nordisk/nornir

nornir-mimir-graph (mimir_graph)

Mímir v2 graph engine (plan PLAN-MIMIR-V2-2026-10-12.md §4/§6/§11, lane 7): a small safe layer over SuiteSparse:GraphBLAS via our own graphblas-sys. Package name nornir-mimir-graph (mimir-graph on crates.io is someone else's); library name mimir_graph.

Needs the system library: sudo apt-get install -y libgraphblas-dev (7.4).

What is here

• Matrix<T> / Vector<T> for bool, u64, f64 — RAII (GrB_free on drop), every call returns Result, Send but not Sync.

• runtime: GrB_init(GrB_NONBLOCKING) once per process (lazy), linked version check (7.x), set_threads / threads, unsafe finalize.

• CSR import:

  • Matrix::from_csr(nrows, ncols, offsets, cols, vals: Option<&[T]>) — validate, one memcpy into malloc'd arrays, GxB_Matrix_pack_CSR (GraphBLAS takes ownership, no further copy). A borrowed slice can't be packed without that copy: GraphBLAS frees packed arrays with its own free.
  • Matrix::from_csr_import — same contract via GrB_Matrix_import (copies inside GraphBLAS); kept for the bench comparison.
  • arrow::from_arrow_edges(batch, from_col, to_col, n_nodes) / from_arrow_edge_batches(&[batch], ..) — UInt64 (from_ord, to_ord) columns: sorted-by-(from,to) fast path (count + cumsum straight into the packed offsets array, to buffer copied once), sorted-by-from with jumbled rows (packed, duplicate-checked), otherwise GxB_Matrix_build_Scalar (GraphBLAS parallel sort, duplicates collapse; reads a single batch's Arrow buffers in place).

• ops: reach, reverse_reach, fan_in, fan_out, min_hops, max_on_path, mask_by, mask_vector, similarity (A·Aᵀ), components, diff — the algebra of each is tabled in src/ops.rs.

Notes from the bench (BENCH-2026-10-12.md)

• reverse_reach (mxv on A, no transpose) is 2–10× (typically ~5×) slower than reach on a materialised Aᵀ. A lens doing several dependents walks should build a.transpose() once and call reach on it.
• Direction-optimising BFS (push/pull switch with Aᵀ) was tried and showed no clear win on the shared box (load 25–38); not kept.
• components is plain label propagation + one pointer jump per round, not FastSV; it is the slowest op (seconds at 10⁷–10⁸ edges).

Not here (yet)

• No op is missing for 7.4 reasons. Not written because no lens needs it yet: general mxm with a caller-chosen semiring, weighted shortest paths (min_hops assumes unit edges), cycle listing (SCCs).
• Bench numbers come from an example runner, not nornir's register_bench! registry: the registry compiles into the nornir lib, which would make every nornir build link libgraphblas.