[Beowulf] LavaLite 1.1.0 – a lightweight open-source HPC scheduler

[Lu Tram]

Hi all,

I wanted to share the new release of LavaLite, a lightweight open-source
HPC batch scheduler written in C.

Version 1.1.0 is now available:

https://github.com/LavaLite-Compute/lavalite
Release tag: `lavalite-1.1.0`

The project started from the idea that a useful HPC scheduler does not
necessarily have to become a very large distributed system. LavaLite tries
to keep the traditional batch scheduling model understandable and
operationally simple while supporting the workloads expected from an HPC
cluster.

It uses familiar HPC batch commands such as bsub, bjobs, bkill, bhosts,
bqueues, and bhist, while the implementation is a new codebase with its own
architecture and design philosophy.

The current scheduler supports CPU, memory, storage, host and GPU
scheduling, cluster-wide consumable resources such as license tokens,
cgroup v2 resource enforcement, durable event-based state, restart and
recovery, accounting, and job history.

The main additions in 1.1.0 are job arrays and job dependencies, including
dependency expressions for building multi-stage workflows.

There is no external scheduler database and relatively little machinery
around the core daemons: one master daemon (mbd) and execution daemons
(sbd) on the compute nodes.

At this stage I am particularly interested in feedback from people who have
operated HPC clusters: what looks useful, what looks wrong, and whether the
deliberately small architecture is interesting in practice.
The repository includes the administrator and user guides, release notes,
manual pages, and the source itself.

Feedback, criticism, testing, and contributions are all very welcome.

Lu Tram

LavaLite
lavalite at lavalite-compute.io
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://beowulf.org/pipermail/beowulf/attachments/20260826/4c3f4816/attachment.htm>