Abstract
In this letter we analyze the economic worth of GPU on analytical processing of GPU-accelerated database management systems (DBMSes). To this end, we conducted rigorous experiments with TPC-H across three popular GPU DBMSes. Consequently, we show that co-processing with CPU and GPU in the GPU DBMSes was cost-effective despite exposed concerns.
Introduction
In the era of big data, Graphical Processing Units (GPUs) are often presented as a silver bullet for accelerating complex analytical queries. However, GPUs represent a significant capital investment, and their memory capacity is often dwarfed by traditional system RAM. For database administrators and budget-conscious organizations, the critical question is not just whether GPUs are faster, but whether they provide a sufficient return on investment. This study conducts a holistic, empirical analysis of three prominent open-source GPU-accelerated DBMSes—BlazingSQL, OmniSci, and PG-Strom—to evaluate their "economic worth," defined as performance delivered per dollar spent on hardware.
Methodology
The researchers employed a rigorous methodology to ensure objective comparisons. They controlled for environmental noise by deactivating non-critical system daemons and utilized NVIDIA’s nvprof tool for fine-grained profiling. To quantify economic worth, the study introduced the MSK (MPS per $1,000) metric:
- MPS (Megabytes Per Second): A measure of data processing throughput.
- Cost (C): Total hardware cost (CPU, RAM, GPU, etc.) based on market prices.
- MSK = MPS / (Cost / $1,000). This metric represents the "bang-for-the-buck" score.
The evaluation used the TPC-H benchmark, scaling from Scale Factor (SF) 1 (~1.37 GB) up to SF 128 (~175 GB) to test how cost-effectiveness evolves as datasets exceed GPU memory limits.
Economic Worth and Scalability
The study revealed that higher hardware costs do not always translate to better economic value. For instance, using a newer RTX 2080 Ti often yielded lower cost-effectiveness than the older GTX 1080 Ti because the performance gains did not offset the price premium. Similarly, adding a second GPU often introduced communication overhead that decreased the overall MSK score at smaller scales.
A primary challenge for GPU DBMSes is the "scalability wall." Engines like OmniSci and BlazingSQL, which prioritize keeping data in GPU memory, see a sharp decline in cost-effectiveness once the dataset size forces frequent swapping with system RAM.
As illustrated in the figure above, while OmniSci maintains high cost-effectiveness at lower scales, its MSK drops significantly as SF increases. Conversely, PG-Strom, which functions as an extension to PostgreSQL, demonstrated superior scalability. Although it was less economical for small datasets, its performance remained stable across larger scales, eventually becoming the most economical choice for datasets exceeding SF 128.
Results and Recommendations
The optimal configuration for analytical processing is highly dependent on the scale of the database. At moderate scales (e.g., SF = 8), OmniSci with a single GTX 1080 Ti proved to be the most economical setup, outperforming CPU-only PostgreSQL by approximately 9x in MSK. However, as the data scales, the recommended architecture shifts.
The table above summarizes the most cost-effective configurations identified by the study. For datasets where SF ≤ 32, OmniSci-based configurations are preferred. For very large-scale processing (SF > 128), PG-Strom on a 1080 Ti emerges as the economic leader.
Conclusion
This experimental study demonstrates that GPU acceleration can be highly cost-effective, but its value is contingent upon the software architecture and the specific scale of the workload. While memory-centric GPU engines offer massive speedups for smaller datasets, more traditional co-processing extensions like PG-Strom provide better economic value for massive-scale analytics. Future work will explore a wider array of hardware, including enterprise-grade GPUs, and diverse analytical workloads to further refine these economic models.
Summary updated