Improvements to Fair Share job prioritisation on Māui

On Thursday 3 September 2020, NeSI updated the way we prioritise jobs on the Māui HPC platform.

Background

Since the start of the year, we have been using Slurm's Fair Tree algorithm on Māui (not yet on Mahuika) to prioritise jobs. This provides a hierarchical structure to Slurm's account management, with the hierarchy representing shares of a total cluster under Slurm's control. This enables control of higher level or aggregate account considerations, such as ensuring a group of projects within a research programme or institution are ensured access to their share of a cluster.

Under our Fair Tree implementation, each of NeSI's four collaborating institutions is assigned a percentage share of Māui, alongside a percentage share for MBIE's Merit allocations (including Postgraduate and Proposal Development allocations), and the remainder as a share to allocations coming from subscriptions.

These six shares, or what we in NeSI call national pools, are then ranked in order, starting with the pool that has been using at the lowest rate compared to its allocated percentage share. See this page (off site) for more details about Slurm's Fair Tree algorithm.

Previously, we had given each pool a hard-coded share of Māui use. These hard-coded shares did not reflect ongoing rounds of allocations given to projects, and so some researchers were suffering from deprioritised jobs. These jobs ended up delayed in the queue, sometimes excessively.

What has changed?

We have now recalculated the shares for each pool to take into account the following:

  • The investments into HPC platforms by the various collaborating institutions and by MBIE;
  • The capacity of each HPC platform;
  • The split of requested time (allocations) by project teams between the Māui and Mahuika HPC platforms, both overall and within each institution's pool.

Under this scheme, any job's priority is affected by the behaviour of other workload within the same project team, but also other project teams drawing on the same pool. In particular, even if your project team has been under-using compared to your allocation, your jobs may still be held up if:

  • Other project teams at your institution (within your pool) have been over-using compared to their allocations, or
  • Your institution has approved project allocations totalling more time than it is entitled to within its pool's share.

What will I notice?

If your institution or pool's ranking has not changed, nothing much will immediately change for you.

However, if your institution or pool's assigned share of the machine has increased, it will become easier to move up the priority rankings, at least in the short term.

Conversely, if your institution or pool's assigned share of the machine has decreased, it will become easier to move down the rankings. This change is one you are more likely to notice over time.

Whenever your institution or pool's ranking changes, whether because of usage or because we adjust the assigned shares based on ongoing rounds of allocations, your job priorities will alter almost immediately. Moving up the rankings will increase your job priorities. Moving down the rankings will decrease your job priorities.

What other changes are NeSI planning on making?

We are looking at introducing Fair Tree on Mahuika as well, though not on Māui ancillary nodes. We will announce this change well ahead of any planned introduction.

We will also adjust the assigned Fair Tree shares on Māui routinely so we don't diverge from allocations across HPC platforms again.

Labels: announcement
Was this article helpful?
0 out of 0 found this helpful