jgarte / 42nd-at-threadmill

A SIMD-accelerated concurrent hash table.

A fast concurrent hash table.

42nd At Threadmill is a nearly lock-free* hash table based on Cliff Click's NonBlockingHashMap, and Abseil's flat_hash_map. We use the general layout of the former, and the fast metadata-based probing trick of the latter.

We are aware of the table being very, very picky with hash functions (Abseil's table is like that too, but I don't think it's that bad), so you might want to hold off using this table in production still.

See A Fast Wait-Free Hash Table and Matt Kulukundis's "Designing a Fast, Efficient, Cache-friendly Hash Table, Step by Step" presentation for an introduction to both tables.

We use SSE2 intrinsics for fast probing, and optionally use AVX2 for faster byte broadcasting. This library requires a post-2.0.5 version of SBCL, so that we can use some instructions introduced to the assembler around then.

*Okay, we effectively lock to resize but it won't deadlock and it appears to be faster than using Click's lock-free resizing technique; so you tell me if that is an acceptable trade-off.

Pictures of a benchmark

Differences from Click's table

We replace copied values with a single +copied+ marker instead of an instance of a Prime class. This change generates less garbage when copying, and leads to slightly faster barrier code; though our table is not wait-free. (Truth be told, we're only barely getting a grip on Click's resize logic now.)

We also removed the <key, tombstone> state, having removes transition back to <key, empty>; a superficial change which doesn't appear to affect anything.

Differences from Kulukundis's table

As Click requires us to pin keys to entries, we don't ever use tombstone metadata. The metadata for a dead entry remains in the metadata table, as we need to be able to find the right key entry to reuse quickly.

We have a somewhat improved load factor, but it is not as extreme as Kulukundis's demonstration. Kulukundis could approach a load factor of 87.5%, but we find that 50% is the best tradeoff between space and throughput with our table. However, we still have improved over Click's table -- Click doubles the table size with 25% live keys, and quadruples with 50% live, in order to "avoid endless reprobing". Faster probing lets us get away with just doubling at 50%.

Previous work

This concurrent hash table is based off the NonBlockingHashMap, and its Common Lisp port in Luckless. It is also based off the linear probing hash table implementation in SICL, as well as its SIMD fork.