Understanding resource usage

This lecture helps you to optimize your resource usage in CSC’s HPC environment.

All materials (c) 2020-2025 by CSC – IT Center for Science Ltd. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 Unported License, http://creativecommons.org/licenses/by-sa/4.0/

Optimal usage on multi-user computing platforms

  • The computing resources are shared among hundreds of your colleagues, who all have different resource needs
  • Resources allocated to your job are not available for others to use
    • Important to request only the resources you need and ensure that the resources are used efficiently
  • Even if you can use more resources, should you?

One resource type will be a bottleneck

  • A single node can host many jobs from different users
  • Different jobs need different resources
  • Typically, the cores run out before the memory does
  • Sometimes a job uses only one core, but will consume all memory
    • No further jobs will fit in the node
    • If the job is not using the memory (just reserving it), resources are wasted

Slurm accounting: batch job resource usage 1/2

  • Resource usage can be queried with seff <slurm jobid>
  • Points to pay attention to:
    • Low CPU Efficiency:
      • Too many cores requested?
      • Cores waiting for other processes?
      • Cores waiting for data from disk?
      • Cores spread over too many nodes?
    • Low Memory Efficiency:
      • Too much memory requested?
      • Lots of caveats here
    • Low GPU efficiency:
      • Better to use CPUs? Disk I/O?

Slurm accounting: batch job resource usage 2/2

  • Not all usage is captured by Slurm accounting
    • If CPU efficiency seems too low, look at the completion time
    • Some applications also print timing data in log files
    • Jobs launched without srun don’t record properly (e.g. orterun)
  • More detailed queries can be tailored with sacct
    • sacct -j <slurm jobid> -o jobid,partition,state,elapsed,start,end
    • sacct -S 2025-01-01 will show all jobs started after that date
    • Note! Querying data from the Slurm accounting database with sacct can be a very heavy operation
      • Don’t query long time intervals or run sacct in a loop/using watch as this will degrade the performance of the system for all users

Billing Units

  • CPU time and storage space consume CPU and Storage Billing Units, respectively
  • BUs are a property of computing projects, not users
  • Monitor the BU usage of your project(s) from the command-line with csc-projects
    • For help/options, try csc-projects -h
  • Batch job billing scheme:
    • Amount of resources allocated: All requested resources are billed, i.e. number of cores, amount of memory, NVMe, …
    • Time allocated: Resources are billed based on the actual (wall) time a job has used, not the reserved maximum time

Applying for Billing Units

BUs are also a metric for comparing usage efficiency

  • Different resources have different rates
    • 1 CPU core hour on Puhti equals 1 CPU BU
    • 1 GPU card hour on Puhti equals 60 GPU BU (+ 1 GPU BU per allocated CPU core)
    • 1 CPU node hour on Mahti equals 100 CPU BU
    • 1 GiB hour of Memory on Puhti equals 0.1 CPU BU
    • 1st TiB of disk quota (/scratch, /projappl) is free-of-charge (0 Storage BU)
      • Applied excess quota is billed by 5 Storage BU/TiBh. (5 Storage Billing Units per TiB per hour)
    • 1 used TiB hour in Allas equals 1 Storage BU (i.e. 1 TiB of data consumes 8760 Storage BU per year)
    • This and other service billing information in Docs CSC

Before starting large-scale calculations

  • Check how the software and your actual input performs
    • Common job errors are caused by typos in batch/input scripts
  • Use short runs in the queue --partition=test to check that the input works and that the resource requests are interpreted correctly
  • Check the output of the seff command to ensure that CPU and memory efficiencies are as high as possible
    • It’s OK if a job is (occasionally) killed due to insufficient resource requests: just adjust and rerun/restart
    • It’s much worse to always run with excessively large requests “just in case”

Parallelizing your workflow

  • There are multiple ways to parallelize your workflow
    • Maybe several smaller jobs are better than one large (task farming)?
    • Is there a more efficient code or algorithm?
    • Is the file I/O slowing you down (lots of read/write operations)?
  • Optimize usage considering single job wall-time, overall used CPU time, I/O
  • Docs CSC: Guidelines for high-throughput computing

Reserving and optimizing batch job resources

seff examples

  • Left: GPU usage ok! (for this example other metrics also ok)
  • Bottom: CPU usage way too low, memory usage too high, job killed