glass: ordered set data structure for client-side order books

TL;DR

This paper introduces a trie-based ordered set data structure optimized for market data, achieving significant speedups over standard implementations through novel caching, hardware acceleration, and market-specific features.

Contribution

The paper presents a new trie-based ordered set implementation with optimizations for market data, including caching, hardware acceleration, and memory management techniques.

Findings

6x-20x faster modifications compared to std::map

30x faster lookups on market data

Significant speedups in real market data scenarios

Abstract

The "ordered set" abstract data type with operations "insert", "erase", "find", "min", "max", "next" and "prev" is ubiquitous in computer science. It is usually implemented with red-black trees, $B$-trees, or $B^+$-trees. We present our implementation of ordered set based on a trie. It only supports integer keys (as opposed to keys of any strict weakly ordered type) and is optimized for market data, namely for what we call sequential locality. The following is the list of what we believe to be novelties: * Cached path to exploit sequential locality, and fast truncation thereof on erase operation; * A hash table (or, rather, a cache table) with hard O(1) time guarantees on any operation to speed up key lookup (up to a pre-leaf node); * Hardware-accelerated "find next/previous set bit" operations with BMI2 instruction set extension on x86-64; * Order book-specific features: the preemption principle and the tree restructure operation that prevent the tree from consuming too much memory. We achieve the following speedups over C++'s standard std::map container: 6x-20x on modifying operations, 30x on lookup operations, 9x-15x on real market data, and a more modest 2x-3x speedup on iteration. In this paper, we discuss our implementation.