LSF Version 7.3 - Using Platform LSF HPC

◆ Kernel-level checkpointing is not available on Linux/QsNet systems.
◆ When LSF selects RMS jobs to preempt, jobs to be preempted are selected from the
list of preemptable candidates based on the topology-aware allocation algorithm.
Allocation always starts from the smallest numbered node on the LSF node and
works from this node up. Some specialized preemption preferences, such as
MINI_JOB and LEAST_RUN_TIME in the PREEMPT_FOR parameter in
lsb.params, and others are ignored when slot preemption is required.
◆ Preemptable queue preference is not supported.
◆ Jobs submitted to a chunk job queue are not chunked together, but run as a normal
LSF job.
◆ User-level account mapping is not supported.
◆ By default, job start time is not accurately predicted for jobs reserving resources
with topology requirements, so the forecast start time shown by
bjobs -l is
optimistic. LSF may incorrectly indicate that the job can start at a certain time, when
it actually cannot start until some time after the indicated time.
For a more accuration start-time estimate, you should configure time-based slot
reservation. With time-based reservation, a set of pending jobs will get future
allocation and estimated start time.
See Administering Platform LSF for more information about time-based slot
reservation.
◆ When a partition is blocked or down, the status of prun jobs becomes
UNKNOWN, and
bjobs shows the jobs as still running.
◆ The administrator must use brun -c to force a job to run on an RMS host. If the
RMS allocation cannot be satisfied for any reason, the job will be dispatched, but
will be requeued and returned to pending state. The administrator can use
brun -c
again to start the job.
◆ In MultiCluster lease model, you should export entire LSF hosts.
Compatibility with earlier releases
In this version of Platform LSF HPC for Linux for Linux/QsNet:
◆ The topology scheduler options (bsub -extsched or queue-level
DEFAULT_EXTSCHED and MANDATORY_EXTSCHED) no longer
determine the host type for the job. Regular Platform LSF resource requirements
(
busb -R or queue-level RES_REQ) or host name selection (bsub -m) dictate
where the job will run.
◆ The default RMS allocation type is RMS_SNODE.
◆ Node-level allocation parameters LSB_RLA_POLICY and
LSB_RMS_NODESIZE in
lsf.conf are no longer needed. Use the
RMS[ptile= cpus_per_node ] option to specify node-level allocation.
◆ For topology scheduler options that use the obsolete syntax:
❖ Only the RMS allocation options (RMS_MCONT, RMS_SLOAD,
RMS_SNODE) are honoured
❖ Options from the queue and from bsub are parsed separately before they are
merged. For example:
bsub -extsched "ptile=1" ...