Overview
Research addressed the application of Graphics Processing Units (GPUs) to Online Transaction Processing (OLTP) workloads, historically considered unsuitable for GPUs. While GPUs have demonstrated substantial potential in accelerating Online Analytical Processing (OLAP) operations, their use in OLTP has presented a gap. The study investigated the adaptation of concurrency control schemes, critical for OLTP performance, to GPU environments, given architectural differences from CPUs.
Research Context
GPUs offer massive parallelism and high memory bandwidth, suggesting their potential for concurrently processing numerous transactions. This capability positions them as candidates for OLTP acceleration. However, the performance of concurrency control schemes, which dictate OLTP system efficiency, may vary significantly due to architectural distinctions between GPUs and CPUs. The central research inquiry focused on evaluating how well CPU-oriented concurrency control schemes perform when adapted to GPU environments.
Approach
The researchers developed and utilized gCCTB, which is described as the first testbed specifically designed for evaluating concurrency control schemes on GPUs. This testbed facilitated the implementation and benchmarking of eight distinct concurrency control (CC) schemes. Among these, six were classic CPU-oriented schemes, while two were designed specifically for GPUs. The evaluation was conducted across two benchmark suites: YCSB and TPC-C. These benchmarks were executed under varying levels of contention and different GPU configurations. The study also examined the impact of GPU-specific parameters, such as the number of threads per warp and warps per block, on overall performance.
Findings
- GPU-optimized concurrency control schemes did not consistently outperform CPU-oriented schemes. This observation was particularly noted under specific workloads and contention levels.
- GPU-specific parameters, including thread count per warp and warps per block, significantly influenced performance. Careful tuning of these parameters was found to be necessary.
- Conflict resolution overhead was identified as a critical factor affecting the performance of CPU-oriented schemes when executed on GPUs.
- Optimistic concurrency control consistently minimized conflict resolution overhead. This scheme demonstrated superior performance compared to other CPU-oriented schemes across all evaluated workloads.
Why This Matters
The findings suggest that GPUs can be viable for OLTP workloads, but their effective utilization requires a nuanced understanding of concurrency control mechanisms. The identification of optimistic concurrency control as a high-performing CPU-oriented scheme on GPUs, and the critical role of GPU parameter tuning, provides specific directions for optimizing OLTP systems leveraging GPU architectures.