could target GPU21.

In summary, these two commands can give us a picture of what partitions have resources available, and then what resources are available on individual nodes within that partition.

Targeting a specific partition

The next step is submitting a job targeting a specific partition. If you’re not sure how to target a specific partition, please visit our SLURM Guide where you will see examples of submission scripts that target different partitions and architectures. Another key part of targeting a specific partition is knowing what partitions you are allowed to use and what “account” and “QOS” you must use to access them. To check what partitions you’re allowed to use and how to access them you can use this command.

acctinfo`whoami`
      User    Account            Partition                  QOS 
---------- ---------- -------------------- -------------------- 
net10004    jth10001               hi-core              general 
net10004    jth10001           general-gpu              general 
net10004    jth10001                 debug              general 
net10004    jth10001               lo-core              general 
net10004    jth10001               general              general
net10004    jth10001          priority-gpu          jth10001gpu

This tells me that I have access to 6 partitions. To access the priority-gpu partition, I need to include the below three flags in the #SBATCH header of my submission script. This will be different for every individual so you will have to modify this with the partitions you have access to and the account and QOS that are associated with your account.

#SBATCH -p priority-gpu     # partition I want to access
#SBATCH -A jth10001         # account I am associated with
#SBATCH -q jth10001gpu      # QOS needed to access partition

If you have further questions about how to check what resources are available and how to target them, please feel free to contact the Storrs HPC admins by sending an email to hpc@uconn.edu.

There are many reasons a job may fail. A good first step is to use the shist command to check the ExitCode SLURM gave it. The command follows this format: shist --starttime YYYY-MM-DD.

Here’s an example of the output for a job that failed immediately with an ExitCode of 1.

JobID         Partition        QOS    JobName      User      State    Elapsed   NNodes      NCPUS        NodeList ExitCode                 End 
------------ ---------- ---------- ---------- --------- ---------- ---------- -------- ---------- --------------- -------- -------------------
73088        priority-+ erm12009g+  submit_rx  jdt10005     FAILED   00:00:00        1         32           gtx21      1:0 2022-11-17T10:05:34 

The ExitCode of a job will be a number between 0 and 255. An ExitCode of 0 means that—as far as SLURM is concerned—the job ran and was completed properly. Any non-zero ExitCode will indicate that the job failed. One could then search the ExitCode on Google to investigate what SLURM thinks caused the job to fail. Sometimes this is helpful but not always. Either way, take note of what you find for future reference. Common ExitCodes are listed below for reference:

• 0 → success

• non-zero → failure

• Exit code 1 indicates a general failure

• Exit code 2 indicates incorrect use of shell builtins

• Exit codes 3-124 indicate some error in job (check software exit codes)

• Exit code 125 indicates out of memory

• Exit code 126 indicates command cannot execute

• Exit code 127 indicates command not found

• Exit code 128 indicates invalid argument to exit

• Exit codes 129-192 indicate jobs terminated by Linux signals

  • For these, subtract 128 from the number and match to signal code

  • Enter kill -l to list signal codes

  • Enter man signal for more information

  • Please note: when a signal was responsible for a job or step's termination, the signal number will be displayed after the exit code, delineated by a colon(:).

The next clue to investigate is the NodeList column. Sometimes a job fails because there is something wrong with the compute node our job was run on. If the compute node is the problem (and Storrs HPC staff haven’t fixed it already), the job should fail again with the same ExitCode. We can submit our job specifically to that same node to see if the job fails again. Try adding this to the #SBATCH header of your script to target a specific node. Here, we target gtx21 because that was the node listed in the NodeList column above.

#SBATCH --nodelist=gtx21

Once you see the job has failed multiple times on the same node but does not fail on other nodes, then you can feel confident that a faulty node is a likely the cause. Please submit a help request to Storrs HPC including a screenshot from the shist output.

Short answer: If you received this error, your job most likely failed because the amount of memory (RAM) it needed was larger than the default. We can request more memory for your job using the --mem or --mem-per-cpu flag.

Long answer: There are several reasons a job may fail from insufficient memory. As of January 2023, the default amount of memory available per CPU is 2 gigabytes. The default mem-per-cpu of 2 GB is okay for many users, but not for all. Users who receive OOM errors need to request more memory when submitting or initiating jobs. Those users can easily override the default using the -mem-per-cpu flag in your submission script. The new line would look something like this:

#SBATCH -n 2                            # asks for 2 cores
#SBATCH --mem-per-cpu=3G                # asks for 3 GB or RAM per core, or 6  GB total

Adding this line will tell SLURM to use 3 gigabytes of RAM per CPU you request. That means if we ask for 2 cores (-n 2), then we’ll be allocated 6 gigabytes of RAM total. Please note that the --mem-per-cpu flag must be used with the -n flag specifying the number of cores you want. Alternatively, users can use the --mem flag to specify the total amount of RAM they need regardless of how many CPUs are rquestedrequested.

#SBATCH --mem=5G                        # asks for 10 gigabytes total

We encourage users to please adjust the --mem-per-cpu or --mem flags in a step-wise fashion. First, we try 3 gigabytes, then 4, then 5, etc. until our jobs start running without failing from memory errors. That strategy helps ensure that every HPC user's jobs get on quickly and run efficiently. For more info on fisbatch, srun, and #SBATCH flags, see this link.

Short answer: Once the job is cancelled canceled by SLURM due to timeout, it cannot be resumed from that point because SLURM sets the exit code to “0” which denotes job completion. As far as SLURM is concerned, the job is now complete, with no state to resume from.

Long answer: One thing you can try is to use the timeout command to stop your program just before SLURM does. You can tell from the return code if the timeout was reached or not. It should set exit code “124”. If so, you can then requeue it with scontrol. Try the following:

In your submission script, add the following:

#SBATCH --open-mode=append
#SBATCH --time=12:00:00

Then, use the timeout command to call your program:

timeout 11h ./your_program
if [[ $? == 124 ]]; then
  scontrol requeue $SLURM_JOB_ID
fi

Note: The --open-mode=append ensures the output of each run is appended to the file specified by #SBATCH --output= to preserve the previous run’s output in the same file.

Disclaimer: This is untested on Storrs HPC; however, it should work as long as everything else is working correctly.