GridAustralia Requirements for GLUE 2.0

GridAustralia Story for the Software Package entity

• Story 1
  • Card: A user wants to know if software package X is installed on a compute resource somewhere on the Grid
• Story 2
  • Card: A user wants to know the available versions of software package X installed on compute resources somewhere on the Grid
• Story 3
  • Card: A user wants to know which sites are running software Package X with version Y
• Story 4
  • Card: A user wants to know if VO V have permissions of submitting a job which will use package X at site A
• Story 5
  • Card: A user wants to know if software package X contain a parallel or sequential version of software executable Z
  • Conversation: Is the a software executable Z under the software package X installed on resource R at Site A support parallel/sequential execution?
• Story 6
  • Card: A user needs to know what module to load to set the user environment is he/she wants to use software package X at site A
• Story 7
  • Card: A user wants to know the maximum number of CPU that can be used if software package X with executable Z is used at site A
• Story 8
  • Card: A user wants to know what type of parallel application software executable Z support (mpi, openmp, etc)
• Story 9
  • Card: A user wants to know the number of free CPUs available for him/her (as a member of VO V) to use when he/she submits a software package X job at site A.
• Story 10
  • Card: A user wants to know the number of free job seats available for him/her (as a member of VO V) to use when he/she submits a software package X job at site A
• Story 11
  • Card: A user wants to know the number of free user seats available for him/her (as a member of VO V) to use when he/she submits a software package X job at site A

GridAustralia Story for the ComputingElement and SubCluster entity

• Story 12
  • Card: A user wants to know which CE he/she can submit his/her jobs to so the jobs run one of the subclusters of the cluster
• Story 13
  • Card: A user wants to know the network interconnect being used by the Cluster/SubCluster, eg GigE, Quadrics, Myrinet. If Infiniband comes to dominate, maybe we need the 4X, 12X designation as well.

GLUE 1.3 Issues/Comments

1. Equating jobslots with CPUs is sloppy use of the language. To a user, a job is a reasonably well defined term, and if my job uses 32 CPUs I don't really expect to think of this as 32 jobslots. Why not use CPUs - it's a natural concept, and if I am using MPI, PVM, OpenMP, etc, I need to know this number anyway. The argument over virtual vs real CPUs is probably moot - I doubt users are allowed this degree of control on production batch systems anyway. I would be interested to know if _any_ production cluster offers more virtual CPUs through hyperthreading than there are real CPUs.
2. If GLUE 1.3 provides preemption information through the preemption (t/f) field in the CE, the definition needs to be clear if it supports resumption or not OR if the definition of this field will depend on the policy of the implementing Grid again. Having a preemptive system that doesn't support resumption will make a big difference. Any system which uses reservation will kill running jobs when the new reservation starts. That's why the big US supercomputer centres like PSC have relatively short time limits (12 hours or so) and expect user jobs to be restartable. So a precise definition of preemption is needed.
3. WorstResponseTime. For a preempting system, this is infinite, as subsequent jobs arriving could have higher priority and therefore prevent my submitted job from starting.
4. Similar to EstimatedResponseTime, many other policy settings are user/VO specific. Although there is the VOView, what happens if a user is not part of a formal VO - can still have individual limits. How will GLUE represent this type of user? Do they need to create another VO for this user?
5. PhysicalCPUs and LogicalCPUs. Not an issue for simple hyperthreading processors probably (for HPC you turn hyperthreading off on x86) but what about more highly scalable threaded processors such as MTA or Sun's Niagra? Is this distinction available on real queueing systems?
6. The storage model assumes POSIX filesystems. What about higher level concepts such as SRB or RDBMS? Where does OGSA-DAI fit in to this?

Last minute stories

• Story 14
  • Card: A user wants to know if software package X supports parallel or sequential execution