SIGMETRICS Experiments
Use this before submission when the evidence is not yet locked. SIGMETRICS reviewers are
performance-evaluation specialists; the evaluation is where a good model or measurement is won or
lost. The organizing principle is evidence proportional to the claim — an analytic claim needs a
proof and validation, a measurement claim needs a methodology a skeptic accepts, and a learning
claim needs a guarantee, not only a benchmark score.
Evaluation audit
- Match evidence to the claim shape. A claim about a bound needs a proof plus a simulation
that shows the analytic curve is right; a claim about a real system needs measured data and a
documented methodology; a claim about a learner needs a regret/convergence guarantee; a claim
about tail behavior needs tail metrics (p99), not means.
- State and test the modeling assumptions. Fit the arrival/service distributions to the real
workload (QQ-plots, goodness-of-fit); a theorem whose M/G/1 assumption is never checked against
the target system invites the "unrealistic model" reject.
- Show analysis-vs-simulation agreement. Overlay the analytic prediction on simulated
measurements; if they disagree, the model or the proof is wrong, and it is better to find that
yourself.
- Use real workloads/traces, sampled or selected by a stated criterion, and describe their
provenance. Toy inputs invite the "does this hold on real systems?" reject.
- Choose fair baselines, including the strongest prior policy/algorithm and a
simple-but-reasonable alternative, tuned with a documented, equal budget. An untuned baseline is a
scored weakness.
- Report statistics for stochastic systems: confidence intervals over independent runs, the
number of runs and the source of variance, warm-up/steady-state handling for simulations, and
effect sizes for comparisons. A single run with no interval is not evidence.
- For learning contributions, report the guarantee (regret/convergence/sample complexity) and
validate it empirically; leaderboard numbers alone route to an ML venue.
Claim-to-evidence design table
| Performance claim |
Matching evidence |
Reject pattern avoided |
| "Our policy bounds the tail" |
Proof under stated assumptions + simulation matching the analytic p99 with CIs |
"A p99 plot with no analytic comparison" |
| "The model captures the system" |
Distribution fit to a real trace + goodness-of-fit |
"Assumed M/G/1, never checked against the workload" |
| "We beat the prior policy" |
Both tuned with equal budget; effect sizes + CIs on real workloads |
"Untuned baseline; toy inputs" |
| "The algorithm has low regret" |
Regret bound (proof) + empirical regret curve vs. the bound |
"Benchmark score with no guarantee" |
| "Scales to realistic load" |
Metrics across realistic arrival rates/sizes with variance reported |
"Only light load tested" |
Validation floor for analytic results
[Proof] full derivation (appendix within the reviewed pages); every case and assumption stated
[Simulation] seeded, steady-state-aware; overlay the analytic prediction; report CIs and #runs
[Agreement] quantify the gap between analysis and simulation; explain any discrepancy
[Robustness] stress an assumption (heavy tails, estimation error) and bound the degradation
Provenance floor for measurement studies
- Record the trace/data source, the collection window, sanitization/anonymization, and access
terms; archive the processed dataset (or document access), not just the collection script.
- State inclusion/exclusion criteria and the resulting sample, with the filtering script in the
artifact.
- Report how outliers, gaps, and measurement artifacts were handled — silent inclusion skews every
downstream metric.
Vignette: evaluating a scheduling policy
Suppose the paper claims a new policy provably reduces p99 latency versus a size-aware optimal. The
matching plan: prove the bound under stated assumptions; fit the service-time distribution to a real
trace and show the fit; simulate both policies with logged seeds and steady-state handling, overlay
the analytic p99 on the simulated p99 with confidence intervals; run a trace-driven evaluation with
both policies tuned equally, reporting effect sizes; and quantify the degradation under fetch-size
estimation error — every number traceable to a logged run in the artifact.
Statistical reporting floor
- Confidence intervals and the number of independent runs for every stochastic measurement.
- Steady-state / warm-up handling stated for simulations.
- The compute and workload scale actually used, not vague feasibility language.
