GPUs rely on a largely synchronous programming and execution model. With increasing need to access data residing on SSDs, GPU threads can incur significant latencies for such accesses when using blocking/synchronous I/O mechanisms. There is little hardware/systems support today to perform non-blocking/asynchronous operations from GPU threads directly to tolerate microsecond level latencies incurred in SSD accesses. To fix this dichotomy, this paper presents the design, implementation and evaluation of AGIO, which provides APIs and a runtime environment for GPU threads to directly perform asynchronous I/O operations (fully GPU-orchestrated without CPU involvement). AGIO decouples, both in time and space, the I/O initiation from its completion, to allow useful computation in-between, in order to hide much of the I/O latency. This is particularly useful in applications with access patterns known at compile time, where similar to prefetching, AGIO I/Os can be introduced ahead of need, to yield 65% better performance than its synchronous counterpart. A non-intuitive benefit of AGIO, particularly in applications with data dependent accesses, is the ability to allow threads to proceed beyond I/O initiation, towards initiating more I/O, even if there is little compute to overlap. Such pro-active I/O issuance increases the I/O parallelism to more fully utilize underlying bandwidths, yielding 32% better performance than the synchronous alternative in data dependent executions. Decoupling initiation from completion makes AGIO more adaptive to dataset characteristics, with the programmer not needing a priori knowledge of inputs for effective performance. We also show that AGIO can meet (or better) the performance using a GPU with fewer than half the compute engines of its synchronous counterparts.

In 1987, Jim Gray and Gianfranco Putzolu introduced the five-minute rule, a simple, storage-memory-economics-based heuristic for deciding when data should live in DRAM rather than on storage. Subsequent revisits to the rule largely retained that economics-only view, leaving host costs, feasibility limits, and workload behavior out of scope. This paper revisits the rule from first principles, integrating host costs, DRAM bandwidth/capacity, and physics-grounded models of SSD performance and cost, and then embedding these elements in a constraint- and workload-aware framework that yields actionable provisioning guidance. We show that, for modern AI platforms, especially GPU-centric hosts paired with ultra-high-IOPS SSDs engineered for fine-grained random access, the DRAM$\leftrightarrow$flash caching threshold collapses from minutes to a few seconds. This shift reframes NAND flash memory as an \emph{active data tier} and exposes a broad research space across the hardware-software stack. We further introduce MQSim-Next, a calibrated SSD simulator that supports validation and sensitivity analysis and facilitates future architectural and system research. Finally, we present two concrete case studies that showcase the software system design space opened by such memory hierarchy paradigm shift. Overall, we turn a classical heuristic into an actionable, feasibility-aware analysis and provisioning framework and set the stage for further research on AI-era memory hierarchy.

The growth rate of the GPU memory capacity has not been able to keep up with that of the size of large language models (LLMs), hindering the model training process. In particular, activations-the intermediate tensors produced during forward propagation and reused in backward propagation-dominate the GPU memory use. This leads to high training overheads such as expensive weight update costs due to the small micro-batch size. To address this challenge, we propose SSDTrain, an adaptive activation offloading framework to high-capacity NVMe SSDs. SSDTrain reduces GPU memory usage without impacting performance by fully overlapping data transfers with computation. SSDTrain is compatible with popular deep learning frameworks like PyTorch, Megatron, and DeepSpeed, and it employs techniques such as tensor deduplication and forwarding to further enhance efficiency. We extensively experimented with popular LLMs like GPT, BERT, and T5. Results demonstrate that SSDTrain reduces 47% of the activation peak memory usage. At the same time, SSDTrain perfectly overlaps the I/O with the computation and incurs negligible overhead. Compared with keeping activations in GPU memory and layerwise full recomputation, SSDTrain achieves the best memory savings with negligible throughput loss. We further analyze how the reduced activation memory use may be leveraged to increase throughput by increasing micro-batch size and reducing pipeline parallelism bubbles.

As the demand for processing larger datasets increases, GPUs need to reach deeper into their (memory) hierarchy to directly access capacities that only storage systems (SSDs) can hold. However, the state-of-the-art mechanisms to reach storage either employ software stacks running on the host CPUs as intermediaries (e.g. Dragon, HMM), which has been noted to perform poorly and not able to meet the throughput needs of GPU cores, or directly access SSDs through NVMe queues (BaM) which does not benefit from lower latencies that may be possible by having the host memory as an intermediate tier. This paper presents the design and implementation of GPU Memory Tiering (GMT) by implementing a GPU-orchestrated 3-tier hierarchy comprising GPU memory, host memory and SSDs, where the GPU orchestrates most of the transfers that are bandwidth/latency sensitive. Additionally, it is important to not blindly transfer pages from the GPU memory to host memory upon an eviction, and GMT employs a reuse-prediction based practical insertion policy to perform discretionary page placement/bypass. An implementation and evaluation on an actual platform demonstrates that GMT performs 50% better than the state-of-the-art 2-tier strategy (BaM) and over 350% better than the state-of-the-art 3-tier strategy that is orchestrated by host CPUs (HMM), over a number of GPU applications with diverse memory access characteristics.

