Grace: A Dell x86 HPC Cluster
|Operating System:||Linux (CentOS 7)|
|Total Compute Cores/Nodes:||44,656 cores|
|Compute Nodes:||800 48-core compute nodes, each with 384GB RAM |
100 48-core GPU nodes, each with two A100 40GB GPU accelerator and 384GB RAM
8 38-core GPU nodes, each with two RTX 6000 24GB GPU accellator and 384 GB RAM
8 48-cores GPU nodes, each with 4 T4 16GB GPU accelerator
8 80-core large memory nodes, each with 3TB RAM
|Interconnect:||Mellanox HDR Infiniband|
|Peak Performance:||2,638.5 TFLOPS|
|Global Disk:||5PB (usable) via DDN appliance for general use |
?PB (raw) via Lenovo's DSS purchased by and dedicated for ??
|File System:||Lustre and GPFS|
|Batch Facility:||Slurm by SchedMD|
|Location:||West Campus Data Center|
|Production Date:||Spring 2021|
Grace is an Intel x86-64 Linux cluster with 950 compute nodes (44,656 total cores) and 5 login nodes. There are 800 compute nodes with 384 GB of memory, and 117 GPU nodes with 384 GB of memory. Among the 117 GPU nodes, there are 100 GPU nodes two A100 40 GB GPU cards, 9 GPU nodes with two RTX 6000 24GB GPU cards, 8 GPU nodes with four T4 16GB GPU cards. These 800 compute nodes and 117 GPU nodes are a dual socket server with two Intel 6248R 3.0GHz 24-core processors. There are 8 compute nodes with 3 TB of memory and four Intel 6248 2.5 GHz 20-core processors.
The interconnecting fabric is a two-level fat-tree based on EDR Infiniband. High performance mass storage of 5 petabyte (usable) capacity is made available to all nodes by DDN.
Get details on using this system, see the User Guide for Grace.
A description of the four types of compute nodes is below:
|Large memory 3TB|
|Processor Type||Intel Xeon 6248R 3.0GHz 24-core||Intel 6248 2.5 GHz 20-core|
|Memory/Node||384 GB DDR4, 3200 MHz||3 TB DDR4, 3200 MHz|
|Accelerator(s)||N/A||2 NVIDIA A100 40GB Accelerator||2 NVIDIA RTX6000 24GB Accelerator||4 NVIDIA T4 16GB Accelerator||N/A|
|Interconnect||Mellanox HDR Infiniband|
|Local Disk Space||1.6TB NVMe|
Usable Memory for Batch Jobs
While nodes on Grace have either 384GB or 3TB of RAM, some of this memory is used to maintain the software and operating system of the node. In most cases, excessive memory requests will be automatically rejected by SLURM.
The table below contains information regarding the approximate limits of Grace memory hardware and our suggestions on its use.
The grace.hprc.tamu.edu hostname can be used to access the Grace cluster. This translates into one of the five login nodes, grace[1-5].hprc.tamu.edu. To access a specific login node use its corresponding host name (e.g., grace2.tamu.edu). All login nodes have 10 GbE connections to the TAMU campus network and direct access to all global parallel (Lustre-based) file systems. The table below provides more details about the hardware configuration of the login nodes.
5PB (usable) with Lustre provided by DDN
"Grade" is named for Grace Hopper.