Skip to content

phased roadmap.md

Latest commit

 

History

History
116 lines (94 loc) · 3.99 KB

phased roadmap.md

File metadata and controls

116 lines (94 loc) · 3.99 KB
Raw

Phased Roadmap (v0.1)

Objective

Sequence execution from architecture spec to credible prototypes with explicit gates and dependencies.

Phase 0: Consolidation (0-8 weeks)

Goal

Lock definitions, claims, and benchmark protocol.

Inputs

  • unified architecture spec.md
  • benchmark disproof matrix.md
  • Existing vision and landscape notes

Deliverables

  • Stable interface glossary
  • Claim-to-benchmark traceability table
  • Prototype scope document (what is in/out for phase 1)

Exit Criteria

  • No unresolved contradictions across core contracts
  • All major claims mapped to measurable tests

Phase 1: seL4-Oriented Runtime Prototype (2-6 months)

Goal

Prove software-only viability without ISA changes.

Scope

  • WorkGraph admission and runtime dependency checks
  • Ownership transition tracking for buffers/pages
  • Capability-scoped execution boundaries
  • Polling/threshold event path for at least one high-rate I/O path

Out of Scope

  • New opcodes
  • Full production driver stack
  • End-user programming language ergonomics

Milestones

  1. P1-M1: WorkNode schema admission pipeline operational
  2. P1-M2: Dependency-safe dispatch and completion signaling
  3. P1-M3: Ownership transfer integrated with I/O path
  4. P1-M4: Baseline benchmark harness executing all 3 workloads

Gate to Phase 2

  • Meet promotion rules in benchmark matrix (minimum threshold pass conditions)
  • No unresolved safety invariant violations

Phase 2: Compiler Bridge (4-10 months, overlaps with late Phase 1)

Goal

Emit kernel-consumable work graph declarations from compiler/runtime IR.

Scope

  • Define intermediate schema mapping from MLIR/TVM/XLA-like graph IR to WorkNode declarations
  • Generate deps, region access descriptors, deadline class, and capability requests
  • Validate fidelity between IR graph and runtime graph

Milestones

  1. P2-M1: IR-to-WorkNode mapping draft
  2. P2-M2: Auto-generated graph declarations for tensor workload
  3. P2-M3: Graph fidelity checks and diagnostics
  4. P2-M4: End-to-end benchmark with generated declarations

Gate to Phase 3

  • Compiler-emitted declarations perform within 5-10% of hand-authored declarations
  • Declaration errors are diagnosable with actionable tooling output

Phase 3: Hardware Extension Track (9-24 months)

Goal

Evaluate which semantic opcodes provide material benefit and justify hardware collaboration.

Scope

  • Emulation/proxy evaluation of intent-bearing opcode classes
  • Prioritize minimal opcode set by measured impact
  • Draft proposal language for open ISA ecosystem (e.g., RISC-V channels)

Candidate Classes

  • Graph completion signaling
  • Deep prefetch intent
  • Tagged branch speculation hints
  • Tensor tile fused operations (where not already available)

Milestones

  1. P3-M1: Opcode candidate shortlist ranked by expected benefit
  2. P3-M2: Software proxy benchmark evidence
  3. P3-M3: Draft extension proposal package

Exit Criteria

  • At least 1-2 opcode classes show material gains beyond software-only ceiling

Cross-Phase Streams

  • Verification stream: Invariant checking and failure classification.
  • Tooling stream: Declaration linting, graph debug visualizer, workload trace export.
  • Collaboration stream: Regular touchpoints across OS, compiler, and hardware contributors.

Dependency Graph

flowchart TD
p0[Phase0Consolidation] --> p1[Phase1seL4RuntimePrototype]
p1 --> p2[Phase2CompilerBridge]
p2 --> p3[Phase3HardwareExtension]
p1 --> verify[VerificationStream]
p2 --> tooling[ToolingStream]
p3 --> collab[CollaborationStream]
verify --> p3
tooling --> p3
collab --> p3
Loading

Risks and Mitigations

  • Risk: Prototype overhead erodes gains.
    • Mitigation: Early ablations and strict disproof conditions.
  • Risk: Graph declaration burden is too high.
    • Mitigation: Prioritize compiler-generated declarations in Phase 2.
  • Risk: Cross-discipline collaboration stalls.
    • Mitigation: Maintain shared artifact contracts and monthly integration reviews.