dynamic resizing of pilot resource allocation

[dsschult]

If the pilot is running multiple task and one goes over memory, consider a priority list for which to keep:

• GPU task
• significantly longer-running task
• high memory task

[gonzalomerino]

Related to this, it might be useful to have the capability to configure the iceprod pilot so that in a gpu node it grabs all resources (brian's submit classsds to request all available cpus...) and then it runs the cpu jobs with less mem requirements, so that the gpu job can grow. Also, for the resource usage check, is Iceprod comparing usage with requested resources or with available resources? Being able to do the latter would be nice. Gonzalo

…

On Thu, Mar 9, 2017 at 12:01 PM David Schultz ***@***.***> wrote: If the pilot is running multiple task and one goes over memory, consider a priority list for which to keep: - GPU task - significantly longer-running task - high memory task — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#179>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHjryeOo7GVynniFx1AZzXVey77E6LMLks5rkD5zgaJpZM4MYbdB> .

[dsschult]

It compares with both requested and available.