Alpine: condo FairShare and resource access¶
This purpose of this document is to provide a detailed overview of Alpine’s FairShare and resource access policies. The intended audience is Alpine’s insitutional-level condo contributors.
Updated November 4, 2022
Goals and philosophy¶
The Alpine supercomputer hosted at CU Boulder and administered by Research Computing at CU Boulder (CURC) supports institutional condo purchases, the first two of which come online in the fall of 2022. This document outlines how jobs affiliated with condo contributors will be assigned priority and billing weight on the system. It does not aim to address FairShare scoring for individual users and allocations, as these involve more detail and complexity than this document intends to cover.
The goals of this policy are to maximize overall system utilization, assign appropriate priority and billing weights proportional to contribution, and maintain a straightforward and accountable configuration. This document describes the initial policy, but changes to any or all of it are possible based on contributor feedback, user issues, CURC’s own data-gathering and monitoring, and other factors.
Node, partition, and FairShare policies¶
All nodes, whether contributed by CU Boulder or by another institution, will be added to a shared partition to which all Alpine users will have access. This does not preclude adding nodes to one or more additional partitions with different settings and attributes. At minimum, CURC will create one dedicated partition and QOS per contributor.
By configuring all nodes into a common pool by default, CURC intends to better maximize system utilization and ensure that users are not locked out of Alpine due to (for example) power or networking issues that only affect their institution’s nodes. Users will always have the option to request specific node attributes for their jobs – for example, presence of GPUs, InfiniBand, or high clock speeds – and those requests will remain part of all scheduling calculations.
Through a combination of QOS’s, trackable resources, partitions, and other Slurm settings, users will receive some preference for running jobs on their own nodes. However, preemption of jobs will not be supported.
As one of several determinants of job priority, CURC will assign a FairShare score in the Slurm database for each condo buy-in on Alpine based on the institution’s total contribution to the system. This will account for total cores, clock speeds, cores on nodes with high memory (>=1TB), and GPUs. All resources will be billed back accordingly at the time of job submission.
To compute the GPU term in the FairShare equation, CURC will multiply total contributed GPUs by an “acceleration factor”. This factor is derived from the MATLAB GPU benchmark series, which CURC has used to measure computational performance on a variety of tasks on a representative GPU against the equivalent on a CPU. As of November 2022, the acceleration factor of an NVIDIA A100 or AMD MI100 GPU will be 108.6, meaning that our benchmarks indicate a GPU will provide 108.6x speedup (on average) over a CPU. Scaling this to a typical Alpine node, the number of SUs allocated for a 64-core node with 3 GPUs and 64 cores would be a factor of 5-6 higher than a 64-core CPU-only node. This figure and the choice of benchmarking software are subject to change in the future based on new information.
The exact score will be derived from the following equation:
FairShare = floor ( (standard node CPU core hours contributed * average clock speed / minimum clock speed on Alpine) + (GPUs contributed * GPU acceleration factor) # changed in this version + (high-memory node CPU core hours contributed * average clock speed in MHz / minimum clock speed available * mem-per-core_high-mem / mem-per-core_standard=256GB))
To distribute the FairShare score, each institutional contributor will receive an account (containing all its affiliated users) that will hold the full score. In turn, each of those accounts will contain a “general” subaccount that will have 20% of the FairShare pool and a “projects” subaccount with 80%. (The general/projects pattern is already in place on Alpine and is a pattern CURC initially set up on Summit.) Contributors wishing for a different accounting arrangement from “general/projects” should reach out to CURC with specific requests.
This crediting and billing system is intended to balance contributions and utilization.
Example and comment on FairShare and billing¶
Suppose the lowest rated clock speed available on any Alpine node is 2.2GHz. Acme University contributes the following order:
- 16x 64-core 3.2GHz CPU nodes with 256 GB RAM
- 2x GPU node with 3 A100 GPUs, same CPU
- 4x 1TB high-memory nodes, same CPU
For this order, Acme would receive the following score:
Acme University FairShare = floor ( 16 * (64+2) * 3200MHz / 2200MHz # 1536 + 2 * 3 GPUs * 108.6/GPU # 651.6 + 4 * 64 * (1028 GB RAM / 256 GB RAM)) # 1028 = 3,215
Note that prior to the change described in this policy, the GPU term was 2 x 3 x 6912 = 41,472.
These scores will be revised upon new nodes being contributed.
GPU, CPU, and memory (along with potentially other resources in the future) will be billed to users in proportion to their weight in the fairshare calculation. For example, a partition might be configured with any of the following resource billing weights:
TRESBillingWeights="CPU=1.0,GRES/gpu:a100=108.6" TRESBillingWeights="CPU=1.0,GRES/gpu:mi100=108.6" TRESBillingWeights="CPU=1.0,Mem=4.0"
Additionally, CURC reserves the authority to modify this policy or specific jobs if users abuse it – if, for example, a contributor purchases many GPU cores and uses their share of “credit” from those GPUs to exploit another contributor’s CPU-only contributions.
CURC will provide regular reports to institutional Alpine partners to demonstrate they are receiving appropriate value for their contributions. The full content of these reports is beyond the scope of this document, but it will include measurements of resource utilization, statistics useful for quality assurance, and the number and size of jobs run.
CURC will generate automated reports monthly and share them through appropriate channels, either written or in-person. CURC will also produce more detailed reports examining the overall value of the system to condo contributors once per quarter. Reports may be addressed to the institution or, if applicable, to the condo buyer within the institution (for example, a PI or department).
Note that some Slurm usage querying tools are also available for general
systems users, including
Public-facing reports and dashboards are available online through RC’s
instance of XDMoD.
Please note the following about the way Slurm calculates priority:
- A job’s priority is based on multiple factors in addition to the FairShare score, including (but not limited to) job age, resources requested, job size, and QOS.
- The FairShare component of the final priority calculation is determined by normalizing the institution’s FairShare against the highest FairShare on the system.
- FairShare scores for sub-accounts (e.g., acme-projects and acme-general) are normalized against each other. These sub-accounts will be assigned scores at an appropriate 80/20 ratio, which will not need to be changed to accommodate expansions or new contributions.
In most but not all cases, changes to FairShare scores, QOSs, and other Slurm settings are straightforward to implement. Systems administrators can apply these changes quickly once CURC and the contributor(s) agree upon them.
For additional details about how Slurm implements this score, please consult the official Slurm documentation (in particular, docs pertaining to the “Multifactor Priority Plugin”) or follow up with CURC with specific questions. This policy is now in effect and is subject to change at any time.
Alpine is jointly funded by the University of Colorado Boulder, the University of Colorado Anschutz, Colorado State University, and the National Science Foundation (award 2201538).
Couldn’t find what you need? Provide feedback on these docs!