Re: [slurm-users] Changing node weights in partitions
Dear Ole, Thanks for your fast reply. I really appreciate that. I had a look at your website and googled about “weight masks” but still have some questions. From your example I see that the mask definition is commented out. How to define what the mask means? If helps, I’ll put an easy example. Node1 have more RAM and clock freq than node2. Partition A should start filling node1, while partition B should start filling node2. Can I accomplish this behavior through weighting the nodes? With your example I’m afraid to say it’s not still clear to me how. Thanks a lot for your help. José > On 22. Mar 2019, at 16:29, Ole Holm Nielsen > wrote: > >> On 3/22/19 4:15 PM, José A. wrote: >> Dear all, >> I would like to create two partitions, A and B, in which node1 had a certain >> weight in partition A and a different one in partition B. Does anyone know >> how to implement it? > > Some pointers to documentation of this and a practical example is in my Wiki > page: > https://wiki.fysik.dtu.dk/niflheim/Slurm_configuration#node-weight > > /Ole >
Re: [slurm-users] Changing node weights in partitions
Hello Chris, You got my point. I want a way in which a partition influences the priority with a node takes new jobs. Any tip will be really appreciated. Thanks a lot. Cheers, José > On 23. Mar 2019, at 03:38, Chris Samuel wrote: > >> On 22/3/19 12:51 pm, Ole Holm Nielsen wrote: >> >> The web page explains how the weight mask is defined: Each digit in the mask >> defines a node property. Please read the example given. > > I don't think that's what José is asking for, he wants the weights for a node > to be different when being considered in one partition to when it's being > considered in a different partition. > > I don't think you can do that though I'm afraid, José, I think the weight is > only attached to the node and the partition doesn't influence it. > > All the best, > Chris > -- > Chris Samuel : http://www.csamuel.org/ : Berkeley, CA, USA >
Re: [slurm-users] Changing node weights in partitions
Dear Ole, Thanks for the support. I think that could help me in the following way: 1. Setting different partitions with the node groups I want to prioritize. 2. Allowing users to submit to several partitions at the same time 3. Through accounting, creating accounts with different priorities from one partition to another. That will allow that each job type, associated to an account, starts differently in different partitions. 4. Once a job start in one partition, the other submitted jobs are killed and get out of SLURM. It’s a bit more work but gets the effect I am looking for: that different nodes prioritize different types of jobs. Is that, specially step 4, possible! Thanks for the help. José > On 24. Mar 2019, at 21:52, Ole Holm Nielsen > wrote: > > Hi José, > >> On 23-03-2019 19:59, Jose A wrote: >> You got my point. I want a way in which a partition influences the priority >> with a node takes new jobs. >> Any tip will be really appreciated. Thanks a lot. > > Would PriorityWeightPartition as defined with the Multifactor Priority Plugin > (https://slurm.schedmd.com/priority_multifactor.html) help you? > > See also my summary in > https://wiki.fysik.dtu.dk/niflheim/Slurm_scheduler#multifactor-priority-plugin-scheduler > > /Ole >
[slurm-users] SLURM in Virtual Machine
Dear all, In the expansion of our Cluster we are considering to install SLURM within a virtual machine in order to simplify updates and reconfigurations. Does any of your have experience running SLURM in VMs? I would really appreciate if you could share your ideas and experiences. Thanks a lot. Cheers José
Re: [slurm-users] SLURM in Virtual Machine
Dear all, thank you for your fast feedback. My initial idea was to run slurmctld and slurmdb in respective KVMs and running while keeping the worker nodes physical. From what I see that is a setup that works without problem. However, I also find interesting some of the suggestions that you mentioned, like having worker nodes in VMs for testing and compilation purposes or even the login node. I will give some thoughts about that. Thanks a lot for the support. You are great. -- José On 12. September 2019 at 19:46:36, Brian Andrus (toomuc...@gmail.com) wrote: Well, technically I have run several clusters all in VMs because it is all in the cloud. I think the main issue would be how resources are allocated and the need. Given the choice, I would not run nodes in VMs because the hypervisor inherently adds overhead that could be used for compute. However, there are definite use cases that make it worthwhile. So long as you allocate enough resources for the node (be it the controller or other) you will be fine. Brian Andrus On 9/12/2019 7:23 AM, Jose A wrote: > Dear all, > > In the expansion of our Cluster we are considering to install SLURM within a virtual machine in order to simplify updates and reconfigurations. > > Does any of your have experience running SLURM in VMs? I would really appreciate if you could share your ideas and experiences. > > Thanks a lot. > > Cheers > > José
