Re: [HTCondor-users] how to use rank with memory

[Todd Tannenbaum <tannenba@xxxxxxxxxxx>]

On 6/24/2021 8:42 PM, Myunggi Yi wrote:

Dear users,

I have installed HTCondor 9.0.1

I want the jobs to run on more memory machines.

I submitted a job with the following script, but the job always goes to node07, which has less memory.

How can I achieve my goal?

    Hi,
    
    A couple of thoughts:
    
    1.  I assume the node with more memory was available (unclaimed) at
    the time you submitted your job?  Realize that setting up a Rank in
    your submit file only sorts amongst resources that are currently
    available (unless you tell HTCondor to allow preemption of jobs
    based on rank - you can do this, but in practice very few sites
    would want to).    Rank is really only helpful on pools that are
    lightly loaded.  On busy pools where most of the nodes are busy
    doing something all of the time, Rank becomes pretty useless without
    preemption, because at any given moment there may only be one or two
    free slots to pick from.  In these scenarios, you really want to use
    Requirements instead of Rank.
    
    2. The administrator of your central manager gets to take first stab
    at ranking slots before the job does.  Basically, it is a
    multi-level sort where matching slots are first sorted by config
    knob NEGOTIATOR_PRE_JOB_RANK, then sub-sorted (i.e. ties from the
    first sort are handled) by the job's Rank _expression_, and then
    sub-sorted by config knob NEGOTIATOR_POST_JOB_RANK (details in the
    Manual at https://tinyurl.com/yzcr5ocq). Suggest you login to your
    central manager and enter the command:
        condor_config_val -v
      NEGOTIATOR_PRE_JOB_RANK
    On my machine, this returns the following:
      NEGOTIATOR_PRE_JOB_RANK = (10000000 *
      My.Rank) + (1000000 * (RemoteOwner =?= UNDEFINED)) - (100000 *
      Cpus) - Memory
         # at: <Default>
    Note that the final clause is referencing CPUs and Memory, in
    an attempt to do depth-first allocation of nodes.  On your pool,
    perhaps you want to get rid of this behavior by setting the
    following in the config of your central manager:   
     
        NEGOTIATOR_PRE_JOB_RANK = (10000000 * My.Rank) + (1000000 *
        (RemoteOwner =?= UNDEFINED))
      and then doing a condor_reconfig as usual.
    
    Hope the above helps,
    Todd