Hi All,
Using OpenPBS v19.1.3 presently - we have a cluster with mixed resources (i.e. different quantity of memory and cores) and a single vnode per host. Is there a way when requesting -l place=exclhost to be able to request the entire quantity of memory within the system without manually defining it? The default if no request is defined is to use default_chunk.mem or the default as defined in the memory.default attribute within the pbs_cgroups.CF hook config.
Users don’t want to specifically request the amount of chunks required to get the default_chunk.mem since the amount is variable per node and I don’t believe you can set the default_chunk.mem on a per-vnode basis as it simply wouldn’t make sense. At present there’s no jobs that would be sharing nodes, so using however much memory within the node is fine and handled well by the application. However, there may be a requirement to share nodes in the future, so I don’t want to disable cgroups support and have to change job scripts if that requirement comes in, it also just feels wrong!
From memory in the case of other schedulers (SGE/Slurm) I’ve used, there’s an option to just request the entire node and not have to worry about requesting memory. I’m guessing this is an intricacy of the multiple-vnode-per-host situation, is there a way around it?
Or is there a way of causing the cgroups memory controller to be disabled in the case of an exclusive node?
Thanks for the help in advance!