技能 硬件工程 计算机架构研究流程规划

计算机架构研究流程规划

v20260724
micro-workflow
本文档规划了计算机架构领域顶级会议的多年论文投稿全流程。它详细描述了从初稿到最终展示的整个研究生命周期,不仅关注时间节点,更强调评估周期(如仿真和计算队列)是整个项目中最核心的难点和时间瓶颈,适用于科研人员进行长期项目排期和风险管理。
获取技能
126 次下载
概览

MICRO Workflow

MICRO owns the autumn slot of computer architecture's four-venue year: an early
April deadline feeding an end-of-October symposium. Because the four sibling venues
interleave, a MICRO plan is really a lattice plan — every outcome has a next move
within roughly four months. Dates below verified 2026-07-08; every one is a
single-cycle snapshot.

The 2026 cycle clock (59th edition, Athens)

2026-03-31  abstract registration       11:59 PM EDT   hard gate
2026-04-07  full paper                  11:59 PM EDT   micro2026.hotcrp.com
2026-06-03 → 06-17  rebuttal/revision   two-week window (micro-author-response)
2026-07-07  notification
2026-09-11  camera-ready               (AE runs in parallel; micro-camera-ready)
2026-10-31 → 11-04  symposium, Athens, Greece

As of this pack's check date (2026-07-08), the 2026 cycle had **just issued
decisions** — the live work is camera-ready + AE for accepts, and lattice routing
for rejects. The next MICRO submission opportunity is the 2027 cycle, expected in
the same early-April position (待核实 when microarch.org posts it).

The four-venue calendar from MICRO's seat

VenueDeadline (verified for cycle shown)ConferenceUse it when
MICRO 2026Apr 7, 2026 (closed)Oct 31–Nov 4, 2026, AthensMechanism-centric core work
HPCA 2027July 24, 2026March 2027, Salt Lake CityFastest next slot after a MICRO reject — 17 days after notification
ASPLOS 2027Sept 9, 2026 (2nd of two; Apr 15 passed)April 2027, CreteReviews said "the interesting part is the software interaction"
ISCA 2027abstracts Nov 10 / papers Nov 17, 2026June 2027, AtlantaFour months of repair headroom; broadest architecture audience
MICRO 2027~April 2027, 待核实autumn 2027Full-year rework or a new project

The rhythm to internalize: **architecture deadlines cluster in April
(MICRO, ASPLOS-1st), July (HPCA), September (ASPLOS-2nd), and November (ISCA)** —
one submission window roughly every quarter. A healthy lab pipeline drafts against
the next two windows simultaneously.

Backward plan for a MICRO 2027 submission

Working back from an assumed early-April 2027 deadline (verify the real date):

WhenMilestoneOwner risk
Aug–Sep 2026Mechanism sketch + micro-topic-selection gate; infrastructure choice (simulator, workloads)Wrong venue discovered late
Oct–Nov 2026Baseline validated against published configs; characterization study doneUntuned baseline poisons all later data
Dec 2026Mechanism implemented in simulator; first end-to-end numbers"It simulates" ≠ "it is modeled correctly"
Jan 2027Ablations + sensitivity sweeps; power/area modeling attachedSweeps are compute-bound — queue early
Feb 2027Full draft; internal red-team review against micro-writing-styleCost accounting missing
Mar 2027Freeze evaluation; micro-reproducibility manifests complete; format passLate "one more feature" restarts sweeps
T-7 daysAbstract registered in HotCRPMissing the registration gate ends everything
T-0Upload (EDT arithmetic!)See micro-submission

The structural insight: in microarchitecture the **evaluation stack is the
schedule**. Simulator bring-up, trace generation, and multi-week sweep queues
dominate; prose is the cheap part. Any plan that allocates most of its calendar to
writing is mis-costed.

Risk register for the compute-bound schedule

RiskEarly symptomMitigation
Sweep queue starvationCluster contention in Feb–Mar (everyone's MICRO crunch)Reserve allocation in January; keep key-tier configs that finish overnight
Baseline invalidated lateA simulator bugfix changes baseline IPCFreeze simulator commit by January; batch fixes into one revalidation
Trace regeneration cascadeNew compiler/flags after traces cutVersion traces; treat regeneration as a one-week line item, not a checkbox
Writing squeezed to two weeks"Evaluation almost done" through FebruaryDraft intro + mechanism sections in parallel with sweeps — they do not depend on final numbers
Deadline-week format panicFirst full format pass happens at T-2Run the micro-submission gate table at T-14 and T-1

Standing weekly cadence during a campaign

Mon   queue review: sweeps finished / stalled / mis-configured; ledger updated
Wed   claims meeting: does yesterday's data still support the paper's sentences?
Fri   risk register walk (table above) + one live-page re-verification
      (deadline, CFP, HotCRP) logged with date — rules change between visits

Outcome branches (July 7 has happened)

  • Accepted: run micro-camera-ready and micro-artifact-evaluation as

parallel tracks converging on September 11; book Athens logistics once
attendance rules are confirmed (待核实 on micro59/attend/).

  • Rejected, methodology wounds: the HPCA July 24 window is realistically only

for papers whose fixes are textual; simulation repairs point at ISCA Nov 17.

  • Rejected, "not microarchitectural enough": ASPLOS Sept 9 with a cross-layer

reframe — see micro-topic-selection routing table.

  • Any branch: archive reviews, update the results ledger, and file rebuttal

promises as the next version's spec.

Multi-paper pipelines and self-competition

Labs running more than one submission per cycle share one sweep queue and one set
of senior readers — schedule both explicitly:

  • Stagger internal red-team reviews (micro-writing-style checklist) so the same

professor is not reading two full drafts in deadline week.

  • Deconflict topics in HotCRP if two submissions are adjacent; two papers from

one lab splitting the same reviewer pool can drag each other down in
double-blind bidding without anyone noticing until the reviews land.

  • After notification, the lattice branches diverge per paper — keep one shared

calendar of the four venues' gates (table above) so a July HPCA sprint for
paper A does not silently consume the cluster paper B needs for its ISCA
repairs.

Output format

[Today] <date> → position in cycle: <phase>
[Active gate] <next deadline, venue, days remaining, timezone stated>
[Lattice] plan A <venue+date> · plan B <venue+date> · trigger condition between them
[Evaluation-stack status] simulator / baseline / traces / sweeps: ready-blocked each
[Backward plan] next 3 milestones with dates and owners
[待核实] <dates assumed but not yet posted on live pages>
信息
分类 硬件工程
名称 micro-workflow
版本 v20260724
大小 6.81KB
更新时间 2026-07-28
语言