Shaheen III#
Shaheen III has a total of 4608 compute nodes, 4 login nodes, 4 datamover nodes and 15 pre-post processing nodes. SLURM scheduler is used to schedule different kinds of workloads as jobs submitted by users. Additionally, Shaheen III’s large storage is also shared between compute nodes and is capable of serving multiple users simultaneously. Policies help streamline the user experience by enforcing quotas and limits on a user or project level. Users must understand these policies before planning their simulation campaigns.
SLURM Policy#
SLURM on Shaheen III has multiple partitions. These partitions exist so that jobs can be routed to different types of computational resources, according to the required of the scheduled workloads.
The table below show the job limits enforced on different partitions.
Note
The term CPUs refers to hyperthreads running on a physical core, i.e. 2 CPUs per physical core
Note that these are enforced on per-user basis
The usage is accounted on the basis of physical core, i.e. for each 2 CPUs allocated for an hour, 1 core hour is billed.
On workq, which is an exclusive allocation of node, full 192 core hours are charged, irrespective of the number of CPUs used.
For access to 72hour partition, the users must send an email request with a justification.
Partition |
Tenancy |
Max CPUs/job |
Max Nodes/job |
Max Walltime/job (hrs) |
Max jobs submitted |
Max jobs running |
---|---|---|---|---|---|---|
workq |
exclusive |
786,432 |
2048 |
24 |
1000 |
200 |
shared |
shared |
64 (4/node) |
16 |
24 |
### |
16 |
debug |
exclusive |
1536 |
4 |
0.5(default)/4(max) |
1 |
1 |
dtn |
exclusive |
1024 |
4 |
1(default)/24(max) |
### |
### |
ppn |
exclusive |
3840 |
15 |
1(default)/24(max) |
### |
### |
72hours |
exclusive |
49152 |
128 |
1(default)/72(max) |
32 |
16 |
Filesystem Policy#
All filesystems on Shaheen III are shared among users. Policies help enable consistent user experiences. It is advised that users read technical documentation provided on this website, attend instructor-led and self-paced trainings by KSL staff, and email in case of any question or ambiguity.
scratch
filesystem#
scratch
is a high performance filesystem on Shaheen III meant for temporary use during the lifetime of the job. For persistent storage, users must move their unused data to project
filesystem.
scratch
has three tiers; please see documentation on Filesystems for more details. There are two quotas enforced on scratch
, capacity and inodes. The table below lists quotas on three tiers:
Tier |
capacity |
inodes |
scope |
---|---|---|---|
Capacity |
10TB |
1 million |
per user |
Bandwidth |
1TB |
1 million |
per user |
IOPS |
50GB |
1 million |
per user |
Note
The inode quota, which is synonymous to number of files per user, is a global policy that governs the files in all tiers of
scratch
. A user can create and maintain no more than 1 million files onscratch
cumulatively.All files expect those in IOPS tier are subject to 60 day purge policy.
Users can check the usage of their quotas on scratch using the following commands:
For overall quota on scratch:
lfs quota -uh <username> /scratch
For quota on capacity tier on scratch:
lfs quota -uh <username> --pool capacity /scratch
For quota on bandwidth tier on scratch:
lfs quota -uh <username> --pool bandwidth /scratch
For quota on IOPS tier on scratch:
lfs quota -uh <username> --pool iops /scratch
project
filesystem#
project
filesystem is a persistent storage for users who are members of a project owned by their respective Principal Investigators (PI). A user can be member of multiple projects on Shaheen III. The ID assigned to each project is also used with SLURM to charging to the account when a job is submitted.
Below are some important policies users must know of:
A PI has a default allocation of 80TB on
project
filesystem. This is shared among the members of the project. A list of users and their usage can be queried using the following command:
sb_user <project-id>
project
is mounted as read-only on compute nodes of Shaheen III, except on data mover nodes indtn
partition of SLURM. Please create jobs for moving data betweenscratch
andproject
to maximize the throughput. There are a number of utilities listed in Data Management documentation. In short, try to use the data mover nodes for movement of data wherever possible.
home
filesystem#
home
filesystem on Shaheen III is available only on login nodes. This is the same home
filesystem you see on Ibex system if you have access to it. On compute nodes of Shaheen III, home
filesystem is redirected to /scratch/<username>/
directory.
home
filesystem has per user quota of 200GB enforced. To query the quota, the following command can be used:
quota -s
home
usually is default filesystem for temporary cache files e.g. when using conda package manager, or pulling images using singularity container platform. It is advisable to clear caches every now and then to free space.