Output format
[Evaluation readiness] strong / adequate / weak
[Claim -> evidence map] <claim: proof? / validation? / measurement? / guarantee?>
[Assumption check] <assumption -> fit to real workload? goodness-of-fit shown?>
[Analysis-vs-measurement] <agreement shown with CIs? discrepancy explained? yes/no>
[Baseline fairness] <baseline -> tuned? equal budget? documented?>
[Decision-critical next run] <one experiment, proof case, or validation to add>
1---2name: sigmetrics-experiments3description: Use when designing or auditing ACM SIGMETRICS evaluations, covering theorem-plus-validation rigor, stating and testing modeling assumptions, analysis-vs-simulation agreement, real workloads and traces, fairly tuned baselines, statistics and confidence intervals for stochastic systems, learning guarantees, measurement provenance, and matching evidence to the shape of each performance claim.4---56# SIGMETRICS Experiments78Use this before submission when the evidence is not yet locked. SIGMETRICS reviewers are9performance-evaluation specialists; the evaluation is where a good model or measurement is won or10lost. The organizing principle is **evidence proportional to the claim** — an analytic claim needs a11proof *and* validation, a measurement claim needs a methodology a skeptic accepts, and a learning12claim needs a guarantee, not only a benchmark score.1314## Evaluation audit1516- **Match evidence to the claim shape.** A claim about a **bound** needs a proof plus a simulation17 that shows the analytic curve is right; a claim about a **real system** needs measured data and a18 documented methodology; a claim about a **learner** needs a regret/convergence guarantee; a claim19 about **tail** behavior needs tail metrics (p99), not means.20- **State and test the modeling assumptions.** Fit the arrival/service distributions to the real21 workload (QQ-plots, goodness-of-fit); a theorem whose M/G/1 assumption is never checked against22 the target system invites the "unrealistic model" reject.23- **Show analysis-vs-simulation agreement.** Overlay the analytic prediction on simulated24 measurements; if they disagree, the model or the proof is wrong, and it is better to find that25 yourself.26- **Use real workloads/traces,** sampled or selected by a stated criterion, and describe their27 provenance. Toy inputs invite the "does this hold on real systems?" reject.28- **Choose fair baselines,** including the strongest prior policy/algorithm and a29 simple-but-reasonable alternative, tuned with a documented, equal budget. An untuned baseline is a30 scored weakness.31- **Report statistics for stochastic systems:** confidence intervals over independent runs, the32 number of runs and the source of variance, warm-up/steady-state handling for simulations, and33 effect sizes for comparisons. A single run with no interval is not evidence.34- **For learning contributions,** report the guarantee (regret/convergence/sample complexity) and35 validate it empirically; leaderboard numbers alone route to an ML venue.3637## Claim-to-evidence design table3839| Performance claim | Matching evidence | Reject pattern avoided |40|---|---|---|41| "Our policy bounds the tail" | Proof under stated assumptions + simulation matching the analytic p99 with CIs | "A p99 plot with no analytic comparison" |42| "The model captures the system" | Distribution fit to a real trace + goodness-of-fit | "Assumed M/G/1, never checked against the workload" |43| "We beat the prior policy" | Both tuned with equal budget; effect sizes + CIs on real workloads | "Untuned baseline; toy inputs" |44| "The algorithm has low regret" | Regret bound (proof) + empirical regret curve vs. the bound | "Benchmark score with no guarantee" |45| "Scales to realistic load" | Metrics across realistic arrival rates/sizes with variance reported | "Only light load tested" |4647## Validation floor for analytic results4849```text50[Proof] full derivation (appendix within the reviewed pages); every case and assumption stated51[Simulation] seeded, steady-state-aware; overlay the analytic prediction; report CIs and #runs52[Agreement] quantify the gap between analysis and simulation; explain any discrepancy53[Robustness] stress an assumption (heavy tails, estimation error) and bound the degradation54```5556## Provenance floor for measurement studies5758- Record the trace/data source, the collection window, sanitization/anonymization, and access59 terms; archive the *processed* dataset (or document access), not just the collection script.60- State inclusion/exclusion criteria and the resulting sample, with the filtering script in the61 artifact.62- Report how outliers, gaps, and measurement artifacts were handled — silent inclusion skews every63 downstream metric.6465## Vignette: evaluating a scheduling policy6667Suppose the paper claims a new policy provably reduces p99 latency versus a size-aware optimal. The68matching plan: prove the bound under stated assumptions; fit the service-time distribution to a real69trace and show the fit; simulate both policies with logged seeds and steady-state handling, overlay70the analytic p99 on the simulated p99 with confidence intervals; run a trace-driven evaluation with71both policies tuned equally, reporting effect sizes; and quantify the degradation under fetch-size72estimation error — every number traceable to a logged run in the artifact.7374## Statistical reporting floor7576- Confidence intervals and the number of independent runs for every stochastic measurement.77- Steady-state / warm-up handling stated for simulations.78- The compute and workload scale actually used, not vague feasibility language.7980## Output format8182```text83[Evaluation readiness] strong / adequate / weak84[Claim -> evidence map] <claim: proof? / validation? / measurement? / guarantee?>85[Assumption check] <assumption -> fit to real workload? goodness-of-fit shown?>86[Analysis-vs-measurement] <agreement shown with CIs? discrepancy explained? yes/no>87[Baseline fairness] <baseline -> tuned? equal budget? documented?>88[Decision-critical next run] <one experiment, proof case, or validation to add>89```
brycewang-stanford