ALTERNATIVE

Best Per-User Historical Averages (sacct) Alternative

SLURM accounting database historical averages for resource prediction

What is Per-User Historical Averages (sacct)?

State-of-the-art approach using SLURM's sacct (accounting database) to track per-user historical resource consumption and apply those averages to new job submissions. Works well for repeated workload patterns but fails when new workload types or code changes are introduced.

✅ What Per-User Historical Averages (sacct) does well

  • • Simple to implement
  • • No external dependencies
  • • Works for repeated workloads

❌ Limitations for Agents

  • • Becomes wildly inaccurate with new workload types
  • • Fails when code-level changes are made
  • • Cannot predict resource needs for novel tasks
  • • No failure prediction capability

Why AI Agents are replacing Per-User Historical Averages (sacct)

Expanse uses multimodal deep learning (source code + hardware telemetry + cluster metadata) to predict resources accurately even for new workloads, outperforming historical averages by 8x

Common Use Cases

GPU cluster resource allocationHPC job schedulingKubernetes workload prediction