High Level Differences from ikt
- Mox is an entirely separate cluster. They share nothing with one another.
- You only get what you ask for, regardless of the resources available on the node. If you ask for 1 CPU, you'll only get one. If you ask for 1GB of RAM, you'll only get 1GB.
- An allocation won't get the same set of nodes all the time, just access to the particular number of nodes to which they're entitled.
- No occasional preemption in ckpt (formerly bf queue) for the moment.
- Preempted jobs get 10s to do something smart before being killed and requeued.
- Please report any problems to firstname.lastname@example.org with Hyak as the first word in the subject. Please also let us know you're using mox not ikt.
SSH = mox.hyak.uw.edu
BBCP = mox1.hyak.uw.edu or mox2.hyak.uw.edu
Internal to Hyak systems
You can copy files at high speed without a password between the Hyak systems using commands like the ones below.
From ikt to mox
ikt1$ hyakbbcp myfile mox1.hyak.uw.edu:/gscratch/MYGROUP/
ikt1$ hyakbbcp -r mydirectory mox1.hyak.uw.edu:/gscratch/MYGROUP/
From mox to ikt
mox1$ hyakbbcp myfile ikt1.hyak.uw.edu:/gscratch/MYGROUP/
mox1$ hyakbbcp -r mydirectory ikt1.hyak.uw.edu:/gscratch/MYGROUP/
There's a scrubbed temporary filesystem available at /gscratch/scrubbed. Files can be removed at any time, but they will be removed on a periodic basis based on creation date (files created 30 days ago or more). The maximum available space is 200TB. That space is shared among all users. Scrubbed storage is for temporary use only. Per month per TB persistent storage is available for purchase.
mmlsquota gscratch:home --block-size G
mmlsquota -j <my short group> gscratch --block-size G
cat /gscratch/<my short group>usage_report.txt
Jobs in Allocation
squeue -p <my short group>
All Jobs in ckpt (was bf)
squeue -p ckpt
Jobs in ckpt from Allocation
squeue -A <my short group>-ckpt
Like ikt you're limited to the number of nodes your group purchased for your allocation.
sbatch -p <my short group> -A <my short group> test-job.sh
Checkpoint Allocation (formerly bf queue)
If you want access to all the nodes in the cluster, you need to use the ckpt partition.
sbatch -p ckpt -A <my short group>-ckpt test-job.sh
Build Allocation - usage limited by core count and time
srun -p build --pty bash -l
srun -p <my short group name> -A
<my short group name> --pty bash -l
Show Allocation Information
Show all allocation information
Show allocation and usage information for specified allocation
hyakalloc <my short group name>
Show Job Info
scontrol show job <jobid>
Show Node Info
scontrol show node <node>
scancel -u <username>
Show Job Resource Information
No Slurm Job Step (generally non-MPI jobs)
Slurm Job Step (generally MPI jobs)
Sample Job Script
Change items that are bold and red. Check items that are bold and blue and change if necessary.
## Job Name
## Allocation Definition
## Tasks per node (Slurm assumes you want to run 28 tasks per node unless explicitly told otherwise)
## Walltime (ten minutes)
## Memory per node
## Specify the working directory for this job
module load icc_<version>-impi_<VERSION>
The sreport command has many options for reporting. For more details refer to the man page. Below you will find two common examples. The usage is reported in TRES minutes which is the equivalent of core minutes. Using all the cores in one twenty-eight core node for one hour is 1680 TRES minutes.
Show formatted per-user utilization data for the month of June 2017 for specified accounts
sreport cluster AccountUtilizationByUser Accounts=<my short group name>,<my short group name>-ckpt Start=2017-06-01 End=2017-07-01
Show easily parsable per-user utilization data for the month of June 2017 for all accounts
sreport --parsable --noheader cluster AccountUtilizationByUser Start=2017-06-01 End=2017-07-01
See articles on mox at below link (scroll down to links with prefix mox):
You can find additional documentation that applies to both mox and ikt on the main Hyak User Wiki.Show utilization data for the month of June 2017 for the hpc and hpc-ckpt accounts