--- name: micro-workflow description: Use when planning a MICRO campaign end to end — the April-to-Athens 2026 cycle clock, the four-venue architecture calendar (MICRO/ISCA/HPCA/ASPLOS) seen from MICRO's October seat, resubmission lattices after each outcome, and starting the MICRO 2027 evaluation stack early enough to matter. --- # 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) ```text 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 | Venue | Deadline (verified for cycle shown) | Conference | Use it when | |---|---|---|---| | **MICRO 2026** | Apr 7, 2026 (closed) | Oct 31–Nov 4, 2026, Athens | Mechanism-centric core work | | HPCA 2027 | **July 24, 2026** | March 2027, Salt Lake City | Fastest next slot after a MICRO reject — 17 days after notification | | ASPLOS 2027 | Sept 9, 2026 (2nd of two; Apr 15 passed) | April 2027, Crete | Reviews said "the interesting part is the software interaction" | | ISCA 2027 | abstracts Nov 10 / papers **Nov 17, 2026** | June 2027, Atlanta | Four months of repair headroom; broadest architecture audience | | MICRO 2027 | ~April 2027, 待核实 | autumn 2027 | Full-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): | When | Milestone | Owner risk | |---|---|---| | Aug–Sep 2026 | Mechanism sketch + `micro-topic-selection` gate; infrastructure choice (simulator, workloads) | Wrong venue discovered late | | Oct–Nov 2026 | Baseline validated against published configs; characterization study done | Untuned baseline poisons all later data | | Dec 2026 | Mechanism implemented in simulator; first end-to-end numbers | "It simulates" ≠ "it is modeled correctly" | | Jan 2027 | Ablations + sensitivity sweeps; power/area modeling attached | Sweeps are compute-bound — queue early | | Feb 2027 | Full draft; internal red-team review against `micro-writing-style` | Cost accounting missing | | Mar 2027 | Freeze evaluation; `micro-reproducibility` manifests complete; format pass | Late "one more feature" restarts sweeps | | T-7 days | Abstract registered in HotCRP | Missing the registration gate ends everything | | T-0 | Upload (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 | Risk | Early symptom | Mitigation | |---|---|---| | Sweep queue starvation | Cluster contention in Feb–Mar (everyone's MICRO crunch) | Reserve allocation in January; keep `key`-tier configs that finish overnight | | Baseline invalidated late | A simulator bugfix changes baseline IPC | Freeze simulator commit by January; batch fixes into one revalidation | | Trace regeneration cascade | New compiler/flags after traces cut | Version traces; treat regeneration as a one-week line item, not a checkbox | | Writing squeezed to two weeks | "Evaluation almost done" through February | Draft intro + mechanism sections in parallel with sweeps — they do not depend on final numbers | | Deadline-week format panic | First full format pass happens at T-2 | Run the `micro-submission` gate table at T-14 and T-1 | ## Standing weekly cadence during a campaign ```text 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 ```text [Today] → position in cycle: [Active gate] [Lattice] plan A · plan B · 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] ```