Skip to content

README.md

Latest commit

 

History

History
56 lines (37 loc) · 3.18 KB

README.md

File metadata and controls

56 lines (37 loc) · 3.18 KB

Crypto.Websocket.Extensions Benchmarks

This folder contains BenchmarkDotNet benchmarks for allocation-sensitive order book hot paths.

Run all benchmarks:

dotnet run --configuration Release --project benchmarks\Crypto.Websocket.Extensions.Benchmarks -- --filter "*"

Run the order book processing benchmarks:

dotnet run --configuration Release --project benchmarks\Crypto.Websocket.Extensions.Benchmarks -- --filter "*OrderBookProcessingBenchmarks*"

Run the order book read/access benchmarks:

dotnet run --configuration Release --project benchmarks\Crypto.Websocket.Extensions.Benchmarks -- --filter "*OrderBookReadBenchmarks*"

Results are written to BenchmarkDotNet.Artifacts\results.

What The Benchmarks Cover

  • OrderBookProcessingBenchmarks streams 1,000 prebuilt L2 diffs through CryptoOrderBookL2 and CryptoOrderBook, with and without an OrderBookUpdatedStream subscriber.
  • OrderBookReadBenchmarks measures snapshot-style accessors such as BidLevels, Levels, and FindBidLevelByPrice after a realistic stream of L2 updates has populated the book.

Use the Allocated column as the primary signal for GC pressure, and Mean/Ratio for throughput changes.

Current Results

Representative results from BenchmarkDotNet ShortRun on Windows, .NET 10.0.5, AMD Ryzen 9 3900X.

Order Book Processing

Benchmark Before After Improvement
CryptoOrderBookL2, no observers 935.4 ns, 545 B 618.4 ns, 161 B 34% faster, 70% less allocation
CryptoOrderBookL2, OrderBookUpdatedStream observer 933.6 ns, 545 B 683.9 ns, 337 B 27% faster, 38% less allocation
CryptoOrderBook, no observers 1,781.2 ns, 1,154 B 1,140.0 ns, 770 B 36% faster, 33% less allocation
CryptoOrderBook, OrderBookUpdatedStream observer 1,629.1 ns, 1,155 B 1,238.4 ns, 947 B 24% faster, 18% less allocation

The processing improvements mainly come from avoiding notification allocation when no stream has subscribers, handling the common single-diff path directly, using an internal single-bulk handoff from OrderBookSourceBase to order books while preserving the public OrderBookStream API, lazily materializing single-source notification metadata only when subscribers read Sources, dispatching built-in bulk updates through indexed IReadOnlyList<T> loops instead of interface enumerators, avoiding repeated dictionary lookups in per-level update/delete paths, skipping temporary collections in buffered/small-batch paths, and caching observable stream wrappers.

Order Book Reads

Benchmark Before After Improvement
CryptoOrderBook.BidLevels 17,822.5 ns, 77,032 B 5,408.8 ns, 4,832 B 70% faster, 94% less allocation
CryptoOrderBook.FindBidLevelByPrice 37.6 ns, 80 B 24.3 ns, 0 B 35% faster, allocation-free
CryptoOrderBook.Levels 3,103.4 ns, 9,600 B 2,964.3 ns, 9,600 B 4% faster
CryptoOrderBookL2.BidLevels 1,721.3 ns, 4,832 B 1,101.1 ns, 4,832 B 36% faster

The read-path improvement for CryptoOrderBook comes from removing OrderedDictionary value enumerator allocations in grouped L3-compatible level access.