Accounting on Snellius
Last updated on 2026-10-27 | Edit this page
Overview
Questions
- What is a system billing unit, and which row of the table sets the price?
- How many SBU does 16 cores for 2 hours cost on
rome? - What does the course cluster charge instead?
Objectives
- Read the Snellius system-details table: nodes, cores, memory, SBU per hour, smallest allocation.
- Compute an SBU cost from cores, hours, and the SBU rate of the partition.
- Say what a full CPU allocation does to idle GPUs, and what a one-core request is rounded up to.
The table you are billed from
Snellius charges system billing units (SBU). The accounting portal is https://portal.cua.surf.nl/.
A rome node is the row most jobs in this course start
from: 521 nodes, 128 cores, 256 GB, AMD Rome 7H12 at 2.6 GHz, 128 SBU
per node-hour. The smallest allocation is 1/8 of a node, which is 16
cores.
The same table has more partitions. GPU counts are per node.
gpu_a100 is 4 NVIDIA A100 40 GB. gpu_h100 is 4
NVIDIA H100 (SXM5) 94 GB.
| Partition | Nodes | Cores | Memory | SBU/hour | Smallest |
|---|---|---|---|---|---|
| rome | 521 | 128 | 256 GB | 128 | 1/8 |
| fat_rome | 72 | 128 | 1,024 GB | 192 | 1/8 |
| himem_4tb | 2 | 128 | 4,096 GB | 256 | 1/8 |
| himem_8tb | 2 | 128 | 8,192 GB | 384 | 1/8 |
| genoa | 737 | 192 | 384 GB | 192 | 1/8 |
| fat_genoa | 48 | 192 | 1,536 GB | 288 | 1/8 |
| gpu_a100 | 63 | 72 | 512 GB | 512 | 1/4 |
| gpu_h100 | 88 | 64 | 768 GB | 768 | 1/4 |
| staging | 10 | 16 | 256 GB | 32 | 1/32 |
Login nodes are a separate row of 3 machines. Do not run a heavy task there. Their hardware is not one of the rows above.
The whole machine in the deck’s summary line is 1,258 thin rome/genoa nodes, plus the fat and high-memory nodes, plus 604 GPUs. Home is 720 TB. Scratch and project together are 12.4 PB, on GPFS.
The formula
Documentation talks about cores. The charge uses the partition’s SBU per node-hour and the number of cores on that node:
cores x hours x (SBU per node-hour / cores per node)
16 cores for 2 hours on rome:
16 x 2 x (128 / 128) = 32
16 cores is 1/8 of 128, so it is the smallest rome job.
At 128 SBU per node-hour that slice is 16 SBU per hour. Two hours is 32
SBU.
A full rome node for one hour is 128 cores, which is 128
SBU. Asking for fewer cores than the smallest unit does not make a
cheaper job. The system rounds the request up to that unit, and the job
gets every core in the unit.
The tightest resource wins
Each node has cores, memory, accelerators, and local disk. The allocation follows the most restrictive thing you asked for.
If you request every CPU core and no GPU, the node is fully allocated. The GPUs sit idle, and nobody else can use them, because the CPUs are taken.
If you request one core, the system rounds up to the smallest
permitted unit. On rome that unit is 1/8 of the node. You
then have every core in that unit, not the single core you typed.
Read the tab for the machine in the welcome note. The Snellius tab is the SBU table above. The course-cluster tab has no project budget.
Cost a rome job
A job asks for 16 cores on rome and runs for 2 hours.
How many SBU is that? Why is 16 the interesting number, rather than
8?
32 SBU.
16 x 2 x (128 / 128) = 32
16 cores is 1/8 of a 128-core rome node, and 1/8 is the
smallest allocation on that row. A request for 8 cores is rounded up to
that same unit, so it does not cost half as much.
Cost a genoa slice
genoa is 192 cores and 192 SBU per node-hour. The
smallest allocation is still 1/8 of a node. A job uses 24 cores for 1
hour. How many SBU?
24 SBU.
24 x 1 x (192 / 192) = 24
24 cores is 192/8, so this is one smallest unit for one hour. One
full genoa node-hour is 192 SBU.