Difference between revisions of "Available Node Features"

From UFRC
Jump to navigation Jump to search
m (Text replacement - "Category:SLURM" to "Category:Scheduler")
 
(8 intermediate revisions by 3 users not shown)
Line 1: Line 1:
[[Category:Scheduler]]
+
[[Category:Scheduler]][[Category:Infrastructure]]
 +
HiPerGator users may finely control selection of compute hardware for a SLURM job like specific processor families, processor models by using the <code>--constraint</code> directive to specify HiPerGator server ''features''.
  
HiPerGator users may finely control selection of compute hardware for a SLURM job like specific processor families, processor models by using the <code>--constraint</code> directive to specify node ''features''.
+
;Example:
  
;Example:
+
Use one of the following commands to specify between rome and milan microarchitectures
  #SBATCH --constraint=westmere
+
  #SBATCH --constraint=rome
or
+
  #SBATCH --constraint=milan
  #SBATCH --constraint=haswell
 
  
 
Basic boolean logic can be used to request combinations of features. For example, to request nodes that have Intel processors '''AND''' InfiniBand interconnect use
 
Basic boolean logic can be used to request combinations of features. For example, to request nodes that have Intel processors '''AND''' InfiniBand interconnect use
Line 12: Line 12:
 
  #SBATCH --constraint='intel&infiniband'
 
  #SBATCH --constraint='intel&infiniband'
  
To request processors from either Intel Haswell '''OR''' skylake CPU family use
+
To request processors from either AMD Rome '''OR''' Milan CPU family use
  
  #SBATCH --constraint='haswell|skylake'
+
  #SBATCH --constraint='rome|milan'
  
 
==All Node Features==
 
==All Node Features==
 
You can run <code>nodeInfo</code> command from the ufrc environment module to list all available SLURM features. In addition, the table below shows automatically updated nodeInfo output as well as the corresponding CPU models.
 
You can run <code>nodeInfo</code> command from the ufrc environment module to list all available SLURM features. In addition, the table below shows automatically updated nodeInfo output as well as the corresponding CPU models.
 
 
{{#get_web_data:url=https://data.rc.ufl.edu/pub/ufrc/data/node_data.csv
 
{{#get_web_data:url=https://data.rc.ufl.edu/pub/ufrc/data/node_data.csv
 
|format=CSV with header
 
|format=CSV with header
Line 24: Line 23:
 
|cache seconds=7200
 
|cache seconds=7200
 
}}
 
}}
{| class="wikitable sortable" border="1"  cellspacing="0" cellpadding="2" align="center" style="border-collapse: collapse; margin: 1em 1em 1em 0; border-top: none; border-right:none; "
+
{| class="wikitable sortable" border="1"  sort=Partition cellspacing="0" cellpadding="2" align="center" style="border-collapse: collapse; margin: 1em 1em 1em 0; border-top: none; border-right:none; "
 
! Partition
 
! Partition
 
! Cores per node
 
! Cores per node
Line 46: Line 45:
 
|}
 
|}
  
'''Note''': the following features correspond to the following cpu models in order from the oldest to the newest - interlagos, magny, sandy-bridge, dhabi, haswell, broadwell, skylake. The 'dhabi' and 'haswell' models are from HPG1 and HPG2 deployments.
+
'''Note''': the bigmem partitions are maintained for calculations requiring large amounts of memory. To submit jobs to this partition you will need to add the following directive to your job submission script.
 +
  #SBATCH --partition=bigmem
 +
 
 +
Since our regular nodes have 1TB of available memory we do not recommend using bigmem nodes for jobs with memory requests lower than that.
 +
 
 +
 
 +
'''Note''': See [[GPU_Access]] for more details on GPUs, such as available GPU memory. The following CPU models are in order from the oldest (HPG2) to the newest (HPG3) - haswell, rome, milan.

Latest revision as of 19:45, 22 April 2024

HiPerGator users may finely control selection of compute hardware for a SLURM job like specific processor families, processor models by using the --constraint directive to specify HiPerGator server features.

Example

Use one of the following commands to specify between rome and milan microarchitectures

#SBATCH --constraint=rome
#SBATCH --constraint=milan

Basic boolean logic can be used to request combinations of features. For example, to request nodes that have Intel processors AND InfiniBand interconnect use

#SBATCH --constraint='intel&infiniband'

To request processors from either AMD Rome OR Milan CPU family use

#SBATCH --constraint='rome|milan'

All Node Features

You can run nodeInfo command from the ufrc environment module to list all available SLURM features. In addition, the table below shows automatically updated nodeInfo output as well as the corresponding CPU models.

Partition Cores per node Sockets Socket Cores Threads/Core Memory,GB Features CPU Model

hpg-dev 64 8 8 1 500 hpg3;amd;milan;infiniband;el8 AMD EPYC 75F3 32-Core Processor
gui 32 2 16 1 124 gui;i21;intel;haswell;el8 Intel(R) Xeon(R) CPU E5-2698 v3 @ 2.30GHz
hwgui 32 2 16 1 186 hpg2;intel;skylake;infiniband;gpu;rtx6000;el8 Intel(R) Xeon(R) Gold 6242 CPU @ 2.80GHz
bigmem 128 8 16 1 4023 bigmem;amd;rome;infiniband;el8 AMD EPYC 7702 64-Core Processor
bigmem 192 4 24 2 1509 bigmem;intel;skylake;infiniband;el8 Intel(R) Xeon(R) Platinum 8168 CPU @ 2.70GHz
hpg-milan 64 8 8 1 500 hpg3;amd;milan;infiniband;el8 AMD EPYC 75F3 32-Core Processor
hpg-default 128 8 16 1 1003 hpg3;amd;rome;infiniband;el8 AMD EPYC 7702 64-Core Processor
hpg2-compute 32 2 16 1 124 hpg2;intel;haswell;infiniband;el8 Intel(R) Xeon(R) CPU E5-2698 v3 @ 2.30GHz
hpg2-compute 28 2 14 1 125 hpg2;intel;haswell;infiniband;el8 Intel(R) Xeon(R) CPU E5-2683 v3 @ 2.00GHz
gpu 32 2 16 1 186 hpg2;intel;skylake;infiniband;gpu;2080ti;el8 Intel(R) Xeon(R) Gold 6142 CPU @ 2.60GHz
gpu 128 8 16 1 2010 ai;su3;amd;rome;infiniband;gpu;a100;el8 AMD EPYC 7742 64-Core Processor
hpg-ai 128 8 16 1 2010 ai;su3;amd;rome;infiniband;gpu;a100;el8 AMD EPYC 7742 64-Core Processor

Note: the bigmem partitions are maintained for calculations requiring large amounts of memory. To submit jobs to this partition you will need to add the following directive to your job submission script.

 #SBATCH --partition=bigmem

Since our regular nodes have 1TB of available memory we do not recommend using bigmem nodes for jobs with memory requests lower than that.


Note: See GPU_Access for more details on GPUs, such as available GPU memory. The following CPU models are in order from the oldest (HPG2) to the newest (HPG3) - haswell, rome, milan.