Page History

Overview

El Gato

Implemented at the start of 2014 and it has been reprovisioned with CentOS 7 and new compilers and libraries. It also uses PBS as the scheduler consistent with Ocelote. El Gato is a large GPU cluster, purchased by an NSF MRI grant by researchers in Astronomy and SISTA. Whilst the Nvidia K20 GPUs are five years old, they are still valuable for single-precision workload. There are 90 nodes with one or two GPU's.

Ocelote

Implemented in the middle of 2016. It is designed to support all workloads on the standard nodes except:

Large memory workloads that do not run within the 192GB RAM of each node can also be handled with the large memory node.
GPU's are available as a buy-in option or windfall and are now available standard on Ocelote.

Tip
Ocelote now has 46 nodes with Nvidia P100's that are available for "standard" and "windfall" queues. See details at GPU Nodes

Puma

Will have a soft opening August 2020. Similarly to Ocelote it will have standard CPU nodes (with 96 cores and 521 GB of memory per node), GPU nodes (with Nvidia V100) and a high-memory node (3 TB). Local scratch storage increased to ~1,6 TB. The new cluster is going to run on CentOS 7 which will improve compatibility with the latest software.

Puma is currently different from ElGato and Ocelote in these ways:

Puma uses SLURM for job scheduling
Puma has many programs installed natively (e.g. Singularity). You won't have to module load Singularity any more!
There's a new alias command: interactive. This automatically places you in a single-cpu interactive session for debugging, testing, and development.
Modules are no long available on the login nodes. To view and test modules, an interactive session is necessary (see above).

Free vs Buy-In

The HPC resources at UA are differentiated from many other universities in that there is central funding for a significant portion of the available resources. Each PI receives a standard monthly allocation of hours at no charge. There is no charge to the allocation for windfall usage and that has proven to be very valuable for researchers with substantial compute requirements.

Research groups can 'Buy-In' (add resources such as processors, memory, storage, etc.) to the base HPC systems as funding becomes available. Buy-In research groups will have highest priority on the resources they add to the system. If the expansion resources are not fully utilized by the Buy-In group they will be made available to all users as windfall.

Details on allocations

Details on buy-in

Test Environment

HPC has a test / trial environment as well as the primary clusters detailed below. This environment is intended to be used for projects that are six months or less in duration and cannot be run on the production systems. Reasons for not being able to be run on the production systems include requiring root access, and hardware or software requirements that cannot be met by one of the production systems. If you have a project in mind that we might be able to support, contact hpc-consult@list.arizona.edu

Feature	Detail
Nodes	16
CPU	Xeon Westmere-EP X5650 Dual 6-core
Memory	128GB
Disk	10TB (5 x 2TB)
Network	GbE and QDR IB

Compute System Details

Name	El Gato	Ocelote	Puma
Model	IBM System X iDataPlex dx360 M4	Lenovo NeXtScale nx360 M5	Penguin Altus XE2242
Year Purchased	2013	2016 (2018 P100 nodes)	2020
Node Count	131	400	192
Total System Memory (TB)	26 TB	82.6 TB	105 TB
Processors	Xeon Ivy Bridge E5-2650 Dual 8-core	Xeon Haswell E5-2695 Dual 14-core Xeon Broadwell E5-2695 Dual 14-core	AMD EPYC 7642 Dual 48-core
Cores / Node	16	28*	94
Total Cores	2160	11528**	19200
Processor Speed (GHz)	2.66	2.3	2.4
Memory / Node (GB)	64 or 256	192 High memory node - 2TB	512 High memory node - 3 TB
Accelerators	137 Nvidia K20x 5 GB video mem 47 nodes with 2 K20x 43 nodes with 1 K20x	46 Nvidia P100 16 GB video mem 15 Nvidia K80 (buy-in only)	24 Nvidia V100 32 GB video mem
/tmp	~840 GB spinning /tmp is part of root filesystem	~840 GB spinning /tmp is part of root filesystem	~1640 GB NVMe /tmp is part of root filesystem
Max Performance (TFLOPS)	46	382
OS	Centos 7.6	CentOS 6.10	CentOS 7
Interconnect	FDR Inifinband	FDR Infiniband for node-node 10 Gb Ethernet node-storage	100 Gb/s Spine/Leaf 2x 25 Gb per compute node Ethernet with RDMA via RoCEv2

* Ocelote includes a large memory node with 2TB of RAM available on 48 cores. More details on the Large Memory Node

** Adjusted for the high memory node

Example Resource Requests

Note Type	ncpus	pcmem	Max mem	Sample Request Statement
ElGato
Standard	16	4gb	62gb	#PBS -l select=1:ncpus=16:mem=62gb:pcmem=4gb
GPU¹	16	16gb	250gb	#PBS -l select=1:ncpus=16:mem=250gb:ngpus=1:pcmem=16gb
Ocelote
Standard	28	6gb	168gb	#PBS select=1:ncpus=28:mem=168gb:pcmem=6gb
GPU^2,3	28	8gb	224gb	#PBS select=1:ncpus=28:mem=224gb:np100s=1:os7=True
High Memory	48	42gb	2016gb	#PBS -l select=1:ncpus=48:mem=2016gb:pcmem=42gb
Puma
Standard	94	5gb	470gb 470gb	#SBATCH --nodes=1 #SBATCH --ntasks=94 #SBATCH --mem=470gb
GPU⁴	94	5gb	470gb 470gb	#SBATCH --nodes=1 #SBATCH --ntasks=94 #SBATCH --mem=470gb #SBATCH --gres=gpu:1
High Memory	94	32gb	3000gb	#SBATCH --nodes=1 #SBATCH --ntasks=94 #SBATCH --mem=3008gb
¹Two GPUs may be requested on ElGato with ngpus=2 ²There is a single node available on Ocelote with two GPUs. To request it, use np100s=2 ³Set os7=False for a CentOS 6 GPU node ⁴Up to four GPUs may be requested on Puma with --gres=gpu=1, 2, 3, or 4

Page tree

Versions Compared

Old Version 77

New Version 78

Key