Back to Gap Map Infrastructure / ML orchestration

IBM Spectrum LSF

closed / Overall score: n/a

Commercial workload manager and job scheduler for distributed high-performance computing. Jobs declare what they need in LSF's own submit schema - a GPU count, mode, model and memory, or a resource requirement string naming physical GPUs and interconnect - and LSF's scheduler places each job on hosts whose advertised resources satisfy it, autoscaling onto cloud capacity where a site has configured it.

IBM publishes LSF's documentation on ibm.com/docs, which returns HTTP 403 to this repository's fetcher on every header combination while serving curl normally - a TLS-level block rather than a policy, and the site is client-rendered besides. The evidence here is drawn from IBM-published sources that are reachable: the product page on ibm.com/products, and IBM's own lsf-python-api examples on GitHub. Verified 2026-09-16 via the IBM Spectrum LSF Suites product page and IBM's lsf-python-api GPU submission example.

Openness

1 high confidence

1.0

license: Proprietary
service: proprietary(IBM commercial suite, licensed per site)
source: closed(no implementation is published

A commercial IBM suite. IBM publishes a Python API wrapper and integration code on GitHub, but those drive the scheduler rather than implement it, and no source for LSF itself is published. Nothing about a license can help where there is no self-hostable implementation, which is why the ladder settles this on source alone.

https://www.ibm.com/products/hpc-workload-management recorded 2026-09-16
IBM's own product page: 'IBM Spectrum LSF Suites is a workload management platform and job scheduler for distributed high performance computing (HPC)', sold as a suite with a separately downloadable Community Edition. Establishes a commercial product with no published implementation.

Adoption

not assessed

No usage figure is published for LSF, and it publishes no countable artifact - no package, repository or registry entry for the scheduler itself. IBM's lsf-python-api repository is a wrapper rather than the product, so its stars would measure a different thing. No level is assigned rather than one being inferred.

https://www.ibm.com/products/hpc-workload-management recorded 2026-09-16
The IBM Spectrum LSF Suites product page, read for a usage disclosure; it names capabilities and a roadmap but publishes no count of sites, clusters or customers.

Capability

4 medium confidence

4.0

Admitted on the accelerator limb: the GPU request is LSF's own submit schema, not a backend spec passed through, and LSF's scheduler consumes it to choose the hosts. Band 4 - it selects the hardware, down to GPU model and memory, and nothing reachable establishes a multi-machine set reserved as one allocation, which is band 5. LSF does have multi-host allocation, but the documentation that would evidence it sits on ibm.com/docs, which this repository cannot fetch; the band records what is citable rather than what is likely.

https://raw.githubusercontent.com/IBMSpectrumComputing/lsf-python-api/master/examples/submit_gpu_job.py recorded 2026-09-16
IBM's own lsf-python-api GPU submission example: submitreq.options4 = lsf.SUB4_GPU_REQ with gpuOpt entries JDATA_EXT_GPU_NUM, JDATA_EXT_GPU_MODE, JDATA_EXT_GPU_MODEL, JDATA_EXT_GPU_MEM and JDATA_EXT_GPU_TILE, and the commented alternative resReq = 'rusage[ngpus_physical=2:gmodel=K80#12G:nvlink=yes]'. Establishes that the accelerator request is expressed in LSF's own schema and submitted to its scheduler.
https://www.ibm.com/products/hpc-workload-management recorded 2026-09-16
IBM's product page describing LSF as a job scheduler for distributed HPC with automated NVIDIA GPU detection and configuration, and autoscaling onto cloud capacity under scheduling policies.

Verified 2026-09-16