Skip to content

GPU-parallel independent search replicates #37

Description

@ms609

Priority: — · Former status: DEFERRED · Blocks:

GPU-parallel independent search replicates. Run 100+ search replicates simultaneously on GPU SMs (one replicate per SM; modern GPUs have 60–128 SMs). Shared read-only character matrix fits in GPU L2 cache. Main obstacle: tree search has highly irregular, data-dependent control flow (rearrangement selection, acceptance decisions, ratchet perturbation) which causes warp divergence and poor GPU utilization. Branch-and-bound in sectorial search has the same problem. CPU multicore parallelism (8–16 cores via future/parallel::mclapply, or 100+ via HPC SLURM array jobs) is far simpler and more efficient per-replicate. GPU replicates only become attractive if per-replicate arithmetic is heavy enough to dominate over control flow overhead (e.g., large Sankoff matrices). References: same as T-290.

Notes. Research: MkPrime .agent-d.md 2026-03-29.


Migrated 2026-08-04 from the retired to-do.md; former task id T-291. Full original row: git show findings-md-final:to-do.md. Task state now lives in this tracker — see dev/red-team/migration-map.tsv for red-team ids.

Metadata

Metadata

Assignees

No one assigned

    Labels

    deferredAssessed and parked; not scheduledtaskPlanned work migrated from the retired to-do.md

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions