...
Partition | Number of nodes | CPU | RAM | Proportional RAM for one GPU | Proportional CPU for one GPU | Accelerator |
---|---|---|---|---|---|---|
plgrid-gpu-a100 | 48 | 128 cores, 2x AMD EPYC 7742 64-Core Processor @ 2.25 GHz | 1024 GB | 128000MB | 16 | 8x NVIDIA A100-SXM4-40GB |
Job submission
...
Accounts and computing grants
Athena uses a new naming scheme of naming Slurm accounts GPU computing grants. GPU computing grants using A100 GPU resources use the for GPU computing accounts, which are supplied by the -A parameter in sbatch command. Currently, accounts are named in the following manner:
Resource | account name |
---|---|
GPU | grantname-gpu-a100 |
suffix. Please mind that sbatch -A grantname
won't work on its own. You need to add the -gpu-a100 suffix! Available computing grants, with respective account names (allocations), can be viewed by using the hpc-grants
command.
Resource allocated on Athena doesn't use normalization, which was used on Prometheus. 1 hour of GPU time equals 1 hour spent using a GPU.on a GPU with a proportional amount of CPUs and memory (consult the table above). The billing system accounts for jobs that use more CPUs or memory than the proportional amount. If the job uses more CPU or memory for each allocated GPU than the proportional amount, it will be billed as it would have used more GPUs. The billed amount can be calculated by dividing the used number of CPUs or memory by the proportional amount per GPU and rounding the result to the closest and larger integer. Jobs on GPU partitions are always billed in GPU hours.
The cost can be expressed as a simple algorithm:
Code Block |
---|
cost_gpu = job_gpus_used * job_duration
cost_cpu = ceil(job_cpus_used/cpus_per_gpu) * job_duration
cost_memory = ceil(job_memory_used/memory_per_gpu) * job_duration
final_cost = max(cost_gpu, cost_cpu, cost_memory) |
Storage
Available storage spaces are described in the following table:
...