Re: [Nelsonlab-analysis] SGE Quotas and Host prioritization (follow up from lab meeting)

[Jordan M. <jme...@uc...>]

I made one additional change, so that external users will only be allowed to
use their 8 machines if they are otherwise idle. This means that when the
cluster is full and people are waiting, the queued jobs of the external
users will have the lowest priority and will not be scheduled unless a node
is otherwise going to sit idle. When the cluster is empty, they will still
only be able to use up to 8 nodes, to prevent them filling up slots that one
of us might otherwise want to use later on.

https://secure.genome.ucla.edu/index.php/Sun_Grid_Engine#Setting_a_user_to_only_use_idle_resources

Jordan

On Tue, Feb 23, 2010 at 6:57 PM, Jordan Mendler <jme...@uc...> wrote:

> Hi all,
>
> I think I resolved two issues, so please keep an eye on the cluster and
> your jobs to make sure nothing breaks:
>
> 1) I added a quota to the external users, so that each can not use more
> than 64 slots at a single time. Each machine has 8 slots, so this equates to
> 8 full machines.
>
> 2) I resolved the issues of host prioritization, without having to add a
> queue. When a user submits a job, they are scheduled as follows:
>     - By default, the job will use a low memory node
>     - If all low memory nodes are full, or the job requires more than 8GB
> of RAM, it will use one of the faster high memory nodes (compute-3-x) so
> that it ends ASAP.
>     - If all compute-3-x nodes are full and the job needs more memory, then
> it will be scheduled to a slower high memory node (compute-2-x).
>
> In my testing, everything seems to work properly, but please let me know if
> you see anything weird happening. If you want to know the gory
> implementation details, I added them to
> https://secure.genome.ucla.edu/index.php/Sun_Grid_Engine
>
> Cordially,
> Jordan
>