LSF Version 7.3 - Administering Platform LSF

Administering Platform LSF 357
Goal-Oriented SLA-Driven Scheduling
SLA goal behavior Only velocity goals are supported in EGO-enabled SLAs. Deadline goals and
throughput goals are not supported. For EGO-enabled SLA, the configured velocity
value is considered to be a minimum number of jobs that should be in run state
from the SLA.
This is different from a regular SLA, where once the velocity is reached, no more
jobs are dispatched by the SLA. Under EGO-enabled SLA, if pending jobs exist, the
SLA will try to run them all. If the goal becomes inactive before all jobs are
dispatched, LSF continues to dispatch the remaining pending jobs until all the jobs
under the SLA are finished, then returns the host resources to EGO after the idle
timeout.
TIP: To ensure a steady flow of jobs through the system for your velocity goals, you should set the
ownership for the consumer representing the SLA to a value equal to the desired velocity.
If the status of the SLA goal is inactive when the job is submitted, the job will remain
pending and LSF will not attempt to get hosts from EGO. This is different from
regular SLAs, where jobs submitted to an inactive SLA are dispatched as if they were
not submitted to any SLA.
How LSF gets host
resources from EGO
Each SLA computes its goal and how many hosts it needs for LSF to request from
EGO. Hosts to be allocated to SLAs are configured in EGO resource groups.
Allocated hosts are added to the LSF host group automatically created for the SLA,
which can then start its jobs.
By default, EGO allocates an entire host to LSF, which uses its own MXJ definition
to determine how many slots are available on the host. Configure
MBD_USE_EGO_MXJ in
lsb.params to configure slot-based allocation.
How host resources
return to EGO
Hosts are returned from LSF to EGO two ways:
Voluntarily All allocated hosts are released to EGO if no jobs are running on the host after the
default idle timeout of 120 seconds (configurable by MAX_HOST_IDLE_TIME in
lsb.serviceclasses). If you configure MBD_USE_EGO_MXJ in lsb.params to
configure slot-based allocation, LSF returns idle slots to EGO.
Reclaim EGO can reclaim hosts allocated to LSF service classes at any time and reallocates
them to other consumers. When LSF receives a reclaim request from EGO, it kills
and requeues the jobs to pending state on the reclaimed host. If you configure
MBD_USE_EGO_MXJ in
lsb.params to configure slot-based allocation, EGO
reclaims the allocated slots.
Configure EGO-enabled SLA scheduling
lsb.params Minimal configuration enables SLA on EGO. Advanced configuration in controls
LSF behavior on EGO.
lsb.serviceclasses Service class configuration maps an SLA to a specific consumer in EGO. Advanced
configuration controls host idle period and SLA evaluation period.