[slurm-users] memory per node default

[Hoot Thompson]

Thank you for the support. I will be back with any additional questions. 
BTW, if it changes or adds to your thoughts, I'm working in AWS on a 
parallelcluster.


Hoot

On 1/21/22 4:12 AM, Ole Holm Nielsen wrote:
> On 1/21/22 10:05, Diego Zuccato wrote:
>> Il 21/01/2022 07:51, Ole Holm Nielsen ha scritto:
>>
>>> There's a nice command to run on any given node which tells you 
>>> slurmd's view of the node:
>>> $ slurmd -C
>>> NodeName=i004 CPUs=16 Boards=1 SocketsPerBoard=2 CoresPerSocket=8 
>>> ThreadsPerCore=1 RealMemory=128691
>>> UpTime=36-17:32:44
>>> Here's an example that I use:
>>> NodeName=i[004-030] Sockets=2 CoresPerSocket=8 ThreadsPerCore=1 
>>> RealMemory=128000
>> @Hoot: note that the line in slurm.conf uses a lower value for 
>> RealMemory. This way you'll avoid having nodes going offline after an 
>> upgrade because the new kernel leaves a couple free MBs less than 
>> older one (been there, done that... :( ).
>
> @Diego: Thanks for reminding me of this!  I've added this information 
> to my Wiki page
> https://wiki.fysik.dtu.dk/niflheim/Slurm_configuration#compute-node-configuration 
>
>
> /Ole
>