IDEAL Lab Euler Share Policy

This page distinguishes IDEAL Lab policy from ETH-wide Euler behavior.

ETH-Wide Facts#

• Euler is shared and schedules non-exclusive jobs on compute nodes.
• The public share allows small CPU jobs, currently up to 48 CPU cores and 128 GiB RAM. It contains no GPUs and may have long waits.
• GPU nodes are available only through shareholder accounts.
• A shareholder entitlement is an average fair share, not exclusive ownership of named physical nodes.
• Users should request resources accurately and normally avoid specifying a partition or hardware model unless the workload depends on it.

Current es_fuge Entitlement#

A read-only my_share_info check on 2026-08-04 reported:

• 784 CPU cores;
• 3136 GiB system RAM;
• 18 GPUs.

The current 18-GPU lab allocation includes two Nvidia RTX PRO 6000 GPUs. RTX 4090 remains the standard student GPU; the RTX PRO 6000 is a special-purpose 96-GiB, CUDA 13-compatible option requested with Slurm identifier pro_6000.

These numbers describe the current shareholder entitlement, not resources that are always idle or exclusively reserved for IDEAL Lab users. Recheck with my_share_info before publishing or relying on these totals.

IDEAL Lab Student Policy#

This subsection is local lab policy for supervised BSc/MSc use of es_fuge:

• Scheduler priority factor: 0.2 relative to regular staff within the lab share.
• Per student: no per-student cap is currently configured.
• Student group collectively: at most 14 GPUs, 1280 CPU cores, and 5120 GiB system RAM.

The collective limits are scheduling ceilings, not reserved capacity or recommended requests. RAM here means system RAM, not GPU VRAM; GPU jobs consume CPU cores and system RAM as well as GPU capacity. Most first jobs should use one CPU, modest memory, no GPU, and a short time limit.

Required Behavior#

• Use --account=es_fuge for GPU jobs and supervised lab-share work.
• Use --account=public only for CPU work that fits public-share limits.
• Never add --gpus to a public job.
• Prefer sbatch for training, sweeps, and unattended computation.
• Keep euler-tunnel allocations short and stop them when interactive work ends.
• Start arrays with a low explicit concurrency cap and count concurrency across every active array and ordinary job, not only within the newest array.
• Inspect completed jobs with sacct and seff; use the evidence to right-size the next request.
• Start with one GPU. Request multiple GPUs only after measuring useful scaling.

Exceptions#

A project blocked by the collective student limits must obtain supervisor approval before requesting a change or temporary exception. The supervisor and lab IT administrator should record its scope, reason, owner, and end date.

Sources#

• ETH HPC shareholders
• ETH HPC Slurm guide
• ETH HPC GPU nodes