perf(mpi): 🧭 GF(2)-linear rank routing — messages per rank flat in R - #296

Open

diagonal-hamiltonian wants to merge 33 commits into

mainfrom

perf/linear-routing-on-wire

Commits on Sep 5, 2026

perf(mpi): 🧭 route through one Router, GF(2)-linear across ranks

authored and
robertodr
committed
16f8b24
perf(mpi): ⏭️ skip the exchange for identity generators

authored and
robertodr
committed
7150164
perf(mpi): 🕸️ exchange point-to-point over the peers routing can reach

authored and
robertodr
committed
1f223b4
perf(mpi): 🧭 route GF(2)-linear by default where the geometry allows it

authored and
robertodr
committed
477d974
test(mpi): 🧭 check scan/find_rank agreement under both routers

authored and
robertodr
committed
06f7bd8
style(tests): 🎨 clang-format mpi_utils_tests

authored and
robertodr
committed
8e049dd
docs(parallelism): 📝 document GF(2)-linear rank routing and its knobs

diagonal-hamiltonian
authored and
robertodr
committed
ec1955e
fix(mpi): 🐛 mask known recv counts through the peer plan

diagonal-hamiltonian
authored and
robertodr
committed
1012e90
refactor(mpi): ♻️ drive every sparse exchange from one pairwise helper

diagonal-hamiltonian
authored and
robertodr
committed
efb214f
feat(mpi): 📈 report a shift set that cannot reach every rank

diagonal-hamiltonian
authored and
robertodr
committed
f3bba01
perf(evolution): ⏭️ skip the scan too for an identity generator

diagonal-hamiltonian
authored and
robertodr
committed
06e05cd
perf(routing): ⚡ take the rank bits from a transposed basis

diagonal-hamiltonian
authored and
robertodr
committed
7ab0e53
perf(mpi): ⚡ complete the sparse exchange through the handle

diagonal-hamiltonian
authored and
robertodr
committed
7f5ccfb
test(mpi): ✅ reach the sparse paths a single peer cannot

diagonal-hamiltonian
authored and
robertodr
committed
d63c7bd
refactor(routing)!: ♻️ collapse the linear-bit dial to a boolean

diagonal-hamiltonian
authored and
robertodr
committed
8e33cb9
perf(routing): ⚡ route the emit path from the generator's shift

diagonal-hamiltonian
authored and
robertodr
committed
63124c2
perf(mpi): ⚡ pack the send side under the count round

diagonal-hamiltonian
authored and
robertodr
committed
ad8d4a5
perf(mpi): ⚡ replay the layer exchange point-to-point

diagonal-hamiltonian
authored and
robertodr
committed
6ddbf28
fix(mpi): 🐛 gate the pairwise replay on routing, not on a rank's row

diagonal-hamiltonian
authored and
robertodr
committed
2c4077a
fix(mpi): 🔀 reconcile the boolean peer plan across the stack

authored and
robertodr
committed
03c6771
feat(mpi): ✨ name the slot window a peer plan can reach

diagonal-hamiltonian
authored and
robertodr
committed
34f9ffe
perf(evolution): ⚡ size the per-generator query path to the peer window

diagonal-hamiltonian
authored and
robertodr
committed
8b94005
perf(mpi): ⚡ narrow HybridComm's serial staging sweeps to the peer ranks

diagonal-hamiltonian
authored and
robertodr
committed
5d79045
test(mpi): ✅ cover derived_wire_plan_ with an empty partition-0 row

diagonal-hamiltonian
authored and
robertodr
committed
3e28e10
fix(mpi): 🐛 export mpi::geometry across the shared-library boundary

diagonal-hamiltonian
authored and
robertodr
committed
563ea9e
fix(tests): 🐛 route the paired-enumeration test through a Router
robertodr
committed
24ddc3c
fix(quality): 🐛 resolve clang-tidy and SonarQube warnings
robertodr
committed
dfe46de
chore(ci): 🩺 unmute MPI ranks for test diagnostics
robertodr
committed
fc5a8d2
test(cpp): ✅ retain stale row-parity storage during growth
robertodr
committed
ca6a8ea
fix(cpp): 🐛 validate Schrodinger cardinality before partitioning
robertodr
committed
aa4f8bf

Commits on Sep 7, 2026

Merge branch 'main' into perf/linear-routing-on-wire
robertodr
authored
8711880

Navigation Menu

Uh oh!

perf(mpi): 🧭 GF(2)-linear rank routing — messages per rank flat in R - #296

perf(mpi): 🧭 GF(2)-linear rank routing — messages per rank flat in R#296
diagonal-hamiltonian wants to merge 33 commits into
mainfrom
perf/linear-routing-on-wire

Commits on Sep 5, 2026

perf(mpi): 🧭 route through one Router, GF(2)-linear across ranks

perf(mpi): ⏭️ skip the exchange for identity generators

perf(mpi): 🕸️ exchange point-to-point over the peers routing can reach

perf(mpi): 🧭 route GF(2)-linear by default where the geometry allows it

test(mpi): 🧭 check scan/find_rank agreement under both routers

style(tests): 🎨 clang-format mpi_utils_tests

docs(parallelism): 📝 document GF(2)-linear rank routing and its knobs

fix(mpi): 🐛 mask known recv counts through the peer plan

refactor(mpi): ♻️ drive every sparse exchange from one pairwise helper

feat(mpi): 📈 report a shift set that cannot reach every rank

perf(evolution): ⏭️ skip the scan too for an identity generator

perf(routing): ⚡ take the rank bits from a transposed basis

perf(mpi): ⚡ complete the sparse exchange through the handle

test(mpi): ✅ reach the sparse paths a single peer cannot

refactor(routing)!: ♻️ collapse the linear-bit dial to a boolean

perf(routing): ⚡ route the emit path from the generator's shift

perf(mpi): ⚡ pack the send side under the count round

perf(mpi): ⚡ replay the layer exchange point-to-point

fix(mpi): 🐛 gate the pairwise replay on routing, not on a rank's row

fix(mpi): 🔀 reconcile the boolean peer plan across the stack

feat(mpi): ✨ name the slot window a peer plan can reach

perf(evolution): ⚡ size the per-generator query path to the peer window

perf(mpi): ⚡ narrow HybridComm's serial staging sweeps to the peer ranks

test(mpi): ✅ cover derived_wire_plan_ with an empty partition-0 row

fix(mpi): 🐛 export mpi::geometry across the shared-library boundary

fix(tests): 🐛 route the paired-enumeration test through a Router

fix(quality): 🐛 resolve clang-tidy and SonarQube warnings

chore(ci): 🩺 unmute MPI ranks for test diagnostics

test(cpp): ✅ retain stale row-parity storage during growth

fix(cpp): 🐛 validate Schrodinger cardinality before partitioning

Commits on Sep 7, 2026

Merge branch 'main' into perf/linear-routing-on-wire

Commits on Sep 8, 2026

Merge branch 'main' into perf/linear-routing-on-wire

Merge branch 'main' into perf/linear-routing-on-wire