Policies and limits

These rules protect a shared research service. Users are responsible for their accounts, jobs and data; principal investigators are responsible for ensuring that group members are trained and that use is authorized for university research or teaching.

Administrator approval required

Numerical values on this page come from MHTML snapshots saved on 4 August 2026. They are presented as a migration baseline, not as proof of the live configuration. The HPC Committee and administrators may change limits. Complete the maintainer checklist before treating this edition as policy.

Acceptable use

• Use only your own account and approved authentication methods. Do not share credentials.
• Submit production computation through Slurm. Login nodes are for access, file management, editing, compilation and short tests.
• Request resources that the application will use. Release interactive allocations promptly.
• Do not attempt to bypass quotas, scheduler controls, access restrictions or licensing.
• Protect confidential, licensed and personally identifiable data according to NU policy and applicable law.
• Report suspected security incidents, compromised credentials and destructive software behavior immediately.

Intentional misuse or repeated noncompliance may lead to account suspension and escalation under university policy. Link the final site to the authoritative NU Acceptable Use Policy; the MHTML archive did not include that policy's text.

Storage

Snapshot quota baseline

System	Home path	Default home quota	Group path	Default group quota
Irgetas	/shared/home/<username>	400 GB	/datahub/<group>	1 TB
Shabyt	/shared/home/<username>	100 GB	/zdisk/<group>	1 TB
Muon	/shared/home/<username>	300 GB	/zdisk/<group>	1 TB

Quota increases require a PI-supported Helpdesk request and are reviewed individually. Capacity units, grace periods, inode limits and group ownership should be stated in the approved operational policy.

Check usage

The snapshot mentions uq, userquota and the following BeeGFS queries:

beegfs-ctl --getquota --uid "$(id -u)"
beegfs-ctl --getquota --uid "$(id -u <username>)"

These commands are filesystem- and permission-dependent. The archive's description of Muon storage does not clearly match the BeeGFS command, so maintainers must document one tested command per filesystem instead of promising that every command works everywhere.

Performance is not capacity

Use home/shared SSD storage for active work and group HDD storage for larger, less latency-sensitive data. The old pages repeatedly say “1 Mbit/s Ethernet”; the hardware inventory names 1 Gb/s Ethernet and InfiniBand links. This edition removes the unsupported throughput promise. Measure representative I/O inside an allocation and confirm the storage topology with administrators.

Backup and recovery

The snapshot says /shared/home is backed up several times per week and that /zdisk and /datahub are not covered. Until an approved service definition states otherwise:

• treat group storage as not backed up;
• keep an independent copy of irreplaceable data;
• do not treat RAID as backup;
• expect storage to be unavailable during a hardware incident;
• verify recovery-point, retention and restore-request procedures before relying on home backup.

Partitions and wall times

System	Partition	Nodes	CPU cores / threads per node	RAM per node	GPUs per node	Snapshot maximum wall time
Irgetas	ZEN4	10	192 / 384	384 GB	—	7 days
Irgetas	H100	6	192 / 384	768 GB	4 × H100	4 days
Shabyt	CPU	20	64 / 128	256 GB	—	14 days
Shabyt	NVIDIA	4	64 / 128	256 GB	2 × V100	2 days
Muon	HPE	10	14 / 28	64 GB	—	14 days

Live checks:

sinfo -o "%P %a %l %D %c %m %G"
scontrol show partition <partition>

The scheduler may reserve some memory for the operating system. Do not infer an exact --mem ceiling solely from installed RAM.

Quality of Service

The snapshot lists the following categories:

Cluster	QoS values	Intended group in snapshot
Irgetas	hpcnc, nu, issai, issai-ext, stud	HPCNC; general NU; ISSAI; ISSAI external collaborators; course students
Shabyt	hpcnc, nu, stud	HPCNC; general NU; course students
Muon	hpcnc, nu	HPCNC; general NU

The old page says Irgetas has four categories but lists five; this edition corrects the count by avoiding the contradictory sentence. Users normally should not select a QoS unless administrators instruct them to do so.

Where permitted, inspect associations with:

sacctmgr show assoc user="$USER" format=Cluster,Account,User,Partition,QOS

Snapshot concurrency limits

The values below reproduce the 4 August 2026 snapshot in a more compact form. “CPU” means the site's configured Slurm CPU unit; confirm whether it maps to a core or hardware thread.

System	QoS	Partition	Running jobs	CPU cores	Threads	GPUs	Relative priority
Irgetas	hpcnc	ZEN4	12	576	1,152	—	5
Irgetas	nu	ZEN4	12	576	1,152	—	5
Irgetas	issai	ZEN4	12	576	1,152	—	5
Irgetas	issai-ext	ZEN4	4	192	384	—	1
Irgetas	stud	ZEN4	4	192	384	—	5
Irgetas	hpcnc	H100	12	576	1,152	12	5
Irgetas	nu	H100	12	576	1,152	12	5
Irgetas	issai	H100	24	1,152	2,304	24	10
Irgetas	issai-ext	H100	12	576	1,152	12	5
Irgetas	stud	H100	4	192	384	4	5
Shabyt	hpcnc	CPU, NVIDIA	40	1,280	2,560	8	10
Shabyt	nu	CPU, NVIDIA	12	256	512	8	5
Shabyt	stud	CPU, NVIDIA	4	128	256	4	5
Muon	hpcnc	HPE	40	140	280	—	10
Muon	nu	HPE	40	140	280	—	10

Treat these as caps, not entitlements or performance targets. Pending work can be constrained by fair share, associations, reservations, licenses, maintenance and the requested combination of resources.

Efficient scheduling

• Ask for realistic wall time and memory; shorter or smaller jobs may fit scheduling gaps.
• Use CPU partitions for CPU-only workloads.
• Use --exclusive only when the job genuinely needs an entire node.
• Use job arrays with a concurrency cap for independent sweeps.
• Check sacct/seff after representative jobs and tune future requests.
• Checkpoint runs that approach a partition's time limit.

Publications

If NU RC resources were essential to published work, include an acknowledgment containing the exact service name Nazarbayev University Research Computing. Examples:

The authors acknowledge the use of computational resources provided by Nazarbayev University Research Computing.

A.B. and C.D. acknowledge the use of the Irgetas HPC cluster at Nazarbayev University Research Computing.

Record publications through the university process if required by the final service policy.

Requesting help

For policy, quota or access changes, use the approved Helpdesk service. For a failed job, send the cluster name, job ID, batch script, module list, relevant sacct output and the smallest error excerpt to hpcadmin@nu.edu.kz. Do not include passwords, tokens, private keys or unnecessary research data.