Graph Neural Networks (GNNs) are emerging as a powerful tool for learning from graph-structured data and performing sophisticated inference tasks in various application domains. Although GNNs have been shown to be effective on modest-sized graphs, training them on large-scale graphs remains a significant challenge due to the lack of efficient storage access and caching methods for graph data. Existing frameworks for training GNNs use CPUs for graph sampling and feature aggregation, while the training and updating of model weights are executed on GPUs. However, our in-depth profiling shows CPUs cannot achieve the graph sampling and feature aggregation throughput required to keep up with GPUs. Furthermore, when the graph and its embeddings do not fit in the CPU memory, the overhead introduced by the operating system, say for handling page-faults, causes gross under-utilization of hardware and prolonged end-to-end execution time. To address these issues, we propose the GPU Initiated Direct Storage Access (GIDS) dataloader, to enable GPU-oriented GNN training for large-scale graphs while efficiently utilizing all hardware resources, such as CPU memory, storage, and GPU memory. The GIDS dataloader first addresses memory capacity constraints by enabling GPU threads to directly fetch feature vectors from storage. Then, we introduce a set of innovative solutions, including the dynamic storage access accumulator, constant CPU buffer, and GPU software cache with window buffering, to balance resource utilization across the entire system for improved end-to-end training throughput. Our evaluation using a single GPU on terabyte-scale GNN datasets shows that the GIDS dataloader accelerates the overall DGL GNN training pipeline by up to 582× when compared to the current, state-of-the-art DGL dataloader.

Graphics Processing Units (GPUs) have traditionally relied on the host CPU to initiate access to the data storage. This approach is well-suited for GPU applications with known data access patterns that enable partitioning of their dataset to be processed in a pipelined fashion in the GPU. However, emerging applications such as graph and data analytics, recommender systems, or graph neural networks, require fine-grained, data-dependent access to storage. CPU initiation of storage access is unsuitable for these applications due to high CPU-GPU synchronization overheads, I/O traffic amplification, and long CPU processing latencies. GPU-initiated storage removes these overheads from the storage control path and, thus, can potentially support these applications at much higher speed. However, there is a lack of systems architecture and software stack that enable efficient GPU-initiated storage access. This work presents a novel system architecture, BaM, that fills this gap. BaM features a fine-grained software cache to coalesce data storage requests while minimizing I/O traffic amplification. This software cache communicates with the storage system via high-throughput queues that enable the massive number of concurrent threads in modern GPUs to make I/O requests at a high rate to fully utilize the storage devices and the system interconnect. Experimental results show that BaM delivers 1.0x and 1.49x end-to-end speed up for BFS and CC graph analytics benchmarks while reducing hardware costs by up to 21.7x over accessing the graph data from the host memory. Furthermore, BaM speeds up data-analytics workloads by 5.3x over CPU-initiated storage access on the same hardware.

Solid State Drives (SSDs) have become increasingly popular due to their superior access latency and bandwidth compared to Hard Disk Drives (HDDs). However, to fully understand the impact of SSD design and microarchitecture on end-to-end application performance, researchers need to move beyond treating SSDs as black-box components. Unfortunately, purchasing multiple SSDs for research is expensive and ineffective since the underlying microarchitecture is still unknown to the system designer. While simulators have become the most popular method for studying SSDs, existing software-based simulators lack real data transfers and cannot simulate the latency from NVMe and PCIe interfaces. Additionally, simulating the entire SSD using software codes is time-consuming and limits the number of experiments that can be run in a reasonable amount of time. To address these issues, we present FSSD, an FPGA-based emulation system that models the latency and access patterns of an actual NVMe SSD. FSSD takes advantage of the flexibility of an FPGA, enabling users to customize SSD microarchitecture features and explore the design space for data-intensive applications. FSSD can be interacting with real operating systems, instead of relying on Virtual Machines like most other software simulators do. Evaluations show that FSSD provides over 1000x speedup compared to software-based simulation using the SimpleSSD simulator. The ability to customize SSD parameters and emulate NAND latency with high precision makes FSSD a valuable platform for SSD research and development. FSSD is also open-sourced to benefit the research community.

