thedixitjain NSDI Related Work
Related work at NSDI answers one reviewer question: what does this design let us do that the systems we already have do not? Coverage gaps read as either ignorance or evasion, and the venue's open-access record removes every excuse for both.
The lanes an NSDI reviewer expects swept
| Lane | Where it lives | What must be said about it |
|---|---|---|
| Direct competitors | NSDI, SIGCOMM, OSDI, SOSP | Mechanism-level delta, ideally measured head-to-head |
| Same problem, other stack layer | SIGCOMM (protocol), OSDI/SOSP (host), FAST (storage) | Why this layer is the right place to solve it |
| Same technique, other domain | broad systems + ML-systems venues | What the networked setting breaks or changes |
| Measurement/characterization basis | IMC, NSDI measurement papers | Which observed phenomena motivated the design |
| Deployed/industrial precedent | operational-track papers, provider tech reports | What production experience the design inherits or contradicts |
| Early/idea-stage neighbors | HotNets, frontiers-track papers | That the idea's grown-up form differs from the sketch |
Both cohorts of the current edition count as literature: NSDI posts spring
accepted-paper lists months before the symposium (e.g.,
usenix.org/conference/nsdi26/spring-accepted-papers), so "it wasn't published yet"
fails for a fall submission when the spring list was public at deadline time. Sweep
the sibling deadlines too — SIGCOMM's accepted list frequently contains the closest
neighbor of an NSDI submission.
Verification: the venue must be proven
Systems folklore misfiles its classics; citing Raft "at NSDI" or B4 "from NSDI"
signals secondhand reading (Raft is USENIX ATC 2014, B4 is SIGCOMM 2013 — see the
guard list in
../../resources/exemplars/library.md). Every
NSDI paper since the venue began is a free PDF on usenix.org, and dblp's
conf/nsdi stream gives the authoritative edition/pages. Verify before the
bibliography freezes:
# Cross-check claimed venues in the .bib against dblp keys / USENIX URLs
grep -E '^@' paper.bib | wc -l # inventory
grep -A3 -iE 'booktitle.*(nsdi|sigcomm|osdi|sosp)' paper.bib | head -40
# For each NSDI claim: the usenix.org/conference/nsdi<yy>/presentation/<key>
# page or the dblp record must exist. No render, no citation-as-NSDI.
Read the PDF of any paper you compare against quantitatively. Comparing against a system's abstract is how "prior work cannot handle churn" ends up in front of the reviewer who wrote the churn section.
Positioning moves that work at this venue
- Delta by mechanism, not by adjective. "X batches at the NIC; we batch at the RPC layer, which preserves per-request deadlines" beats "unlike X, our approach is more flexible."
- Concede the overlap, then locate the divergence. NSDI reviewers usually know the neighbor papers personally; a generous one-clause summary of the competitor buys credibility for the distinction that follows.
- Use operational precedent as evidence, not decoration. If a provider's operational-track paper reports the failure mode your design removes, that citation is motivation-grade, not related-work filler.
- Position against the measurement literature. A design justified by workload properties should cite where those properties were measured — or measure them in §2.
Double-blind mechanics for citations
- Self-citations stay, in third person. "We extend Ballast [12]" reveals; "This work extends Ballast [12]" written as if by strangers does not. Omitting your own prior paper to hide identity is explicitly against the CFP.
- Concurrent NSDI submissions are cited with the CFP's fixed sentence: "Under submission. Details omitted for double-blind reviewing."
- Tech reports and arXiv versions of your own work: cite in third person like any prior art; do not link author-revealing artifacts in the text.
- Operational-track nuance: the anonymization exception covers your paper's system and company names — it does not license identifying the author list through possessive phrasing.
Sequencing the sweep across the two-deadline year
Because NSDI publishes cohort lists mid-cycle, the sweep has a rhythm:
- At project start: lanes 1-3 in depth; identify the two or three systems the
paper must beat or distinguish, since they shape the evaluation
(
nsdi-experiments). - Four weeks before the gate: re-sweep the current edition's earlier cohort list and the most recent SIGCOMM/OSDI/SOSP programs; new neighbors get a mechanism-delta sentence, not a panicked rewrite.
- At bibliography freeze: run the venue-verification pass below; fix misattributions before reviewers find them.
- In a one-shot revision: re-check whether the reviewers' cited omissions have
themselves been published since — the same reviewers will notice both presence
and absence (
nsdi-author-response).
Failure patterns
- The wall of one-line citations ("Much work has studied datacenter transport [3-19]") — coverage theater with no deltas.
- Comparing only against academic prototypes when a widely deployed system solves an adjacent problem — reviewers from industry will supply the omission.
- Treating a spring-cohort paper of the same edition as invisible.
- Venue misattribution in the bibliography — cheap to fix, expensive in credibility.
- A related-work section that never mentions the paper's own limits relative to prior work; positioning is bidirectional.
Where the section lives
NSDI drafts split positioning across two homes, both inside the 12 pages: a short contrast in the introduction (the two nearest neighbors, one delta sentence each) and the fuller section — conventionally late, before or after discussion — carrying the lane-by-lane sweep. Whichever home a comparison occupies, the quantitative version belongs in the evaluation: a related-work sentence claiming superiority that §6 never measures is a self-inflicted reviewer question.
Output format
[Lane coverage] table lane -> covered / gap (paper to add)
[Nearest neighbor] <paper, venue-verified> + mechanism delta
[Cohort sweep] current-edition spring list checked? sibling deadlines?
[Verification] # citations venue-proven / total NSDI-venue claims
[Blind check] self-cites third-person? concurrent-submission wording?
[Fix list] ordered corrections before bibliography freeze
Source: brycewang-stanford/Awesome-Journal-Skills → NSDI-Skills/skills/nsdi-related-work/SKILL.md