NSDI Writing Style

An NSDI paper argues that a nameable design idea, embodied in a working networked system, holds up under realistic traffic. The prose exists to make that chain checkable. Format facts below are the NSDI '27 CFP as rendered on 2026-07-08.

The opening arc

NSDI first pages that work tend to execute five beats in order (see the rebuilt example in ../../resources/worked-examples/01-introduction.md):

1. Operational pain, evidenced — a failure shape, cost, or limit shown in a trace or deployment, not asserted from intuition.
2. Why the networked setting causes it — the distributed or stack-level reason local fixes fail. This beat is also the scope defense; a paper that cannot write it is drifting out of the CFP's "design of networked systems or the networking stack."
3. The design principle, named — the transferable idea, stated so another team could apply it to a different system.
4. The built system — what was implemented, where it runs, what was not built.
5. Claims with evidence addresses — each headline number tied to a section, a workload, and a percentile.

Budget arithmetic before sentence polish

The '27 rule: 12 pages including footnotes, figures, and tables; references and appendices may extend beyond. Two consequences shape the writing:

• Figures are inside the cap, so every plot pays rent. Cut the third bar chart saying "ours is faster again"; keep the breakdown that explains why.
• Appendices exist but are overflow, not annex: reviewers judge the 12 pages. Nothing load-bearing (a baseline, the key workload description, a correctness argument) may live only past the references (nsdi-supplementary).

A workable allocation for a 12-page systems paper:

Sentence-level calibration

Systems reviewers at this venue distrust superlatives and read distributions. Editing passes that consistently improve NSDI drafts:

• Replace every "up to N×" with the percentile, workload, and range: "2.1-3.4× at p99.9 on the replayed incident weeks."
• Replace "significantly / dramatically / novel" with the measured delta or the named mechanism. Adjectives signal missing numbers.
• Convert passive mechanism descriptions into agent-explicit ones — which component decides, on what signal, at what timescale.
• State the boundary honestly, in the body: the regime where the design loses (incast, single replica, cold caches). NSDI culture treats a named limitation as competence, and an unnamed one found by a reviewer as concealment.
• Keep terminology bijective: one name per component, one component per name, figures and text agreeing exactly.

Weak:  "Our novel scheduler significantly reduces latency (up to 4.1x)
        compared to state-of-the-art approaches."
NSDI:  "Shifting retries off congested paths cuts p99.9 latency 2.1-3.4x
        vs tuned exponential backoff across the three incident classes
        (§6.2); gains vanish when RTT < the lease renewal interval (§7)."

Anonymity constraints on prose

Double-blind shapes phrasing, with one NSDI-specific fork: research-track papers neutralize identifying deployment details ("a large provider's fleet"), while operational-track papers may keep real system and company names with author names still withheld. Write self-citations in the third person; never delete a reference to hide identity; cite a concurrent NSDI submission as "Under submission. Details omitted for double-blind reviewing." The CFP endorses AI tools for grammar and clarity of human-written text — the ideas, and the responsibility, stay human.

Revision-letter prose (one-shot resubmissions)

A resubmitted revision is read by the same reviewers against their own issue list, so the writing task inverts: instead of impressing a stranger, show a known reader that each demanded change exists. Keep new text visually traceable — the auxiliary change-highlighted PDF must let a reviewer jump from issue to fix in seconds (nsdi-author-response). Prose that quietly rewrites unrelated sections invites re-litigation of settled points.

Figures as argument, not decoration

Because figures bill against the 12 pages, each one should be designed like a paragraph with a topic sentence:

• Caption states the takeaway ("Leases cap retry amplification at 1.3× during all three incident classes"), not the axes ("Latency vs. time").
• One visual idea per figure; a plot needing three sentences of legend explanation is two plots or zero.
• CDFs for distribution claims, timelines for dynamic behavior, breakdowns for "why" claims — the form should match the claim type (nsdi-experiments).
• Readability at print scale: systems reviewers annotate paper printouts; 6-point axis labels get skipped, and a skipped figure is a page of rent paid for nothing.

Self-edit checklist

• First page executes all five beats; a stranger can quote the design principle.
• Every claim has an evidence address (section + workload + metric).
• No superlative survives without a number; tails reported where tails matter.
• Failure regimes named in the body, not confessed in the appendix.
• 12-page count verified on the built PDF, figures included.
• Anonymization style matches the declared track.

Output format

[Arc check] beats 1-5 present? weakest beat + fix
[Budget] pages by section vs allocation; what to cut first
[Calibration] list of unbacked superlatives / missing percentiles
[Boundary] named failure regimes: <list or MISSING>
[Track-consistent anonymity] research-neutralized / operational-named
[Edit order] highest-leverage rewrite first