Developed by Seren Soner and Can Ozturan from Bogazici University, Istanbul, Turkey as a part of PRACE WP9.2
Contact [email protected], [email protected] if necessary
-
IMPORTANT NOTE: sched currently works with only CR_CPU with GresTypes=gpu (optional) nodes should be SHARED, overcommit is not allowed.
-
USE CPLEX WHENEVER AVAILABLE. In order to use cplex, first go to the plugins/sched/lpsched directory, copy Makefile.am.cplex to Makefile.am. If you wish to use lpsolve, copy Makefile.am.lpsolve to Makefile.am. Afterwards, run autogen.sh, ./configure, and make install. If CPLEX is selected, the license address of cplex should be given in slurm.conf in the following way:
SchedulerParameters=cplex_lic=/share/apps/ILOG/CPLEX_Studio/AcademicResearch122/licenses/access.ilm -
max_job_count can also be set to limit the number of jobs handled in every single ILP solve of the scheduler. By default, it is 200.
- from the queue, do not start the first job, but take a set of jobs
- test all of them at once to see selection of which of these jobs would yield the best result
-
patches on gres.c and gres.h add two functions, which return number of gres types defined, and the number of gpu's a specific job requests. also, one function is modified to return number of gpu's available in that node, instead of not returning anything.
-
patch on job_scheduler.c changes the job scheduler so that, when a job is submitted, its not immediately considered by the job scheduler; but waits for the next run of IP solve to be allocated.
-
patch on makefile.am files allow the makefiles to be generated automatically.
- Feb 6, 2011: Restructuring.
- Jan 31, 2011: Bug fixes.
- Jan 10, 2011: cplex and lpsolve can now be compiled dynamically.
- Dec 5, 2011: gpu support added. for this, common/gres.c was modified and some new functions were added. it returns the job's gpu request and node's availab le gpu count to the lpsched scheduler.
- Dec 4, 2011: submitted jobs are not scheduled right away currently. they should first go through solve_allocation(). this required a change in the core cod e in the job_mgr.c line 2852. It does not call select_nodes() anymore.