It's been almost two months since we turned on full accounting with just SLURM and our own slurm-bank set of wrapper scripts. It's mostly working as we expected with allocating hours to people and projects.

However we have noticed a few things, messing with the QOS settings in slurmdbd is a global option, you can hose all your jobs across all your clusters if you aren't careful. That said, you could probably implement charge rates for a higher QOS level for users who wish to have their jobs queued up ahead of everyone elses

One of the flip sides of doing full accounting with just SLURM is that fairshare doesn't quite work anymore. Since the decay and reset period is set to one year (or never). The whole system pretty much turns into a FIFO queue where jobs that wait longer gets bumped up. We will need to tweak the settings a bit.

Overall using just slurm without GOLD has been great, things tend to just work faster with out GOLD or MAUI slowing things down with scheduling. Having just SLURM also meant that we now have access to the multi-cluster capabilities for submitting and checking jobs across multiple clusters. So far two thumbs up for SLURM and slurm-bank.


Related posts:

Alternatives to using GOLD, Maui for accounting and banking for HPC systems
Posted

The ultimate sandbox game, ever!
Posted

Astro, Particle, High Energy physics is cool
Posted

Creating a git branch with no ancestry
Posted

Using gource to visualise projects stored in version control
Posted

Bookmark and Share