guest ::
| Home > Publications database > The Art of Process Pinning: Turning Chaos into Core Harmony |
| Home > Publications database > The Art of Process Pinning: Turning Chaos into Core Harmony |
| Poster (After Call) | FZJ-2025-02926 |
; ; ; ; ;
2025
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2025-02926
Abstract: High-Performance Computing (HPC) centres face growing challenges as user numbers and application diversity increase, requiring systems to manage a wide range of workflows. While users prioritise scientific output over specific configurations, administrators strive to maintain fully utilised systems with optimised jobs, avoiding resource waste. However, no single default environment can address the diverse needs of users and applications due to the complex landscape of unique use cases. Process pinning - the binding of tasks and threads to specific CPU cores - is a vital yet often overlooked optimisation that significantly improves job performance. This technique benefits both CPU-intensive and GPU-enabled jobs. Proper pinning prevents process migration, ensures efficient memory access, and enables faster communication, improving system performance by simply adjusting workload manager parameters (e.g., Slurm) without altering code. Metrics from various applications and benchmarks show that suboptimal pinning can drastically reduce performance, with production scenarios likely impacted even more. Achieving optimal process pinning is challenging due to three interrelated factors: - System side: Application layers and libraries (e.g., MPI, OpenMP, Slurm) interact with hardware architectures, affecting task and thread placement. Updates to these components can disrupt the expected pinning behaviour. - User side: Users must consider system architecture and configuration options, such as how to split processes and threads or distribute them across cores. Even with the same core usage pattern, distribution can vary based on workload options (e.g., Slurm `cpu-bind` and `distribution` values). Portability across systems is not guaranteed, often leading to suboptimal performance. - Operator side: Administrators and support staff must monitor systems to ensure effective resource utilisation and address issues proactively. Identifying problematic jobs is difficult due to the variety of characteristics, with inefficiencies often hidden in core usage patterns. We developed tools and processes based on investigations across diverse HPC systems to address these challenges. These solutions enhance overall system throughput by identifying binding errors, guiding users in optimisation, and monitoring core usage. Our solutions include: - A workflow that validates pinning distributions by running automated test jobs, periodically or manually, via the GitLab-CI framework. Results are compared to expected outputs, with summaries generated and full comparison displayed on the provider-targeted part of the JuPin pinning tool (https://go.fzj.de/pinning). Tests help HPC providers address issues pre-production, update documentation, and notify users of changes. - A user-targeted interactive visualisation functionality of JuPin enables users to test pinning options, visualise task distributions, and generate Slurm-compatible commands. Though focused on Slurm, it can be adapted for other workload managers. - LLview (https://go.fzj.de/llview), an open-source monitoring and operational data analytics tool, has been extended to monitor core usage patterns, providing statistics and aggregated computing times. This helps identify inefficiencies and intervene proactively. JuPin and LLview collectively improve node utilisation, reduce waste, and simplify achieving optimal pinning. These advancements translate to delivering more results in less time. We published JuPin as open-source software on GitHub in May 2025 (https://github.com/FZJ-JSC/jupin). In conclusion, resolving pinning challenges is critical for optimising HPC systems. Our tools establish a foundation for scaling operations, including preparations for the JUPITER exascale supercomputer.
|
The record appears in these collections: |
PDF
Fulltext