Below please find the wishlist and questions the main SLAC batch system users and administrators posed.
Followup Questions
- What kind of support (transition/longterm) do we get for the quoted price?
- How much downtime do we incur if we need to "restart the system"
- What sort of activities require such a "restart" (e.g. creating a new queue?)
- Do large numbers of short jobs (about 1 minute) cause problems?
- Please provide additional information about support for virtualization, and any plans for future enhancements to that support.
- What would be involved in adding support for Mac OSX or Windows? Would that be considered normal "support"?
- Where are the bottlenecks in the system likely to occur? For example, can a user stress the system by repeatedly and frequently querying job stats?
Original Wishlist
User Group 1
- Automatic job preemption/suspend/resume?
- Support for multiple-levels of job preemption (e.g. 3-queue hierarchy)?
- Job environment propagation (including limits like "stacksize")?
- Subgroup-specific priority calculation (queue-specific priority formula)?
- Capability to delegate subgroup administration privileges (adjust job priorities, suspend, resume, kill) to subgroup administrators?
- Cross-queue fairshare (with cpu-speed weighting)?
- CPU advanced reservations for MPI?
- GPU support?
- Ability to submit jobs to hosts where we don't have accounts/home-directories?
- Avoid bad behavior when MPI head node reboots: slave node processes get "forgotten" ?
- How well does the system scale?
- Number of cores, queues, queued and running jobs?
User Group 2
- Please list supported operating systems (for submission hosts and for execution hosts)
- Explicit, site-specific-naming resource specification (e.g., rhel5-64, amount of /scratch space, etc.) at job submission time?
- Is there an API for "time remaining" query (to save state near end of job)?
- Do submission / management hosts require a license?
- ability to submit/monitor jobs from any machine at SLAC (not just those with licenses as is the case with LSF)
...
- Integration with AFS (i.e. it must be possible for single-node and MPI jobs to acquire the user's kerberos and AFS identities)?
- Does the system (at a minimum) provide hooks to implement such a feature?
- Real-time and historical monitoring , accounting and analytics (performance, utilization, resource consumption, etc) available through web interface, command line, other APIs (e.g. for embedding in other web applications).
- It is fine if this is an add-on capability
- Short list of resource allocation / scheduling schemes (such as fairshare, backfill, etc.) with brief explanations
- Job submission and management through a web portal?
- Duration and limitations of trial licenses?
User Group 8
- How are licenses managed (i.e. is there a requirement for a license server, if yes, does it allow for redundancy)?
- Ability to schedule jobs based on file server load?