Overview
The batch system at SLAC uses the IBM Platform Load Sharing Facility, LSF, and is made up of
- a general farm of batch servers that is open to all SLAC users
- an mpi farm which requires that you request access (send email to unix-admin) and run only mpi jobs
various other farms of machines that belong to individual scientific computing groups.
Various scientific groups at SLAC contribute to the purchase of the general farm batch systems and jobs submitted to run in the general farm are scheduled according to a fairshare priority structure. If you are not part of a computing group that has a fairshare, your jobs will run according to the priority of the group called AllUsers. More information about fairshare scheduling at Slac can be found here https://confluence.slac.stanford.edu/display/SCSPub/Fairshare+Scheduling
Submitting Jobs to the General Farm
To submit jobs to the general farm you will need to
- have a SLAC unix account http://www2.slac.stanford.edu/comp/slacwide/account/account.html
- login to a SLAC public machine using ssh
The public and general farm batch machines are described here http://www.slac.stanford.edu/comp/unix/public-machines.html . The simplest command to submit a job to the general farm is
bsub <your job>
This command will submit your job to the first available general farm machine and will be able to run for 20 minutes of wall clock time. It is recommended to run your job specifying a wall clock time since LSF will use that information to make use of windows of opportunity on systems that are accumulating cores for mpi jobs that are not yet running, and your jobs will schedule more quickly. To specify a wall clock time:
bsub -W <time in minutes> <your job>
bsub argument | requirement | description |
---|---|---|
-q <queue_name> | optional | Specifies a job submission queue. Not required for running jobs in the general shared clusters. |
-W <[hours]:minutes> | required | Wallclock runtime limit that is not normalized for CPU differences. Essential for efficient job scheduling. Jobs will be terminated if they exceed this runlimit |
-We <[hours]:minutes> | optional | Wallclock estimated runtime limit that is not normalized for CPU differences. The system will consider this estimate for scheduling purposes. Jobs may be not terminated immediately if they exceed this limit |
-c <[hours]:minutes> | optional | cputime limit that is normalized by the CPU factor of the assigned core. Intended to prevent runaway jobs but it is not used to schedule time on cores. |
General Queue Run Times
Queue
| Default Runlimit | Maximun Runlimit | Priority |
---|---|---|---|
EXPRESS | 4 min | 4 min | 200 |
SHORT | 20 min | 60 min | 185 |
MEDIUM | 30 min | 2 days | 180 |
LONG | 30 min | 5 days | 175 |
IDLE | 12 hours | 12 hours | 5 |
bulletmpi * | 15 min | 7 days | 187 |
bulletmpi-large * | 15 min | 1 day | 187
|
*The bulletmpi and bulletmpi-large queues are for mpi jobs only. Access can be requested by email to unix-admin.
Fermi/GLAST and batch
Information about using batch at SLAC for the Fermi/GLAST user can be found here: http://www.slac.stanford.edu/exp/glast/wb/prod/pages/installingOfflineSW/usingSlacBatchFarm.htm
Documentation
There are manpages available for the various batch commands and there is full documentation for LSF here https://www.ibm.com/support/knowledgecenter/SSWRJV_10.1.0/lsf_welcome/lsf_welcome.html .