Step 3: Ap­p­ly for Re­sources

• Please don't forget to consider memory usage in your resource request/justification. In the accounting of used compute time, the memory reserved (i.e., the memory allocation specified in the job request) counts towards your resource usage of CPU-core-hours. For example, if you proportionally allocate more of the node's memory than CPU cores, then the higher of the two determines the effective number of CPU-core-hours used by the job. This is done so that the memory allocation in jobs is accounted for in a fair way.

• Example: Allocating 4 CPU cores and 120 GB of main memory on an Otus normal node (730 GB usable main memory) – e.g., `sbatch --mem 120000 -N 1 --ntasks-per-node 4 --cpus-per-task 1`– would “cost” you 31.56 CPU-core-hours for a one-hour job (1 hour times 192 cores per node times 120 GB / 730 GB) instead of 4 CPU-core-hours.

   

• Please consider using the specialized resource justification table, which is designed in terms of the common units of ns/day.

• Also, workloads with unfavorable parallel scaling (e.g., those not scaling beyond a few CPU cores or single GPUs) can be granted computation time if the HPC systems offered by NHR are required to achieve progress for the given scientific problem that would not be possible with other HPC systems.
• The same holds for embarrassingly parallel workloads (high-throughput computing, e.g., parameter studies) where the total computational need requires a large HPC system offered by NHR. In these cases, no scalability benchmarks are needed.
   

FPGA projects are typically of two different kinds

1. FPGA experiments and jobs with active FPGA offloading (where most of the time is spent on FPGA nodes equipped with FPGA cards).
2. Development of FPGA designs, including hardware build steps (so-called synthesis). Here, most of the time is spent on regular CPU nodes, and from time to time, the design is tested with real FPGAs on FPGA nodes.

To estimate the resources for FPGA projects, please check the examples provided here.