Resources and shares
Methodology to calculate core hours and GPU hours
Formula for calculating available core hours per year:
NumberOfNodes * NumberOfCPUsPerNode * NumberOfCoresPerCPU * NumberOfThreadsPerCore * 24 hours per day * 365 days
All HPC Vega nodes have two CPUs per node, so NumberOfCPUsPerNode is set to two. All CPUs on HPC Vega have two CPU threads, but NumberOfThreadsPerCore in formula is set to one, although Hyper-threading (HT) on the system is turned On (so it should be two).
Slurm always reserves the entire core for job. If the user requests one CPU, Slurm allocates the whole core with two threads. However, if e.g. the user requests 64 cores, he gets 32 cores or 64 CPU threads.
Formula for calculating available GPU hours per year:
NumberOfGPUNodes * GPUsPerGPUnode * 24 hours per day * 365 days
All available resources in Core/GPU hours
Available resources in core hours per year (without HT) are listed in the following table:
| Partition | Nodes (c/CPU) | Cores | Node hours | Core hours | GPUs | GPU hours |
|---|---|---|---|---|---|---|
| CPU partition | 960 (64) | 122.880 | 8.409.600 | 1.076.428.800 | N/A | N/A |
| CPU standard (256 GB) | 768 (64) | 98.304 | 6.727.680 | 861.143.040 | N/A | N/A |
| CPU large memory (1 TB) | 192 (64) | 24.576 | 1.681.920 | 215.285.760 | N/A | N/A |
| CPU login nodes | 4 (64) | 512 | 35.040 | 4.485.120 | N/A | N/A |
| GPU partition (CPUs) | 60 (64) | 7.680 | 525.600 | 67.276.800 | 240 | 2.102.400 |
| GPU login nodes | 4 (32) | 184 | 17.520 | 2.242.560 | 4 | 8.764 |
| SUM | 1.028 | 131.256 | 8.935.200 | 1.143.705.600 | 244 | 2.111.164 |
Usage shares of HPC Vega
For the HPC Vega, the following shares of usage are intended on a yearly basis:
- 65% of all capacities – national share, divided:
- up to 70% for Open Access
- up to 20% for Industry Access
- from 10% upwards is Hosting Entity (IZUM) reservation for:
- Community Access
- Urgent Priority Usage (e.g. priority use of national importance)
- Continuous Maintenance (Updates/Upgrades/Repairs)
- 35% of all capacities – EuroHPC JU share:
- Users and projects approved by EuroHPC JU
Shares presented in number of nodes1, core and GPU hours per year and storage2 (Ceph):
| Share | Partition | Nodes | Node hours | Core hours | GPU hours | Storage |
|---|---|---|---|---|---|---|
| National (65%) | CPU standard | 499 | 4.372.992 | 559.742.976 | N/A | 11,7 PB |
| CPU large memory | 125 | 1.095.000 | 140.160.000 | N/A | \ | |
| GPU partition | 39 | 341.640 | 43.729.920 | 1.366.560 | \ | |
| EuroHPC JU (35%) | CPU standard | 269 | 2.354.688 | 301.400.064 | N/A | 6,3 PB |
| CPU large memory | 67 | 586.920 | 75.125.760 | N/A | \ | |
| GPU partition | 21 | 183.960 | 23.546.880 | 735.840 | \ |