GPU Storage Approach Targets Big Data Bottlenecks
An emerging storage technology aims to leverage faster GPUs by creating a direct path between local and remote storage, thereby overcoming I/O bottlenecks that are slowing the crunching of AI and HPC data sets.
Nvidia (NASDAQ: NVDA) is touting the potential bottleneck-breaker called GPUDirect Storage as a way to keep its graphics processors humming, thereby boosting overall application performance.
GPUDirect Storage and an Nvidia-customized approach called GPUDirect RDMA improved bandwidth and latency by avoiding extra copies using a “bounce buffer” in CPU memory. The approach enables direct memory access (DMA) nearer to storage “to move data on a direct path into or out of GPU memory – all without burdening the CPU or GPU,” Nvidia engineers explained this week in a blog post.
Direct memory access uses a copy engine to move large blocks of data over a PCI Express interface that connect NVMe storage as well as GPUs and CPUs. That approach offloads processing, allowing GPUs to crunch big data. “If a DMA engine in an NVMe drive or elsewhere near storage can be used to move data instead of the GPU’s DMA engine, then there’s no interference in the path between the CPU and GPU,” Nvidia noted.
The company is betting the GPU storage approach will gain traction for data analytics, deep learning and graph analytics applications handling ever-larger data sets that must be pulled in from storage. For applications such as training neural networks, in which many file sets are accessed and read multiple times, “Optimization of data transfer to [the] GPU, in this case, could have a significant and beneficial impact on the total time to train an AI model,” Nvidia said.
The GPU leader indicated it is nearing deployment of GPUDirect Storage feature, including new set of APIs that will be added to its CUDA developer platform. Those APIs would support the shift to the emerging storage feature along with native integration into Nvidia’s RAPIDS data science library.