isca-related-work

ISCA Related Work

Architecture is an old field with a long memory. ISCA has run since 1973, its committee members have often worked an area for twenty years, and the fastest way to lose a reviewer is to present a mechanism they saw under another name in a proceedings they attended. Positioning at this venue is therefore archaeology first, differentiation second.

Search across three catalogs, because ISCA lives in two publishers

ISCA's dual ACM/IEEE sponsorship means its proceedings are split across publisher libraries by edition (the 52nd, Tokyo 2025, was ACM-published; the 2026 edition carried IEEE branding — the pattern alternates). Consequences for search:

• dblp (dblp.org/db/conf/isca/) is the reliable spine: complete, per-edition, publisher-neutral. Start every sweep there.
• ACM DL and IEEE Xplore each hold only part of the canon; searching one and concluding "no prior work" is a classic self-inflicted wound.
• Sweep the sibling venues as first-class citizens — MICRO, HPCA, ASPLOS — plus the journals where architecture ideas resurface (TACO, CAL, TC, IEEE Micro including its annual Top Picks selections).
• Finish with a recent-window arXiv pass: concurrent preprints will not sink you, but a reviewer mentioning one you never saw costs credibility in the rebuttal.

The lineage paragraph reviewers actually want

For the two or three mechanisms closest to yours, write lineage, not lists:

Decoupled tracking originates with [x], which [y] extended to multi-level hierarchies at the cost of . Both retain . We remove that assumption, which is what enables <consequence §4>.

The differentiation must be structural — a different trigger, storage organization, or contract — and it must be checkable against your design section. "Unlike prior work, our approach is efficient" differentiates nothing; every cited author believed the same of theirs.

Where comparison belongs

Double-blind mechanics, ISCA 2026 wording

The verified 2026 guidelines are specific and slightly unusual:

• References must not be omitted or anonymized — the bibliography stays intact. Blinding happens in the prose voice, not the reference list.
• Your own prior papers are cited in third person: "extends the tracker of Chen et al. [7]" — never "our prior design [7]". Audit for possessives near citations.
• Links to your artifacts (repos, project pages) must be fully anonymized; a github.com/ URL anywhere in the paper is an identity leak.

# Blind-voice audit over the LaTeX source
grep -nEi 'our (prior|previous|earlier|recent) (work|paper|design|study)' *.tex
grep -nEi 'we (previously|earlier) (showed|proposed|presented)' *.tex
grep -nE  'github\.com|gitlab\.com|bitbucket|\.edu/~[a-z]' *.tex *.bib
# Coverage sanity: in-text citation keys that never resolve
grep -oE '\\cite[tp]?\{[^}]*\}' *.tex | tr ',' '\n' | sort -u | wc -l

Building the map before writing the section

Maintain a positioning table in the repo from week one; the section then writes itself and the rebuttal reuses it:

prior-work-map.tsv
mechanism            venue/year   trigger        storage        differs-from-ours-by
StreamTrack [x]      ISCA'19      miss-driven    per-core CAM   global stall on rollback
LazyMerge  [y]       MICRO'22     epoch-driven   shared SRAM    no cross-level visibility
<ours>               —            demand+decay   per-bank tags  removes rollback entirely

Rows come from the dblp sweep; the last column must name a mechanism-level delta. If you cannot fill that column for a row, that row is your real problem — solve it in the design or the framing, not in adjectives.

Using the unlimited reference pages

ISCA 2026 imposed no limit on reference pages, and architecture papers routinely carry 60-90 citations. Never trim the bibliography for space — trim prose. A thin bibliography in an old field signals a shallow sweep to reviewers who know the depth exists. Depth means: the origin paper of each idea family, the strongest recent representative, and the sibling-venue variants — not twenty near-duplicate cites of one group.

Timing and the November reality

The literature moves between your search and the deadline: MICRO's October program lands weeks before ISCA's November gate, and its papers are fair game for reviewers to mention. Re-run the sweep in the final two weeks, scoped to the newest MICRO/HPCA/ASPLOS programs and recent arXiv, and slot any hit into the lineage paragraphs rather than bolting on a "recent work also..." afterthought.

Section self-check

• dblp sweep done for ISCA + MICRO + HPCA + ASPLOS, last five years minimum, plus the idea family's origin papers regardless of age.
• Every close mechanism: lineage sentence + structural delta + evaluation treatment decided.
• Zero first-person self-citations; zero identifying URLs; references intact.
• Final-fortnight refresh completed against the newest sibling programs.

Positioning output

[Idea family + origin paper] <family, first-named-at venue/year>
[Closest 3 mechanisms] <name, venue/year, structural delta vs ours>
[Evaluation treatment] quantitative baseline: <which> · qualitative: <which>
[Concurrent-work watch] <arXiv/sibling-program hits from the final sweep>
[Blind-voice audit] first-person self-cites: 0 · identifying URLs: 0

Blinding rules and reference policy verified 2026-07-08 against the ISCA 2026 guidelines (../../resources/official-source-map.md); reconfirm for 2027 — 待核实.