Skills Engineering Ensuring Scientific Reproducibility in Systems Papers

Ensuring Scientific Reproducibility in Systems Papers

v20260724
socc-reproducibility
This comprehensive guide outlines best practices for achieving high levels of scientific reproducibility in systems papers, especially for cloud measurement studies. It teaches authors how to map every claim in their paper to verifiable evidence, detailing the entire testbed setup, reproducing tail latency and costs accurately, and maintaining consistency between reported numbers and artifacts.
Get Skill
58 downloads
Overview

SoCC Reproducibility

Use this before submission and again before camera-ready. SoCC — the joint SIGMOD+SIGOPS cloud symposium — is read by reviewers who expect an inspectable measurement trail: the SIGOPS half wants to believe the system runs, and the SIGMOD half wants to believe the numbers. The goal is that a competent reader with a comparable testbed could rebuild your evidence and reach your conclusions — including the tail latency and cost, not just the average.

Evidence map

  • Map each cloud claim and reported number to a verifiable location — a paper section, a figure generated from logged runs, or a script in the artifact.
  • For systems, give enough of the mechanism, configuration, and testbed (node counts, instance types, OS/kernel, network) that a reader could re-deploy and re-measure.
  • For measurement/trace studies, report the trace provenance, extraction date, filtering, the workload replay, metrics, and the analysis scripts.
  • Reproduce tail and cost: percentiles (p95/p99/p99.9), the pricing model behind any cost claim, and the number of runs and variance — the mean alone is not a cloud result.
  • Keep the paper and the artifact consistent: a number in the PDF that no script regenerates is the contradiction reviewers read as carelessness.

Reproducibility statement audit

Claim in the paper Weak answer SoCC-ready answer
"We evaluate on a production trace" "Trace available on request" Anonymized (then released) trace + the replay harness and extraction date
"Our system improves throughput" "Code will be released" Anonymized, runnable system with a testbed description and a small demo
"We cut cost by X" A single cost number The pricing model, the instance-seconds logged, and the script that computes it
"p99 stays within target" Mean latency only Per-run tail percentiles with variance and run count
"Scales to N nodes" One large run A scaled reproduction path plus the full-scale logs

"Available on request" is treated as not available; convert every such line into a concrete, anonymized (then released) artifact or an explicit, justified exception (e.g., a confidential production trace, with a synthetic generator provided instead).

Provenance pinning

[Measurement] pin commit SHAs; record trace extraction dates; archive the replayed trace or a
              faithful generator, not just a query or a pointer
[Testbed]     record node counts, instance types, OS/kernel versions, network, and the run count;
              a cloud result that cannot be re-deployed cannot be reproduced
[Cost]        state the pricing model and the source of every cost figure so a reader can recompute
[Tail]        log per-request or per-run latency distributions, not only aggregates
[Randomness]  log seeds for stochastic components; say what is and is not deterministic

Degrees of reproducibility (state the one you achieved)

  • Turnkey: one documented command regenerates each figure (including tail and cost) from logged runs, at least at reduced scale.
  • Scripted: scripts exist but require a specific testbed, documented manual steps, or restricted trace access.
  • Descriptive: prose and configuration detailed enough that a competent reader could rebuild the deployment and measurement pipeline.

For SoCC, aim turnkey for anything an evaluator could rerun at small scale (a short trace replay, a tail/cost plot from logged runs); full-cluster or proprietary-trace results may stay scripted with access clearly documented. Stating the achieved level honestly beats promising turnkey behavior that fails on someone else's testbed.

Vignette: a scheduling measurement paper

Consider a paper measuring a new scheduler on a replayed production trace. Its reproducibility spine: the scheduler code with pinned SHAs; the replay harness and the (anonymized, then released) trace with its extraction date; the testbed description (nodes, instance types, kernel); the measurement scripts that turn raw logs into the throughput, p99, and cost figures; the run count and variance; and one honest sentence about the parts (a confidential production trace, the full cluster) that cannot be shared and what synthetic or scaled substitute is provided.

Consistency and camera-ready pass

  • Before submission: every reported number traces to the artifact; the reproducibility statement matches reality; the artifact is anonymized (no cluster names, provider hints, or trace provenance that reveals identity).
  • Before camera-ready: swap anonymized links for permanent, DOI-issuing archives, and align the statement with the ACM badges you are pursuing if the edition offers evaluation (socc-artifact-evaluation).

Output format

[Claim inventory] <claim -> evidence location>
[Reproducibility statement] concrete / vague / missing
[Provenance gaps] <trace SHAs+dates / testbed description / cost model / tail logging / seeds>
[Tail + cost] reproducible, not just the mean? yes/no
[Reproducibility level] turnkey / scripted / descriptive, stated honestly
[Paper fixes] <must appear in the PDF>
[Artifact fixes] <additions before upload>
Info
Category Engineering
Name socc-reproducibility
Version v20260724
Size 5.52KB
Updated At 2026-07-29
Language