asplos-topic-selection
GitHub指导论文投稿ASPLOS还是其他单一社区会议。通过删除测试判断软硬件耦合度，利用路由表匹配ISCA、SOSP等目标 venue，并确认2027年截稿日期及跨层研究主题。
触发场景
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill asplos-topic-selection -g -y
SKILL.md
Frontmatter
{
"name": "asplos-topic-selection",
"description": "Use when deciding whether a project belongs at ASPLOS or at a single-community venue — applying the cross-layer deletion test, matching against ASPLOS's architecture\/PL\/OS intersection identity, routing to ISCA\/MICRO\/HPCA, PLDI\/POPL, SOSP\/OSDI\/EuroSys, SC, or MLSys instead, and choosing between the April and September deadlines."
}
ASPLOS Topic Selection
ASPLOS's self-description is unusually operational: it seeks work that improves the core disciplines — operating systems, programming languages, computer architecture — and their intersection, and its 2027 rapid-review round explicitly prioritizes submissions at that intersection (CFP, checked 2026-07-08). Venue fit is therefore not a prestige question here; it is a screening variable. A technically strong paper that lives cleanly inside one community can be filtered in the first round not because it is bad but because a single-community venue owns its reviewers.
The deletion test
Ask, for each layer the paper touches: if this layer's content were deleted, would a publishable paper remain?
- Delete the software half and a full paper remains → the work is an ISCA/MICRO/HPCA candidate wearing an ASPLOS template.
- Delete the hardware half and a full paper remains → it is a SOSP/OSDI/EuroSys or PLDI paper.
- Delete either half and nothing defensible remains → the coupling is real; ASPLOS is a first-choice target.
The exemplars library shows what passing looks like in award-validated form: an
accelerator argued from data movement, a scheduler premised on hardware
heterogeneity, an interrupt-delivery boundary redrawn between hypervisor and silicon
(see ../../resources/exemplars/library.md).
Routing table
| If the core claim is... | Stronger first target | ASPLOS becomes right when... |
|---|---|---|
| A microarchitectural structure (predictor, cache policy, pipeline) | ISCA / MICRO / HPCA | The structure exists to serve a named software behavior, and software changes with it |
| A language design, type system, or compiler optimization | PLDI / POPL / CGO / OOPSLA | The compiler co-evolves with an ISA, accelerator, or OS interface |
| An OS/distributed-systems design on commodity hardware | SOSP / OSDI / EuroSys / ATC | Hardware assumptions are load-bearing: new memory technology, custom hints, near-data compute |
| HPC-scale performance engineering | SC / PPoPP | The result generalizes into an architectural or systems principle beyond one machine |
| ML training/serving systems on stock stacks | MLSys | The ML workload drives a hardware/software interface change, not just better scheduling |
| Security via hardware mechanisms | IEEE S&P / USENIX Security / CCS | The defense requires co-designed architecture + systems support and its evaluation is performance-shaped |
Scope confirmation: the 2027 topics list spans heterogeneous architectures and accelerators; cloud and datacenters; memory, storage, networking, and I/O; power, energy, and thermal management; profiling, debugging, and testing; security, reliability, and availability; parallelism and big-data systems; virtualization; and experimental methodologies — with non-traditional topics welcome when centered in the core disciplines.
Fit interrogation script
Run this before any writing begins:
Q1 Name the interface/boundary the work changes (ISA extension, OS/hardware
contract, runtime/accelerator ABI, ...). No boundary named -> weak fit.
Q2 Deletion test per layer (above). Both halves survive alone -> route away.
Q3 Who must adopt this? If one community can adopt it without the other
noticing, the intersection claim is decorative.
Q4 What evidence class does the claim need — real silicon, FPGA, simulator?
If the answer is "none of these", the paper may be a PL/theory paper.
Q5 Can the contribution be stated in one sentence that a rapid reviewer
reading only two pages will believe? Draft that sentence now.
Choosing between the two 2027 deadlines
ASPLOS 2027 offers two submission gates — April 15, 2026 (passed as of 2026-07-08) and September 9, 2026 — framed by the CFP as "submit when ready." Treat the choice as an evidence-maturity decision, not a calendar accident:
- Submit September only if the evaluation matrix is complete; the two-deadline model removed the mid-year round, so a miss costs roughly half a year.
- A Major Revision outcome effectively adds a private third gate (revision due at the camera-ready deadline, six weeks after notification) — but only for papers whose core was sound. Do not submit a hollow evaluation hoping for the revision path; the revision counts as a submission on your record with that PC.
Distinctive-fit signals worth writing down
- The paper's insight is information that crosses a layer (what hardware knows that software needs, or vice versa) — the most reliably ASPLOS-shaped idea.
- The evaluation needs both a hardware-accuracy instrument (simulator/FPGA/silicon) and a software-realism instrument (kernel patch, runtime, full applications).
- Related work must be argued against two or three communities at once; if one community's literature suffices, so would its venue.
Intersection anti-patterns
Recurring shapes that look ASPLOS-flavored but fail the screen:
- The garnish paper — a complete single-community design plus one paragraph of "hardware support could accelerate this." The intersection must be load-bearing on page 1, not speculative in section 8.
- The stapler paper — an architecture result and an OS result about the same workload, bound in one PDF with no shared mechanism. Two rejectable halves do not sum to one acceptable whole.
- The tuning paper — heroic performance engineering on existing interfaces. Without a changed boundary or a transferable principle, it routes to ATC-style venues or an industrial track.
- The premature-generality paper — a mechanism evaluated only where it was conceived, claiming the whole intersection. Scope the claim to the evidence and let the routing table above pick the venue for the scoped claim.
Repairing a near-miss instead of rerouting
If the deletion test fails narrowly, a reframing pass is often cheaper than a venue change: identify the strongest cross-layer consequence of the existing result (a new OS policy the hardware data enables; a compiler contract the mechanism makes checkable), promote it from implication to evaluated contribution, and rebuild pages 1-2 around the promoted claim. Budget this as real research work — it usually needs one new experiment — and re-run the interrogation script afterward rather than assuming the patch took.
Output format
[Fit verdict] first-choice ASPLOS / viable but single-community-stronger / route away
[Boundary changed] <the named interface>
[Deletion test] hardware-only remainder: Y/N · software-only remainder: Y/N
[Route-away target] <venue + one-line reason, or n/a>
[Deadline call] September 9, 2026 readiness: evidence gaps listed
[Two-page sentence] <the one-sentence contribution a rapid reviewer must believe>
版本历史
- 9f86f09 当前 2026-07-19 14:19
