ankerl::unordered_dense

Martin Ankerl's single-header C++ hash map and the uncontested iteration champion. It stores all key-value pairs in a contiguous std::vector rather than the sparse slot array of swiss-table designs, making iteration as fast as scanning an array — 2.55× faster than abseil-flat-hash-map in Jackson Allan's benchmark. Robin Hood probing with backward-shift deletion sidesteps tombstone degradation entirely.

The dense storage trick

A separate index table holds slot metadata; the actual entries live packed in a std::vector. Iteration walks the vector linearly with full hardware prefetching. Erase is a swap-and-pop that keeps the vector contiguous. The cost: insertions are slightly more expensive (the index update plus the vector append, 1.78× the leader on integer insert), and lookups touch one extra cache line (1.15× on lookup-hit — still very competitive).

When it wins

• Iteration-bound workloads — analytics, snapshotting, walking a map repeatedly. Nothing else comes close.
• Small-to-medium maps (200–2,000 entries) where dense layout fits in L1/L2.
• Single-header drop-in — simplest possible integration. No Boost, no Abseil dependency.
• Sustained insert/erase cycling — Robin Hood backward-shift means no tombstone rot.

When to pick something else

For pure single-threaded throughput on insert/erase-heavy workloads with little iteration, boost-unordered-flat-map wins. For lowest memory footprint, folly-f14. For concurrent access, see parlayhash / papaya / dashmap.

The broader landscape

ankerl::unordered_dense sits in the Jackson Allan benchmark alongside the Swiss Table family. Its dense-layout idea is conceptually closer to ECS archetype storage or SoA layouts — pay a small lookup cost to make the dominant access pattern (sequential traversal) cache-perfect. See fastest-hash-map-2025 for full context.