FrozenDictionary

C#/.NET 8's read-optimized immutable dictionary, taking a different angle on the "fastest hash map" question: analyze the keys at construction time and pick a lookup strategy specialized to that key set. The result is 43–69% faster string lookups than mutable Dictionary<TKey, TValue> — about 4.3 ns versus 10–18 ns for small dictionaries.

The construction-time analysis

When you call dict.ToFrozenDictionary(), the runtime inspects the keys and chooses among several internal layouts: integer-key tables, length-bucketed string tables, ordinal-comparison string tables, single-substring-hash tables, and small-array linear search for tiny dictionaries. The chosen layout is then specialized — different code paths for different key shapes.

This is conceptually the dual of swiss-table: instead of a single architecture that handles all key types well, FrozenDictionary picks the best architecture for this key set. The cost is amortized at build time and frozen forever after.

When it pays off

The break-even point is roughly 5,000–630,000 reads depending on dictionary size. Below that, the construction cost outweighs the per-lookup savings. Above that, FrozenDictionary wins decisively.

Right choice for:

• Configuration loaded at startup and queried for the process lifetime.
• Compiled lookup tables (HTTP headers, MIME types, error code maps).
• Routing tables, feature flag maps, anything built once and read forever.

Wrong choice for:

• Anything mutable.
• Dictionaries that turn over frequently.
• Cold dictionaries hit only a handful of times.

How it relates to perfect hashing

FrozenDictionary stops short of full perfect-hashing — it does not guarantee zero collisions. It picks a good layout per key set, not an optimal one. The construction is much faster than perfect-hash builders (PTHash, RecSplit), and the lookup cost is correspondingly slightly higher. For most application workloads it is the right pragmatic point on the curve: significant speedup, fast enough to build at startup, no special tooling.

Other languages

• C++: similar role played by frozen (Quentin Quintet), a constexpr perfect-hash library that builds at compile time.
• Rust: phf for compile-time perfect hashing of static maps. OnceLock<HashMap> for runtime-built read-only maps without specialized layout.
• Java: no equivalent; primitive-specialized maps (Koloboke, fastutil) provide their own speedup axis.

See fastest-hash-map-2025 for context and perfect-hashing for the upper-bound version of this idea.