About IBM Spectrum LSF Session Scheduler
LSF Session Scheduler enables users to run large collections of short duration tasks within the allocation of a single LSF job using a job-level task scheduler that allocates resources for the job once, and reuses the allocated resources for each task. LSF Session Scheduler implements a hierarchical, personal scheduling paradigm that provides very low-latency execution. With very low latency per job, LSF Session Scheduler is ideal for executing very short jobs, whether they are a list of tasks, or job arrays with parametric execution.
While traditional LSF job submission, scheduling, and dispatch methods such as job arrays or job chunking are well suited to a mix of long and short running jobs, or jobs with dependencies on each other, LSF Session Scheduler is ideal for large volumes of independent jobs with short run times.
As clusters grow and the volume of workload increases, the need to delegate scheduling decisions increases. LSF Session Scheduler improves throughput and performance of the LSF scheduler by enabling multiple tasks to be submitted as a single LSF job.
Each LSF Session Scheduler is dynamically scheduled in a similar manner to a parallel job. Each instance of the ssched command then manages its own workload within its assigned allocation. Work is submitted as a task array or a task definition file.
- Minimize the latency when scheduling short jobs
- Improve overall cluster utilization and system performance
- Allocate resources according to LSF policies
- Support existing LSF pre-execution, post-execution programs, job starters, resources limits, etc.
- Handle thousands of users and more than 50000 short jobs per user
System requirements
- Supported operating systems
- LSF Session Scheduler is delivered in the following distribution:
- lsf10.1.0_ssched_lnx26-libc23-x64.tar.Z
- Required libraries
Note: These libraries may not be installed by default by all Linux distributions.
On Linux 2.6 (x86_64), the following external libraries are required:- libstdc++.so.6
- libpthread-2.3.4.so or later
- Compatible Linux distributions
- Certified compatible distributions include:
- Red Hat Enterprise Linux AS 3 or later
- SUSE Linux Enterprise Server 10
- IBM Spectrum LSF
-
LSF Session Scheduler is included with IBM Spectrum LSF Advanced Edition and is available as an add-on for other editions of IBM Spectrum LSF:
- If you are using IBM Spectrum LSF Advanced Edition, download the LSF Session Scheduler distribution package from the same download page as the IBM Spectrum LSF Advanced Edition distribution packages.
- If you are using other editions of IBM Spectrum LSF, purchase LSF Session Scheduler as a separate add-on, then download the distribution package from the LSF Session Scheduler download page.
LSF Session Scheduler terminology
- Job
- A traditional LSF job that is individually scheduled and dispatched to sbatchd by mbatchd and mbschd
- Task
- Similar to a job, a unit of workload that describes an executable and its environment that runs on an execution node. Tasks are managed and dispatched by the LSF Session Scheduler.
- Job Session
- An LSF job that is individually scheduled by mbatchd, but is not dispatched as an LSF job. Instead, a running LSF Session Scheduler job session represents an allocation of nodes for running large collections of tasks
- Scheduler
- The component that accepts and dispatches tasks within the nodes allocated for a job session.
Architecture
LSF Session Scheduler jobs are submitted, scheduled, and dispatched like normal LSF jobs.
When the LSF Session Scheduler begins running, it starts one LSF Session Scheduler execution agent on each host in its allocation.
The LSF Session Scheduler then reads in the task definition file, which contains a list of tasks to run. Tasks are sent to an execution agent and run. When a task finishes, the next task in the list is dispatched to the available host. This continues until all tasks have been run.
Tasks submitted through LSF Session Scheduler bypass the LSF mbatchd and mbschd. The LSF mbatchd is unaware of individual tasks.
Components
LSF Session Scheduler comprises the following components.
LSF Session Scheduler command (ssched)
The ssched command accepts and dispatches tasks within the nodes allocated for a job session. It reads the task definition file and sends tasks to the execution agents. ssched also logs errors, performs task accounting, and requeues tasks as necessary.
sservice and sschild
These components are the execution agents. They run on each remote host in the allocation. They set up the task execution environment, run the tasks, and enable task monitoring and resource usage collection.
Performance
| Average Runtime (seconds) | Recommended maximum allocation size (slots) |
|---|---|
| 0 | 12 |
| 5 | 64 |
| 15 | 256 |
| 30 | 512 |