GPU-Initiated On-Demand High-Throughput Storage Access in the BaM System Architecture

GPU Graph Processing on CXL-Based Microsecond-Latency External Memory

“…As mentioned in Section 1, the host DRAM-based method EMOGI [31] is generally faster than the SSD-based method BaM [33] in terms of graph processing time. If EMOGI's runtime includes the time for loading graph data onto the host DRAM from the SSDs (data loading time) in addition to the time for running the algorithm on the GPU (graph processing time), BaM is shown to be competitive for some benchmark workloads.…”

“…With 16 drives, the system well supports the required random read speed of 93.75 MIOPS. To evaluate BaM, we replace XLFDDs with NVMe SSDs that collectively offer 6-MIOPS random read performance to match the number used in [33]. As with BaM, we place submission queues (SQs) and data buffers in the base address register (BAR) section of the GPU memory in order to control storage devices directly from the GPU.…”

“…This paper has shown that EMOGI stays as performant even if the external memory latency is longer than the host DRAM, up to a few microseconds. GPU graph processing on storage: BaM introduced a first GPUcentric storage access method that does not involve CPU intervention [33]. While there are several prior works in GPU-centric approaches [28,37,39,40,43], they rely on the CPU to handle storage access and use the GPU memory as a staging buffer for their data transfer.…”

“…In order to handle ever-growing data sizes in these applications beyond the relatively limited capacity (tens of GBs) of GPU onboard memory, the use of external memory such as the host DRAM and solid-state drives (SSDs) can be a cost-effective approach compared with pooling multiple GPUs' memory together [9-11, 18, 22, 28, 31, 33, 37, 39, 40, 43]. In particular, GPU-centric external memory access methods have been shown to yield the stateof-the-art runtime performance in workloads involving on-demand, fine-grained random access such as graph analytics [31,33]. That is, when small pieces of data to be read next depend on the current processing results and cannot be a priori determined, it is more efficient to have the GPU initiate data requests than to have the CPU control the data flow between the GPU and external memory.…”