LSF Version 7.3 - Administering Platform LSF

Administering Platform LSF 523
Running Parallel Jobs
Configure parallel fairshare
To configure parallel fairshare so that the number of CPUs is considered when
calculating dynamic priority for queue-level user-based fairshare:
NOTE: LSB_NCPU_ENFORCE does not apply to host-partition user-based fairshare. For
host-partition user-based fairshare, the number of CPUs is automatically considered.
1 Configure fairshare at the queue level as indicated in Fairshare Scheduling on
page 295.
2 To enable parallel fairshare, set the parameter LSB_NCPU_ENFORCE=1 in
lsf.conf.
3 To make your changes take effect, use the following commands to restart all
LSF daemons:
# lsadmin reconfig
# lsadmin resrestart all
# badmin hrestart all
# badmin mbdrestart
How Deadline Constraint Scheduling Works For Parallel Jobs
For information about deadline constraint scheduling, see Using Deadline
Constraint Scheduling on page 275. Deadline constraint scheduling is enabled by
default.
If deadline constraint scheduling is enabled and a parallel job has a CPU limit but
no run limit, LSF considers the number of processors when calculating how long
the job takes.
LSF assumes that the minimum number of processors are used, and that they are all
the same speed as the candidate host. If the job cannot finish under these
conditions, LSF does not place the job.
The formula is:
(deadline time - current time) > (CPU limit on candidate host / minimum number
of processors)
Optimized Preemption of Parallel Jobs
You can configure preemption for parallel jobs to reduce the number of jobs
suspended in order to run a large parallel job.
When a high-priority parallel job preempts multiple low-priority parallel jobs,
sometimes LSF preempts more low-priority jobs than are necessary to release
sufficient job slots to start the high-priority job.
The PREEMPT_FOR parameter in
lsb.params with the MINI_JOB keyword
enables the optimized preemption of parallel jobs, so LSF preempts fewer of the
low-priority parallel jobs.
Enabling the feature only improves the efficiency in cases where both preemptive
and preempted jobs are parallel jobs.