nsdi-supplementary

NSDI Supplementary Material

NSDI gives authors four containers with different review status. Misfiling content — a load-bearing baseline in an appendix, revision evidence missing from the auxiliary packet — costs more than any phrasing error. Rules below are the NSDI '27 CFP as rendered on 2026-07-08.

The four containers

Two contrasts worth internalizing. Against sibling venues that ban submission-time appendices outright, NSDI permits them — the trap here is overuse, not existence. And the auxiliary-material slot is not a junk drawer: the CFP frames it for content critical to the paper that lives online, plus the structured revision documents below.

The placement rule

Ask of every artifact of content: would a reviewer who skipped this be unable to judge a claim?

• Yes → body. Key workload description, baseline identity and tuning, the mechanism a headline number depends on, the failure regimes. If it does not fit, the paper is over-claiming for its budget — cut claims, not their evidence (nsdi-writing-style).
• Supports but does not carry → appendix. Full parameter tables, additional workload results consistent with in-body trends, protocol message formats, extended proofs of secondary properties.
• Lives online and matters → auxiliary material, anonymized. A demo, a dataset sample, an interactive visualization. Never a raw GitHub link in the text — repo URLs identify authors (nsdi-submission).

Self-containment is a hard constraint in both directions: the 12 pages must stand alone if every appendix page is ignored, and no appendix may quietly introduce a new claim the body never made.

Appendix design, when used

• Order appendices by likely reviewer interest; a skeptical reviewer goes to the appendix to check one specific thing, not to read linearly.
• Give each appendix a one-line pointer from the body ("full per-workload results: App. B") so their existence is discoverable inside the reviewed pages.
• Regenerate appendix tables from the same run ledger as body figures (nsdi-reproducibility); internal contradictions between body and appendix are a reviewer's fastest route to distrust.
• Anonymize appendices to the same standard as the body — leaks hide in config listings, hostnames, and dataset paths.

The one-shot revision packet

NSDI's most distinctive supplementary requirement. A resubmitted one-shot revision must include auxiliary material containing:

1. a version of the paper highlighting all changes made, and
2. a high-level explanation of the major changes.

Build both as living documents during the revision, not as a deadline-night diff:

# Maintain the highlighted version continuously with latexdiff
latexdiff submitted/main.tex revised/main.tex > diff/main-diff.tex
pdflatex -output-directory diff diff/main-diff.tex
# Change memo skeleton: one block per required issue
#   R2.3 (baseline tuning): re-ran §6.2 with vendor-recommended config;
#   deltas now 2.1-2.9x (was 2.1-3.4x); text + Fig 7 updated. [diff pp. 8-9]

The same reviewers re-read the packet against their issue list; a memo that maps issue → change → diff location does their verification for them. A revision whose packet obscures what changed — or that ignores a required issue — is rejected with no further revision option (nsdi-author-response).

Triage drill for an overfull draft

When the draft runs to 13.5 pages, resist moving the overflow wholesale to an appendix — demote by evidence role:

1. List every figure/table with the claim it supports; anything supporting no in-body claim leaves the paper entirely.
2. Secondary workloads whose trend matches the primary: one summarizing sentence in the body + full table in the appendix.
3. Mechanism details needed to reimplement but not to judge (message formats, tuning constants): appendix.
4. Anything a skeptical reviewer needs to believe a number: stays in-body, and a claim gets cut instead if space still lacks.

Pre-upload sweep

• Body stands alone with all appendices deleted.
• No load-bearing evidence outside the 12 pages.
• Every appendix pointed to from the body; no orphan claims.
• Auxiliary uploads anonymized (metadata, paths, owners) and actually critical.
• Revision packets: highlighted PDF + change memo present, issue-complete.
• Combined upload sizes within whatever HotCRP enforces this cycle (check the form; caps 待核实).

Cross-container consistency

The containers are read at different times by different people — reviewers now, revision reviewers months later, artifact evaluators after acceptance — so consistency is a supply-chain property: numbers quoted in the body match the appendix tables they summarize; the auxiliary demo shows the same system version the paper measures; and the revision packet's diff derives from the exact PDF the reviewers previously read, not an intermediate draft.

Output format

[Placement audit] item -> body / references / appendix / auxiliary (+ moves needed)
[Self-containment] body survives appendix deletion? orphan claims?
[Appendix hygiene] pointers present / regenerated from ledger / anonymized
[Revision packet] highlighted PDF ok? memo maps all required issues?
[Blocking issues] <ordered before upload>