sosp-experiments
GitHub用于设计或审计SOSP论文实验评估，确保每个声明都有对应实验支持。指导选择公平基线、混合微基准与端到端测试、诚实报告尾部延迟与开销，并隔离机制影响，保证评估严谨性。
触发场景
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill sosp-experiments -g -y
SKILL.md
Frontmatter
{
"name": "sosp-experiments",
"description": "Use when designing or auditing the evaluation of a SOSP paper — mapping every claim to an experiment, choosing baselines a systems PC will accept as fair, mixing microbenchmarks with end-to-end and failure runs, reporting tails and overheads honestly, and isolating the mechanism the design credits."
}
SOSP Experiments
Use this while the evaluation is being designed — it determines whether the paper is writable at all. SOSP reviews concentrate their negative weight on evaluation soundness, and the venue's response phase forbids new experiments (CFP, 2026 cycle, checked 2026-07-08), so a hole discovered in review stays a hole. The evaluation's job is to make each claim's evidence findable, fair, and attributable to the mechanism the paper credits.
Claims first, experiments second
Write the claim list before designing runs; every evaluation subsection should open by naming the claim it discharges.
| Claim type | Required evidence | Classic hole |
|---|---|---|
| "Foo improves throughput/latency" | End-to-end runs vs tuned state-of-the-art baselines, real + synthetic workloads | Baseline in default config while Foo is hand-tuned |
| "The gain comes from mechanism M" | Ablation: Foo with M disabled, or M grafted onto the baseline | Whole-system win credited to one component on faith |
| "Overhead is acceptable" | Cost measured where it hurts: memory, CPU, write amplification, the workload M does not help | Only the favorable workload reported |
| "Foo survives failures" | Kill/partition/restart runs with recovery-time distributions | Availability claimed, only steady state measured |
| "Foo scales" | Sweep to the knee, with the bottleneck at the knee identified | Straight-line plot ending before saturation |
| "Correctness holds under concurrency" | Stress + targeted schedules, or a verification argument | "We ran it for days" |
Baselines a systems PC accepts
- Compare against the published state of the art, not only your own strawman variants; if the strongest system is unrunnable (dead code, unobtainable hardware), say so in the paper and compare against its published numbers with the configuration deltas spelled out.
- Tune the baseline with the same care as your system, and document both tunings. "We used X's defaults" invites the PC member who built X to explain what the defaults are for.
- Match the fairness surface: same kernel, same mitigations, same NIC firmware,
same warm-up policy. An OS-level artifact changes the platform under everything,
so baseline runs need the environment pinned too (see
sosp-reproducibility).
Workloads: three layers, each with a job
- Microbenchmarks isolate mechanisms and expose costs; they prove why.
- Application-level or end-to-end workloads (real applications on top of the system) prove the why matters; a syscall-path win that vanishes under a real application is a finding, not a failure to report.
- Trace-driven runs on production or public traces prove relevance; when the trace is private, characterize it and pair it with a matched synthetic generator (the tiering that artifact evaluation later formalizes).
Tails, variance, and the numbers reviewers check first
Systems effects live in distributions, and SOSP reviewers reach for the tail:
- Report p50/p99 (p999 where the claim warrants) with repetition counts and spread; a single-run p99 will be called folklore.
- State the load point for every latency number — latency without offered load is meaningless, and latency-throughput curves beat point pairs.
- Account for the denominator: speedups over a baseline that is itself misconfigured are the fastest way to lose the room at the PC meeting.
Evaluation matrix (one row per claim; freeze before the writing sprint)
claim experiment workload/trace baseline+tuning metric+spread figure
throughput at scale e2e sweep to knee YCSB-B + trace T1 X v2.3 tuned §6.1 ops/s, 5 runs, CI Fig 6
gain attributable to M ablation Foo-noM YCSB-B Foo itself delta ops/s Fig 7
recovery independent of kill @ 3 log sizes trace T1 X p50/p99 recov, 10x Fig 9
log size
overhead where M is idle read-only run YCSB-C X CPU%, mem, 5 runs Tab 4
Honesty as a strategy
The evaluation section is also where credibility is banked for the PC meeting. Report the workload where the design loses, and explain the regime boundary — a measured loss with a mechanism story reads as understanding; a suspiciously uniform sweep of wins reads as curation. Reviewers at this venue have written these sections themselves; the fastest way to earn a champion is an evaluation that anticipates the attack they were about to write.
Output format
[Claim map] every paper claim -> experiment row? orphans: <list>
[Baseline audit] SOTA present? tuning documented both sides?
[Layer coverage] micro / end-to-end / trace — gaps
[Tail discipline] spreads + load points on all latency numbers?
[Attribution] ablations isolate the credited mechanism?
[Adverse results] where the design loses + is it in the paper?
版本历史
- 9f86f09 当前 2026-07-19 17:38
