Introduction
Modern GPU architectures, such as the NVIDIA A100, provide unprecedented compute and memory bandwidth for data analytics. Features like Multi-Instance GPU (MIG) allow for hardware partitioning, enabling isolated execution of concurrent tasks. However, database administrators face a significant "right-sizing" challenge: determining the optimal GPU resource allocation for specific queries. Over-provisioning wastes expensive hardware cycles, while under-provisioning leads to missed performance targets. This research presents a cross-stack analysis of five GPU database systems to replace manual trial-and-error with a formal modeling framework.
Methodology
The study evaluates five prominent GPU database systems: Crystal, HeavyDB, BlazingSQL, TQP, and PG-Strom. Using profiling tools like NVIDIA Nsight, the researchers tracked micro-architectural metrics including integer operations, DRAM throughput, and cache hit rates.
The core contribution is a dual-modeling approach:
- Crystal-Opt (White-Box): An analytical model for the Crystal system that incorporates real-world utilization factors for DRAM and caches.
- Extended Roofline Model (Black-Box): An adaptation of the classic Roofline Model that introduces L2 cache bandwidth as a distinct bottleneck. This extension is crucial for highly optimized databases where small working sets often saturate L2 bandwidth before DRAM.
Experiments were conducted using the Star Schema Benchmark (SSB) with Scale Factors (SF) up to 16, representing the memory limits of current GPU database implementations.
Key Insights: Bottlenecks and Scaling
A primary finding is that performance does not scale linearly with hardware resources. As shown in the figure below, increasing the GPU resource fraction yields diminishing returns, particularly for memory-bound queries.
The L2 Cache and Kernel Fusion
Specialized systems like Crystal and HeavyDB leverage kernel fusion to combine multiple database operations into a single GPU kernel. This minimizes intermediate result writes to memory. However, this optimization shifts the bottleneck from DRAM to the L2 cache. Queries with hash tables that fit within the cache saturate L2 bandwidth long before hitting DRAM limits. Conversely, general-purpose systems like TQP and BlazingSQL suffer from high overhead due to non-fused operations and frequent intermediate data movement.
Performance Comparison and Model Accuracy
The evaluation reveals a wide performance gap between specialized prototypes and general-purpose frameworks. Crystal remains the performance leader, while HeavyDB offers a competitive balance between speed and SQL support. PG-Strom exhibits the highest latency, primarily due to CPU-GPU data transfer overhead.
The proposed extended roofline model significantly outperforms naive linear scaling. It achieved a median error of 22.73%, whereas the linear baseline reached errors exceeding 139% in some scenarios. By applying this framework to optimize concurrency via MIG, the authors achieved throughput gains of up to 6.43× for smaller datasets.
Conclusion and Future Directions
While the extended roofline model provides a robust foundation for resource allocation, it currently requires an initial "warm" profiling run. Future research will focus on mitigating "noisy neighbor" interference in shared cache environments and validating the model's generalizability across newer GPU architectures like NVIDIA Hopper and Blackwell, as well as hardware from other vendors.
Summary updated