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 help@uw.edu with Hyak as the first word in the subject. Please also let us know you're using mox not ikt.

Connecting

SSH = mox.hyak.uw.edu

BBCP = mox1.hyak.uw.edu or mox2.hyak.uw.edu

File Transfers

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/

Filesystems

Scrubbed

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.

Quotas

Home Directory

mmlsquota gscratch:home --block-size G

Allocation Directory

mmlsquota -j <my short group> gscratch --block-size G

cat /gscratch/<my short group>/usage_report.txt

Slurm Primer

Show Queue

All Jobs

squeue

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

Submit

Own Allocation

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

Interactive Session

Build Allocation - usage limited by core count and time

srun -p build --mem=100G --time=2:00:00 --pty bash -l

Own Allocation

srun -p <my short group name> -A <my short group name> --mem=100G --time=2:00:00 --pty bash -l

Show Allocation Information

Show all allocation information

hyakalloc

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>

Other Commands

Cancel Jobs

scancel <jobid>

scancel -u <username>

Show Job Resource Information

No Slurm Job Step (generally non-MPI jobs)

sstat <jobid>.batch

Slurm Job Step (generally MPI jobs)

sstat <jobid>

Sample Job Script

Change items that are bold and red. Check items that are bold and blue and change if necessary.

#!/bin/bash
## Job Name
#SBATCH --job-name=test-job ## Allocation Definition #SBATCH --account=MYSHORTGROUP #SBATCH --partition=MYSHORTGROUP ## Resources
## Nodes
#SBATCH --nodes=2
## Tasks per node (Slurm assumes you want to run 28 tasks, remove 2x # and adjust parameter if needed) ###SBATCH --ntasks-per-node=28 ## Walltime (two hours)
#SBATCH --time=2:00:00 # E-mail Notification, see man sbatch for options

##turn on e-mail notification

#SBATCH --mail-type=ALL

#SBATCH --mail-user=your_email_address

## Memory per node
#SBATCH --mem=100G
## Specify the working directory for this job
#SBATCH --chdir=/gscratch/MYGROUP/MYUSER/MYRUN

module load icc_<version>-impi_<VERSION>
mpirun /gscratch/MYGROUP/MYMODEL/MYMODEL-BIN

Usage Reporting

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