Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hive Hash Table: A Warp-Cooperative, Dynamically Resizable Hash Table for GPUs

Md Sabbir Hossain Polak, David Troendle, Byunghyun Jang

TL;DR

Hive hash table tackles the challenge of high-throughput, concurrent hash tables on GPUs by introducing a warp-cooperative, dynamically resizing open-addressing cuckoo design. It uses a cache-aligned 64-bit packed KV bucket layout, WABC for fast slot claiming, WCME for coordinated updates, and a load-aware warp-parallel linear hashing resizing strategy to avoid global rehashing. The approach delivers lock-free fast paths with bounded eviction costs, achieving sustained high occupancy and superior throughput compared with state-of-the-art GPU hash tables under balanced and mixed workloads. The results on an RTX 4090 show scalable performance with negligible resizing overhead, enabling efficient GPU-accelerated data processing; the authors also plan an open-source release for reproducibility.

Abstract

Hash tables are essential building blocks in data-intensive applications, yet existing GPU implementations often struggle with concurrent updates, high load factors, and irregular memory access patterns. We present Hive hash table, a high-performance, warp-cooperative and dynamically resizable GPU hash table that adapts to varying workloads without global rehashing. Hive hash table makes three key contributions. First, a cache-aligned packed bucket layout stores key-value pairs as 64-bit words, enabling coalesced memory access and atomic updates via single-CAS operations. Second, warp-synchronous concurrency protocols - Warp-Aggregated-Bitmask-Claim (WABC) and Warp-Cooperative Match-and-Elect (WCME) - reduce contention to one atomic operation per warp while ensuring lock-free progress. Third, a load-factor-aware dynamic resizing strategy expands or contracts capacity in warp-parallel K-bucket batches using linear hashing, maintaining balanced occupancy. To handle insertions under heavy contention, Hive hash table employs a four-step strategy: replace, claim-and-commit, bounded cuckoo eviction, and overflow-stash fallback. This design provides lock-free fast paths and bounded recovery cost under contention determined by a fixed eviction depth, while eliminating ABA hazards during concurrent updates. Experimental evaluation on an NVIDIA RTX 4090 shows Hive hash table sustains load factors up to 95% while delivering 1.5-2x higher throughput than state-of-the-art GPU hash tables (Slab-Hash, DyCuckoo, WarpCore) under mixed insert-delete-lookup workloads. On balanced workload, Hive hash table reaches 3.5 billion updates/s and nearly 4 billion lookups/s, demonstrating scalability and efficiency for GPU-accelerated data processing.

Hive Hash Table: A Warp-Cooperative, Dynamically Resizable Hash Table for GPUs

TL;DR

Hive hash table tackles the challenge of high-throughput, concurrent hash tables on GPUs by introducing a warp-cooperative, dynamically resizing open-addressing cuckoo design. It uses a cache-aligned 64-bit packed KV bucket layout, WABC for fast slot claiming, WCME for coordinated updates, and a load-aware warp-parallel linear hashing resizing strategy to avoid global rehashing. The approach delivers lock-free fast paths with bounded eviction costs, achieving sustained high occupancy and superior throughput compared with state-of-the-art GPU hash tables under balanced and mixed workloads. The results on an RTX 4090 show scalable performance with negligible resizing overhead, enabling efficient GPU-accelerated data processing; the authors also plan an open-source release for reproducibility.

Abstract

Hash tables are essential building blocks in data-intensive applications, yet existing GPU implementations often struggle with concurrent updates, high load factors, and irregular memory access patterns. We present Hive hash table, a high-performance, warp-cooperative and dynamically resizable GPU hash table that adapts to varying workloads without global rehashing. Hive hash table makes three key contributions. First, a cache-aligned packed bucket layout stores key-value pairs as 64-bit words, enabling coalesced memory access and atomic updates via single-CAS operations. Second, warp-synchronous concurrency protocols - Warp-Aggregated-Bitmask-Claim (WABC) and Warp-Cooperative Match-and-Elect (WCME) - reduce contention to one atomic operation per warp while ensuring lock-free progress. Third, a load-factor-aware dynamic resizing strategy expands or contracts capacity in warp-parallel K-bucket batches using linear hashing, maintaining balanced occupancy. To handle insertions under heavy contention, Hive hash table employs a four-step strategy: replace, claim-and-commit, bounded cuckoo eviction, and overflow-stash fallback. This design provides lock-free fast paths and bounded recovery cost under contention determined by a fixed eviction depth, while eliminating ABA hazards during concurrent updates. Experimental evaluation on an NVIDIA RTX 4090 shows Hive hash table sustains load factors up to 95% while delivering 1.5-2x higher throughput than state-of-the-art GPU hash tables (Slab-Hash, DyCuckoo, WarpCore) under mixed insert-delete-lookup workloads. On balanced workload, Hive hash table reaches 3.5 billion updates/s and nearly 4 billion lookups/s, demonstrating scalability and efficiency for GPU-accelerated data processing.
Paper Structure (23 sections, 1 theorem, 6 equations, 9 figures, 4 algorithms)

This paper contains 23 sections, 1 theorem, 6 equations, 9 figures, 4 algorithms.

Table of Contents

  1. Introduction
  2. Related Works
  3. Hive hash table
  4. Memory Layout and Bucket Design
  5. Hash Table Data Structure
  6. Hash Functions
  7. Supported Concurrent Operations
  8. Warp-Aggregated-Bitmask-Claim (WABC)
  9. Warp-Cooperative Match-and-Elect (WCME)
  10. Implementation Details
  11. Insert & Replace
  12. Lookup & Delete
  13. Dynamic Resizing Strategy
  14. Expansion (Split Phase)
  15. Contraction (Merge Phase)
  16. ...and 8 more sections

Key Result

Theorem 1

Let $n$ keys be mapped independently and uniformly at random into $m$ buckets. For bucket loads $L_b$ and total collisions $Y=\sum_{b=1}^{m}(L_b-1)_+$: For $n << m$ with $\lambda=n/m$, $L_b \sim \mathrm{Poisson}(\lambda)$, $\mathbb{E}[\text{empty}] \approx m e^{-\lambda}$, and $\mathbb{E}[Y]\approx n^2/(2m)$.

Figures (9)

  • Figure 1: Bucket Memory Layout
  • Figure 2: Hive hash table data structure.
  • Figure 3: Collision Speedup Ratio (CSR) of hash functions across varying key counts for $m=512^2$ buckets.
  • Figure 4: WABC Approach from Warp's perspective.
  • Figure 5: Insertion throughput comparison of lookup-based and computation-based hash function pairs in Hive hash table. Results show two-hash configurations consistently outperform three-hash alternatives.
  • ...and 4 more figures

Theorems & Definitions (1)

  • Theorem 1: Uniform hashing occupancy and collisions