Rutherford-Boeing

Rutherford-Boeing (RB) is a fixed-format ASCII file format for storing sparse matrices, defined at the Rutherford Appleton Laboratory around 1997 by Iain Duff, Roger Grimes and John Lewis. It is the direct successor to the older Harwell-Boeing format (see Harwell-Boeing), sharing the same authors and much of the same column-oriented layout, and it was introduced as the exchange format for what was renamed the Rutherford-Boeing sparse matrix collection. Because sparse matrices are dominated by zeros, RB stores only the nonzero entries, making it far more compact than a dense representation.

A matrix can be held in one of two representations. The common case is compressed column storage (CCS), used for general and symmetric assembled matrices, where an array of column pointers is followed by row indices and, unless the matrix is pattern-only, the numerical values. The second representation is a finite-element form for unassembled matrices, which are expressed as a sum of small elemental contributions rather than a single assembled array. Each file begins with a small header: a title/identifier line, a line of line-counts, a three-character type code, and a line of Fortran fixed-format edit descriptors (for example an integer format for the pointers and indices and a floating format for the values) that tell a reader exactly how to parse the columns of data that follow.

The three-letter type code (often mirrored in the file extension) encodes the matrix's properties. The first character gives the arithmetic — r for real, c for complex, p for pattern-only; the second gives the symmetry — s symmetric, u unsymmetric, h Hermitian, z skew-symmetric, r rectangular; and the third is a for assembled or e for elemental/unassembled. A distinguishing improvement over Harwell-Boeing is a companion supplementary-file convention that lets associated data such as right-hand sides, starting vectors or eigenvectors be stored and cleanly associated with a matrix, instead of being crammed into the single matrix file.

In the graph-data world RB matters because a sparse matrix is, structurally, the adjacency or incidence matrix of a graph, so RB files double as a compact encoding of weighted directed or undirected graphs — much like other matrix-based graph formats such as Matrix Market (with which RB is frequently paired as an alternate serialization), Harwell-Boeing, and coordinate lists like DIMACS. Its strengths are compactness, long provenance, and a canonical role in numerical-linear-algebra benchmarking. Its limitations for general graph interchange are real: the reliance on Fortran fixed-format columns makes hand-editing and generic tooling awkward, it carries no vertex/edge attributes or labels beyond numeric weights, and it is essentially unknown outside the scientific-computing community, so most graph software cannot read it without a dedicated converter.