Graph neural networks (GNNs) have shown high potential for a variety of real-world, challenging applications, but one of the major obstacles in GNN research is the lack of large-scale flexible datasets. Most existing public datasets for GNNs are relatively small, which limits the ability of GNNs to generalize to unseen data. The few existing large-scale graph datasets provide very limited labeled data. This makes it difficult to determine if the GNN model's low accuracy for unseen data is inherently due to insufficient training data or if the model failed to generalize. Additionally, datasets used to train GNNs need to offer flexibility to enable a thorough study of the impact of various factors while training GNN models.

Modern analytics and recommendation systems are increasingly based on graph data that capture the relations between entities being analyzed. Practical graphs come in huge sizes, offer massive parallelism, and are stored in sparse-matrix formats such as compressed sparse row (CSR). To exploit the massive parallelism, developers are increasingly interested in using GPUs for graph traversal. However, due to their sizes, graphs often do not fit into the GPU memory. Prior works have either used input data pre-processing/partitioning or unified virtual memory (UVM) to migrate chunks of data from the host memory to the GPU memory. However, the large, multi-dimensional, and sparse nature of graph data presents a major challenge to these schemes and results in significant amplification of data movement and reduced effective data throughput. In this work, we propose EMOGI, an alternative approach to traverse graphs that do not fit in GPU memory using direct cache-line-sized access to data stored in host memory. This paper addresses the open question of whether a sufficiently large number of overlapping cache-line-sized accesses can be sustained to 1) tolerate the long latency to host memory, 2) fully utilize the available bandwidth, and 3) achieve favorable execution performance. We analyze the data access patterns of several graph traversal applications in GPU over PCIe using an FPGA to understand the cause of poor external bandwidth utilization. By carefully coalescing and aligning external memory requests, we show that we can minimize the number of PCIe transactions and nearly fully utilize the PCIe bandwidth with direct cache-line accesses to the host memory. EMOGI achieves 2.60X speedup on average compared to the optimized UVM implementations in various graph traversal applications. We also show that EMOGI scales better than a UVM-based solution when the system uses higher bandwidth interconnects such as PCIe 4.0.

Using flash-based solid state drives (SSDs) as main memory has been proposed as a practical solution towards scaling memory capacity for data-intensive applications. However, almost all existing approaches rely on the paging mechanism to move data between SSDs and host DRAM. This inevitably incurs significant performance overhead and extra I/O traffic. Thanks to the byte-addressability supported by the PCIe interconnect and the internal memory in SSD controllers, it is feasible to access SSDs in both byte and block granularity today. Exploiting the benefits of SSD's byte-accessibility in today's memory-storage hierarchy is, however, challenging as it lacks systems support and abstractions for programs. In this paper, we present FlatFlash, an optimized unified memory-storage hierarchy, to efficiently use byte-addressable SSD as part of the main memory. We extend the virtual memory management to provide a unified memory interface so that programs can access data across SSD and DRAM in byte granularity seamlessly. We propose a lightweight, adaptive page promotion mechanism between SSD and DRAM to gain benefits from both the byte-addressable large SSD and fast DRAM concurrently and transparently, while avoiding unnecessary page movements. Furthermore, we propose an abstraction of byte-granular data persistence to exploit the persistence nature of SSDs, upon which we rethink the design primitives of crash consistency of several representative software systems that require data persistence, such as file systems and databases. Our evaluation with a variety of applications demonstrates that, compared to the current unified memory-storage systems, FlatFlash improves the performance for memory-intensive applications by up to 2.3x, reduces the tail latency for latency-critical applications by up to 2.8x, scales the throughput for transactional database by up to 3.0x, and decreases the meta-data persistence overhead for file systems by up to 18.9x. FlatFlash also improves the cost-effectiveness by up to 3.8x compared to DRAM-only systems, while enhancing the SSD lifetime significantly.

Recent advancements in deep learning techniques facilitate intelligent-query support in diverse applications, such as content-based image retrieval and audio texturing. Unlike conventional key-based queries, these intelligent queries lack efficient indexing and require complex compute operations for feature matching. To achieve high-performance intelligent querying against massive datasets, modern computing systems employ GPUs in-conjunction with solid-state drives (SSDs) for fast data access and parallel data processing. However, our characterization with various intelligent-query workloads developed with deep neural networks (DNNs), shows that the storage I/O bandwidth is still the major bottleneck that contributes 56%--90% of the query execution time.