arXiv: SysPilot tunes performance across libraries using reusable resource contracts

arXiv: SysPilot tunes performance across libraries using reusable resource contracts

Yineng Yan, Haowei Chen and Christopher J. Rossbach present SysPilot, a system for tuning applications that combine independently developed libraries. They argue that libraries can compete for shared CPU cores, GPUs and memory bandwidth, while existing autotuners do not directly examine which resources are shared or which runtime signals indicate interference. SysPilot uses a Library Performance Contract (LPC) to capture a library’s resource demands, assumptions, controls and observability signals. It synthesizes contracts from documentation and code, combines contracts for an application’s libraries, forms resource-conflict hypotheses, profiles the application to test them and selects candidate configurations. The authors’ central idea is reusable performance knowledge: a library’s contract can serve across applications that depend on it. On a benchmark suite of cross-library contention scenarios, the authors report performance matching autotuners that use application-specific manual information. They also report up to an additional 38% improvement in benchmark performance metrics over the best configurations found by OpenTuner and SMAC3, using 3x fewer runs. These are reported benchmark results; the abstract does not establish how they carry over to other workloads.

Publication: e5f8ac31-613c-4dbc-8680-4adb8b3c556b

attribution — pass: The reading names the authors listed on the arXiv page. It frames the problem, design and results as the authors' claims and reported findings, not as established fact.

Authors:Yineng Yan, Haowei Chen, Christopher J. Rossbach

summary — pass: The reading leads with the actor and the development. It explains the problem in plain terms: libraries competing for CPU cores, GPUs and memory bandwidth. It defines a Library Performance Contract by what it captures and outlines the pipeline. It also conveys the authors' central idea of reusable performance knowledge. The title and question are accurate and non-leading. Autotuner is not formally defined, but context makes its meaning clear enough.

This design makes the Library Performance Contract a reusable unit of performance knowledge: a contract for a library can be reused across many applications that depend on that library.

date — pass: The originalPublished date of 2026-11-04 matches the arXiv v1 submission date. The later feed date does not affect this.

[Submitted on 4 Nov 2026]

support — pass: Every claim matches the abstract. This includes the contention problem, the contract-based pipeline, parity with manual-information autotuners, and the up-to-38% gain with 3x fewer runs versus OpenTuner and SMAC3. The results are kept to the benchmark suite and the up-to qualifier is preserved. The limitation about generalization is appropriately hedged.

SysPilot improves benchmark performance metrics by up to an additional 38% over the best configuration found by these autotuners, while using 3x fewer runs.

media — pass: The selected media is the verified direct link to the arXiv abstract page, which is the original source. No better source media was found.