BaM: A Case for Enabling Fine-grain High Throughput GPU-Orchestrated Access to Storage

Accelerators like Graphics Processing Units (GPUs) have been increasingly deployed in modern data centers because of their compute capabilities and memory bandwidth. These accelerators have traditionally relied on the "application host code" and the OS running on the CPU to orchestrate their access to the data storage devices. CPU orchestration of storage data accesses works well for classic GPU applications, like dense neural network training, where data access patterns are predefined, regular, dense, and independent of the data values, enabling the CPU to partition the storage data into coarse-grain chunks and coordinate the storage device accesses and data transfers to the accelerators. Unfortunately, such a CPU-centric strategy causes excessive CPU-GPU synchronization overhead and/or I/O traffic amplification, diminishing the effective storage bandwidth for emerging applications with fine-grain data-dependent access patterns like graph and data analytics, recommender systems, and graph neural networks. In this work, we make a case for enabling GPUs to orchestrate high-throughput, fine-grain accesses into NVMe Solid State Drives (SSDs) in a new system architecture called BaM. BaM mitigates the I/O traffic amplification by enabling the GPU threads to read or write small amounts of data on-demand, as determined by the compute. We show that (1) the BaM infrastructure software running on GPUs can identify and communicate the fine-grain accesses at a sufficiently high rate to fully utilize the underlying storage devices, (2) even with consumer-grade SSDs, a BaM system can support application performance that is competitive against a much more expensive DRAM-only solution, and (3) the reduction in I/O amplification can yield significant performance benefit.