Priority Policy

GridUnesp uses a priority system based on Fair Share to ensure balanced use of computational resources among all users and research groups.

Fair Share Concept

Fair Share works based on the following principles:

  • Each research group is entitled to a fair portion of the resources

  • Each user within the group is entitled to a fair portion

  • Priority decreases as the user/group uses more resources

  • Priority is gradually “reset” over time (14-day half-life)

Simplified formula:

  • Group: 1/N of the resources, where N = total number of groups

  • User: 1/M of the group’s resources, where M = number of users in the group

Viewing Fair Share

sshare Command

The sshare command shows the current Fair Share status of all users:

sshare -a

Example output:

Account           User  RawShares  NormShares    RawUsage  EffectvUsage  FairShare
------------ --------- ---------- ----------- ----------- ------------- ----------
   startrek                    10    0.008065    40079089      0.008656
    startrek      kirk          1    0.333333     1761930      0.043961   0.220126
    startrek    picard          1    0.333333    38237843      0.954060   0.219227
    startrek     spock          1    0.333333       79315      0.001979   0.221024
   hogwarts                    10    0.008065     8028212      0.001734
    hogwarts    potter          1    0.250000           0      0.000000   0.278527
    hogwarts  hermione          1    0.250000     8028212      1.000000   0.275831
    hogwarts voldemort          1    0.250000           0      0.000000   0.278527
    hogwarts    malfoy          1    0.250000           0      0.000000   0.278527

Interpretation of the columns:

  • Account: Group name

  • User: User name

  • RawShares: Assigned weights (the higher, the more resources)

  • NormShares: Normalized share

  • RawUsage: Recent cumulative usage

  • EffectvUsage: Effective usage (0-1)

  • FairShare: Current priority (the higher, the better)

Filtering the View

To see only your group:

sshare -a | grep group_name

To see only your user:

sshare -a | grep $USER

How Priority Affects Jobs

The priority calculated by Fair Share directly influences the execution order of jobs in the queue.

Listing Jobs in Priority Order

View pending jobs sorted by priority
squeue -o "%.18i %.9Q %.8j %.8u %.10V %.6D %R" --sort=-p,i --states=PD

Example output:

 JOBID       PRIORITY      NAME     USER  SUBMIT_TIM  NODES NODELIST(REASON)
   2847404     296816  force1.s     yoda  2022-03-31      1 (Priority)
2838783_12     222821  dark1.sh    vader  2022-03-15      1 (Priority)
2838783_14     222821  dark1.sh    vader  2022-03-15      1 (Priority)
   2844502     185817  crucio1. valdemo+  2022-03-28      1 (Priority)
   2848555      50480  vulcano.    spock  2022-04-01      4 (Priority)

Notes:

  • The PRIORITY column shows the numeric priority value

  • Jobs with higher priority (larger number) run first

  • Priority is dynamic and changes as:

    • New jobs are submitted

    • Running jobs finish

    • Time passes (14-day half-life)

Factors That Influence Priority

  1. Recent cluster usage

    • The more you used in the last 14 days, the lower your priority

    • Every 14 days, the “counter” is gradually reset

  2. Number of users in the group

    • Groups with many users have smaller slices per person

    • Encourages groups to manage usage internally

  3. Job size

    • Jobs that request many resources may have their priority adjusted

    • The system seeks a balance between large and small jobs

  4. Wait time

    • Jobs that wait a long time gradually gain priority

    • Prevents starvation (jobs that never run)

    • However, jobs that require a short processing time have a greater chance of obtaining resources more quickly.

Tips to Improve Your Priority

✅ Do:

  1. Request only what you need

    #SBATCH -t 02:00:00     # If your job runs in 1h30, don't request 24h
#SBATCH --mem=8G        # If it uses 6G, don't request 64G

  2. Distribute usage over time

    • Avoid submitting hundreds of jobs at once

    • Spread submissions over days/weeks

  3. Use job arrays for many small jobs

    #SBATCH --array=1-1000   # 1000 small jobs
#SBATCH --ntasks=1
#SBATCH --time=01:00:00

  4. Monitor your Fair Share regularly

    sshare -U $USER

❌ Avoid:

  • Monopolizing resources with very large and long jobs

  • Requesting resources far above what is needed

  • Leaving jobs running that could have already finished

Understanding the 14-Day Half-Life

The system uses an exponential decay with a 14-day half-life:

  • Day 0: Maximum usage (minimum priority)

  • Day 14: The “weight” of the usage drops by half

  • Day 28: The “weight” drops to 1/4 of the original value

  • And so on…

This means that:

  • If you used many resources today, your priority tends toward the initial value as the days pass

  • The system gradually “forgets” old usage

  • Everyone has a chance to recover priority after periods of inactivity

Practical Example

Let’s follow a user over time:

Week 1: João submits few jobs: high priority (0.9)

sshare -U joao
# FairShare = 0.901

Week 2: João submits many jobs: priority drops

sshare -U joao
# FairShare = 0.452

Week 3: João keeps using a lot: very low priority

sshare -U joao
# FairShare = 0.187

Week 4: João reduces usage: priority starts to recover

sshare -U joao
# FairShare = 0.223

Weeks 5-6: With moderate usage: priority tends toward the initial value

sshare -u joao
# FairShare = 0.854

Command Summary

Commands for priority management

Command

Description

sshare -a

View Fair Share of all users

sshare -U $USER

View only your Fair Share

sshare -a | grep group

View your group’s Fair Share

squeue -U $USER --sort=-p

View your jobs sorted by decreasing priority (-p)

squeue -o "%.18i %.9Q %.8j %.8u %.10V %.6D %R" --sort=-p,i --states=PD

List pending jobs in order of decreasing priority (-p)

See also