Service Unit Policies

Service Unit Definition

The heterogeneous nature of H2P means that some resources are more valuable than others. For example, the SMP nodes have 256 GB of RAM and the MPI_OPA nodes only have 64 GB. Some GPU nodes have K40 cards, while others have Titan X and GTX 1080. The CRC team attempted to create a fair system. First, 1 service unit (SU) is roughly defined as using 1 CPU with the default RAM for 1 hr. To avoid users asking for 1 CPU and all of the RAM on the machine, we actually take the maximum of the charge for the CPU and RAM. The default RAM for your job is calculated as:

Default RAM = Number of Cores * Total RAM / Number of Cores

The RAM charge is scaled such that 1 hr with the default RAM is 1 SU. Some example charges:

  • Use 12 cores with default RAM for 4 hours (SMP, 12 cores is all the cores on machine):
    • 4 hours * maximum(12 SUs [CPU], 12 SUs [RAM]) = 48 SUs

** Insert CPU prefactors table here **

GPUs are charged as card hours, not core hours, e.g.:

  • Use 2 cards for 4 hours:
    • card pre-factor * 4 * 2

** Insert card prefactors table here **

Service Unit Allocations

SUs are allocated on a per account (group) basis. For all groups on H2P the first 10,000 SUs are on us. If that is enough for your research projects then great! If you need more SUs, you will need to submit a proposal. See  guidelines for allocation proposal for more details.

To view your remaining SUs

  • H2P: crc-usage.py
  • Frank: mybalance -h