SIGCOMM Reproducibility
Use this before submission and again before the artifact deadline. SIGCOMM's culture treats a
result as a claim a stranger should be able to rebuild; reproducibility here is largely about
whether the network conditions behind a number are recorded well enough to recreate.
Reopen the current Call for Artifacts to confirm the process for this edition.
Evidence map
- Map each reported number to a rebuildable location: a script, a logged run, a config, and
the figure it feeds.
- For every measurement, record the topology (physical or emulated), the traffic
(workload generator or trace, with provenance), the configuration (every parameter),
and the environment (kernel, NIC, switch firmware, data-plane toolchain).
- For any stochastic result, report the run count, the seed or workload driver, and the
variance behind each percentile — a lone 99th-percentile number with no replication is not
reproducible.
- Decide early what can legally ship: production traces and topologies often cannot, so plan a
substitute with matching statistical character rather than discovering the block at the
deadline.
- Keep the paper and the artifact consistent; a figure the artifact cannot regenerate is a
reproducibility gap reviewers and the AEC both notice.
The four ledgers
| Ledger |
What it records |
Failure it prevents |
| Topology |
Nodes, links, capacities, buffer sizes, emulation vs. hardware |
"Which topology produced Figure 6?" |
| Traffic |
Workload distribution or trace, provenance, which figure used it |
Numbers that cannot be tied to an input |
| Configuration |
Every parameter and its value per experiment |
Silent knob changes between runs |
| Run |
Seeds, replication counts, timestamps, variance |
Percentiles with no reproducible basis |
Maintaining these while you run is cheap; reconstructing them after the deadline is
error-prone and often impossible.
Degrees of reproducibility
Turnkey : one command rebuilds each figure from logged runs on a documented setup
Scripted : scripts exist but need hardware, private data, or documented manual steps
Descriptive : prose detailed enough that a competent networker could rebuild the pipeline
Fragile : results depend on unrecorded conditions -> a red flag to fix before submission
For SIGCOMM, downscaled testbed results should be turnkey because evaluators actually run
them; full-scale hardware results may stay scripted with the deviation documented. State the
achieved level honestly rather than promising turnkey behavior that fails on a clean machine.
Vignette: a wide-area measurement result
A paper characterizes loss and latency on a production overlay, then proposes a routing tweak.
Its reproducibility spine: the measurement window and vantage points, the sampling method, the
anonymization applied before any release, a public-trace or synthetic substitute for the parts
that cannot ship, and a driver that regenerates the loss-latency figures from logged samples —
plus one honest sentence about which production condition the substitute cannot capture.
Legal and ethical release
- Real user or third-party-infrastructure measurement needs its handling addressed in the
paper; anonymize before release and record what was removed.
- Prefer an archival host with a DOI for anything you can ship, and document the exact reason
for anything you cannot.
Output format
[Claim inventory] <number -> rebuildable location>
[Ledger status] topology / traffic / configuration / run — complete or gaps
[Reproducibility level] turnkey / scripted / descriptive / fragile
[Data-release plan] shippable / substitute (+ provenance) / withheld (+ reason)
[Paper fixes] <must appear in the main PDF>
[Artifact fixes] <what the package still needs>
1---2name: sigcomm-reproducibility3description: Use when strengthening the reproducibility evidence of an ACM SIGCOMM paper — topology and testbed ledgers, traffic workload and trace provenance, configuration and version pinning, tail-percentile run counts and variance, legal data-release decisions, and consistency between the paper's claims and the artifact that backs them.4metadata:5 source: https://skillmd.com/skills/brycewang-stanford/sigcomm-reproducibility6---78# SIGCOMM Reproducibility910Use this before submission and again before the artifact deadline. SIGCOMM's culture treats a11result as a claim a stranger should be able to rebuild; reproducibility here is largely about12whether the **network conditions** behind a number are recorded well enough to recreate.13Reopen the current Call for Artifacts to confirm the process for this edition.1415## Evidence map1617- Map each reported number to a rebuildable location: a script, a logged run, a config, and18 the figure it feeds.19- For every measurement, record the **topology** (physical or emulated), the **traffic**20 (workload generator or trace, with provenance), the **configuration** (every parameter),21 and the **environment** (kernel, NIC, switch firmware, data-plane toolchain).22- For any stochastic result, report the **run count**, the seed or workload driver, and the23 variance behind each percentile — a lone 99th-percentile number with no replication is not24 reproducible.25- Decide early what can legally ship: production traces and topologies often cannot, so plan a26 substitute with matching statistical character rather than discovering the block at the27 deadline.28- Keep the paper and the artifact consistent; a figure the artifact cannot regenerate is a29 reproducibility gap reviewers and the AEC both notice.3031## The four ledgers3233| Ledger | What it records | Failure it prevents |34|---|---|---|35| Topology | Nodes, links, capacities, buffer sizes, emulation vs. hardware | "Which topology produced Figure 6?" |36| Traffic | Workload distribution or trace, provenance, which figure used it | Numbers that cannot be tied to an input |37| Configuration | Every parameter and its value per experiment | Silent knob changes between runs |38| Run | Seeds, replication counts, timestamps, variance | Percentiles with no reproducible basis |3940Maintaining these while you run is cheap; reconstructing them after the deadline is41error-prone and often impossible.4243## Degrees of reproducibility4445```text46Turnkey : one command rebuilds each figure from logged runs on a documented setup47Scripted : scripts exist but need hardware, private data, or documented manual steps48Descriptive : prose detailed enough that a competent networker could rebuild the pipeline49Fragile : results depend on unrecorded conditions -> a red flag to fix before submission50```5152For SIGCOMM, downscaled testbed results should be turnkey because evaluators actually run53them; full-scale hardware results may stay scripted with the deviation documented. State the54achieved level honestly rather than promising turnkey behavior that fails on a clean machine.5556## Vignette: a wide-area measurement result5758A paper characterizes loss and latency on a production overlay, then proposes a routing tweak.59Its reproducibility spine: the measurement window and vantage points, the sampling method, the60anonymization applied before any release, a public-trace or synthetic substitute for the parts61that cannot ship, and a driver that regenerates the loss-latency figures from logged samples —62plus one honest sentence about which production condition the substitute cannot capture.6364## Legal and ethical release6566- Real user or third-party-infrastructure measurement needs its handling addressed in the67 paper; anonymize before release and record what was removed.68- Prefer an archival host with a DOI for anything you can ship, and document the exact reason69 for anything you cannot.7071## Output format7273```text74[Claim inventory] <number -> rebuildable location>75[Ledger status] topology / traffic / configuration / run — complete or gaps76[Reproducibility level] turnkey / scripted / descriptive / fragile77[Data-release plan] shippable / substitute (+ provenance) / withheld (+ reason)78[Paper fixes] <must appear in the main PDF>79[Artifact fixes] <what the package still needs>80```