Apply a uniform power cap across many cluster nodes using Ansible
Reference BIOS tweaks for a student supercomputing competition cluster
Disable NUMA balancing across nodes for HPC workloads
Requires a real multi-node cluster with BIOS access and power monitoring; README is Chinese and config is trimmed for reference only.
This repository is a small collection of server configuration scripts that the authors used in a student supercomputing competition called ASC26. The team is from the University of Electronic Science and Technology of China, abbreviated UESTC in the project name. The competition gives each team a fixed power budget for the cluster, so a lot of the preparation work is about tuning the machines to stay under that limit while still running benchmarks as fast as possible. The README lists four folders. One holds basic BIOS settings, one holds the actual power-limit configuration, one is for running stress tests, and a fourth turns off a Linux feature called NUMA balancing that the author says was not useful in their case. NUMA balancing is a kernel behaviour that moves memory around between processor sockets, and the team chose to disable it. The configuration is delivered through Ansible. Ansible is described in the README as an automation tool that connects to many servers over SSH and runs the same setup steps on each of them. The author notes that they were rushed during the competition and let an AI assistant generate the Ansible playbooks on the spot, but the result was good enough to push power-limit settings out to the whole cluster. The author also clarifies that a 4990 watt stable limit mentioned in the files was not a planned target. It was discovered by accident during a test run, where the cluster happened to stay stable at that level without tripping. The repository is a trimmed-down version of their working setup, kept for reference only. The README is written in Chinese.
Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.