Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.nyc-ai.app/llms.txt

Use this file to discover all available pages before exploring further.

CSI HPCC operates several computer systems grouped into three service tiers: free tier (FT), advanced tier (AT), and condo tier (CT). Its main hybrid cluster is Arrow. Jobs target different resources depending on your project, QOS, and partition. When in doubt, Arrow is the current default for new users.

​
Architecture at a glance

Diagram showing how Chizen and Karle gateways connect users to Arrow and other clusters, with shared home and scratch file systems
Only Chizen and Karle are reachable from outside the CSI network. All other systems require logging into a gateway first. Bulk data transfer runs through cea.

​
Arrow

Arrow is the current flagship hybrid CPU + GPU cluster and is where most new workloads are placed.
  • Scheduler: SLURM.
  • Module system: LMOD (Lua-based, supports hierarchies).
  • GPU config: up to 8 GPUs per node on Arrow. Request them with --gres=gpu:N.
  • Workspace: jobs must start from /scratch/<username>. Launching from /global/u/<username> is not supported.
Exact per-node specifics (CPU model and core count, RAM, GPU model and memory, interconnect) should be filled in here against authoritative HPCC sources before publishing. The unaudited research draft for this site cited figures like “62 nodes, 64 cores/node, 8× NVIDIA A40 80GB, 2 TB RAM”. Treat those as placeholders until confirmed on the HPCC Wiki or by a sysadmin.

​
Other systems

CSI HPCC has historically operated additional named subsystems and servers. Depending on what is currently in service for your project you may use one or more of:
  • Penzias: an older hybrid cluster that the HPCC Wiki describes as partially retired.
  • Appel: an operational NUMA system for massive parallel, sequential, and OpenMP work; the HPCC Wiki lists it without GPUs.
  • Salk: a legacy Cray XE6m system that the HPCC Wiki describes as fully retired.
Service status changes over time as systems are retired or repurposed. Before directing users to a specific subsystem, confirm it’s still available for your project. The canonical list lives on the HPCC Wiki.

​
GPUs available

HPCC partitions host several generations of NVIDIA data-center accelerators. Request the type you need on the account form, and constrain on SKU at submit time with --constraint='gpu_sku:<name>' (see GPU with a specific type).

NVIDIA V100

NVIDIA V100 GPU
Volta generation. Available in 16 GB and 32 GB HBM2 variants.

NVIDIA A30

NVIDIA A30 GPU
Ampere data-center GPU. 24 GB HBM2.

NVIDIA A40

NVIDIA A40 GPU
Ampere visualization/compute GPU. 24 GB GDDR6.

NVIDIA L40

NVIDIA L40 GPU
Ada Lovelace generation. 48 GB GDDR6.

NVIDIA A100

NVIDIA A100 GPU
Ampere flagship. Available in 40 GB and 80 GB HBM2e variants.

​
Service tiers

Jobs across HPCC are organized by service tier and partition:
TierPurpose
Free tier (FT)General-purpose allocation available to all approved users.
Advanced tier (AT)Higher-priority allocation for approved projects.
Condo tier (CT)Dedicated hardware purchased by a PI or group.
The tier that applies to your project determines which QOS and partition you pass in your SLURM scripts. Ask your PI or the HPC Helpline if you’re unsure.

​
File systems

  • /global/u/<username>: your home directory, hosted on DSMS, backed up to tape. Small quota, but durable.
  • /scratch/<username>: your scratch workspace, large but ephemeral. Not backed up. Files can be purged when the filesystem exceeds ~70% full or after about two weeks.
  • Project directories may be provisioned for group work under request.
See Storage & quotas for sizes, purge rules, and how to move data in and out.

​
Next steps

Software & modules

Use LMOD to get compilers, MPI, and applications into your environment.

Job submission

SLURM templates that pair with each system and tier.