As GPUs continue to grow in popularity for accelerating demanding applications, such as high-performance computing and machine learning, GPU architects need to deliver more powerful devices with updated instruction set architectures (ISAs) and new microarchitectural features. The introduction of the AMD RDNA architecture is one example where the GPU architecture was dramatically changed, modifying the underlying programming model, the core architecture, and the cache hierarchy. To date, no publicly-available simulator infrastructure can model the AMD RDNA GPU, preventing researchers from exploring new GPU designs based on the state-of-the-art RDNA architecture.

Graph Neural Networks (GNNs) have emerged as a promising class of Machine Learning algorithms to train on non-euclidean data. GNNs are widely used in recommender systems, drug discovery, text understanding, and traffic forecasting. Due to the energy efficiency and high-performance capabilities of GPUs, GPUs are a natural choice for accelerating the training of GNNs. Thus, we want to better understand the architectural and system-level implications of training GNNs on GPUs. Presently, there is no benchmark suite available designed to study GNN training workloads. In this work, we address this need by presenting GNNMark, a feature-rich benchmark suite that covers the diversity present in GNN training workloads, datasets, and GNN frameworks. Our benchmark suite consists of GNN workloads that utilize a variety of different graph-based data structures, including homogeneous graphs, dynamic graphs, and heterogeneous graphs commonly used in a number of application domains that we mentioned above. We use this benchmark suite to explore and characterize GNN training behavior on GPUs. We study a variety of aspects of GNN execution, including both compute and memory behavior, highlighting major bottlenecks observed during GNN training. At the system level, we study various aspects, including the scalability of training GNNs across a multi-GPU system, as well as the sparsity of data, encountered during training. The insights derived from our work can be leveraged by both hardware and software developers to improve both the hardware and software performance of GNN training on GPUs.

Programming on a GPU has been made considerably easier with the introduction of Virtual Memory features, which support common pointer-based semantics between the CPU and the GPU. However, supporting virtual memory on a GPU comes with some additional costs and overhead, with the largest being from the support for address translation. The fact that a massive number of threads run concurrently on a GPU means that the translation lookaside buffers (TLBs) are oversubscribed most of the time. Our investigation into a diverse set of GPU workloads shows that TLB misses can be extremely high (up to 99%), which inevitably leads to significant performance degradation due to long-latency page-table walks. Our profiling of TLB-sensitive workloads reveals a high degree of page sharing across the different cores of a GPU. In many applications, a page can be accessed in temporal proximity by multiple cores, following similar memory access patterns. To support the inherent sharing present in GPU workloads, we propose Valkyrie, an integrated cooperative TLB prefetching mechanism and an inter L1-TLB probing scheme that can efficiently reduce TLB bottlenecks in GPUs. Our evaluation using a diverse set of GPU workloads reveals that Valkyrie is able to achieve an average speedup of 1.95x, while adding modest hardware overhead.

As transistor scaling becomes increasingly more difficult to achieve, scaling the core count on a single GPU chip has also become extremely challenging. As the volume of data to process in today's increasingly parallel workloads continues to grow unbounded, we need to find scalable solutions that can keep up with this increasing demand. To meet the need of modern-day parallel applications, multi-GPU systems offer a promising path to deliver high performance and large memory capacity. However, multi-GPU systems suffer from performance issues associated with GPU-to-GPU communication and data sharing, which severely impact the benefits of multi-GPU systems. Programming multi-GPU systems has been made considerably simpler with the advent of Unified Memory which enables runtime migration of pages to the GPU on demand. Current multi-GPU systems rely on a first-touch Demand Paging scheme, where memory pages are migrated from the CPU to the GPU on the first GPU access to a page. The data sharing nature of GPU applications makes deploying an efficient programmer-transparent mechanism for inter-GPU page migration challenging. Therefore following the initial CPU-to-GPU page migration, the page is pinned on that GPU. Future accesses to this page from other GPUs happen at a cache-line granularity - pages are not transferred between GPUs without significant programmer intervention. We observe that this mechanism suffers from two major drawbacks: 1) imbalance in the page distribution across multiple GPUs, and 2) inability to move the page to the GPU that uses it most frequently. Both of these problems lead to load imbalance across GPUs, degrading the performance of the multi-GPU system. To address these problems, we propose Griffin, a holistic hardware-software solution to improve the performance of NUMA multi-GPU systems. Griffin introduces programmer-transparent modifications to both the IOMMU and GPU architecture, supporting efficient runtime page migration based on locality information. In particular, Griffin employs a novel mechanism to detect and move pages at runtime between GPUs, increasing the frequency of resolving accesses locally, which in turn improves the performance. To ensure better load balancing across GPUs, Griffin employs a Delayed First-Touch Migration policy that ensures pages are evenly distributed across multiple GPUs. Our results on a diverse set of multi-GPU workloads show that Griffin can achieve up to a 2.9× speedup on a multi-GPU system, while incurring low implementation overhead.

The rapidly growing popularity and scale of data-parallel workloads demand a corresponding increase in raw computational power of Graphics Processing Units (GPUs). As single-GPU platforms struggle to satisfy these performance demands, multi-GPU platforms have started to dominate the high-performance computing world. The advent of such systems raises a number of design challenges, including the GPU microarchitecture, multi-GPU interconnect fabric, runtime libraries, and associated programming models. The research community currently lacks a publicly available and comprehensive multi-GPU simulation framework to evaluate next-generation multi-GPU system designs.

An emerging class of applications attempt to make use of both the CPU and GPU in a heterogeneous system. The peak performance for these applications is achieved when both the CPU and GPU are used collaboratively. However, along with this increased gain in performance, power and energy management is a larger challenge. In this paper we address the issue of executing applications that utilize both the CPU and GPU in an energy efficient way. Towards this end, we propose a power management framework named Airavat that tunes the CPU, GPU and memory frequencies, synergestically, in order to improve the energy efficiency of collaborative CPU-GPU applications. Airavat uses machine learning-based prediction models, combined with feedback based Dynamic Voltage and Frequency Scaling to improve the energy efficiency of such applications. We demonstrate our framework on the NVIDIA Jetson TX1 and observe an improvement in terms of Energy Delay Product (EDP) by 24% with negligible performance loss.

GPUs have been shown to deliver impressive computing performance, while also providing high energy efficiency, across a wide range of high-performance and embedded system workloads. However, limited support for efficient communication and synchronization between the CPU and the GPU impacts our ability to fully exploit the benefits of heterogeneous systems. Recently, the Heterogeneous System Architecture (HSA) was introduced to address these issues with synchronization and communication, but given the low-level nature of HSA, it was not easily adopted by the broader programming community. In 2016, AMD described the Radeon Open Compute (ROC) platform that brings high-level programming frameworks such as OpenCL, HC++, and HIP to end users. These high-level programming frameworks offer a simpler programming experience by wrapping complex HSA APIs, while still delivering the power of HSA. To date, there has been little evaluation of the potential performance benefits and trade-offs of leveraging the ROC platform. In this work, we evaluate the performance of the ROC platform using the Hetero-Mark and DNNMark benchmark suites. Equipped with Hetero-Mark, we compare the performance of different programming frameworks, including OpenCL, HC++, and HIP on both integrated APUs and discrete GPUs. We also present three new CPU-GPU collaborative patterns and employ three new benchmarks to evaluate system-level atomics. With DNNMark and a new DNN Face Detection benchmark, we evaluate the performance of ROC libraries including rocBLAS and MIOpen. We also provide guidance on best practices to programmers when developing applications leveraging the ROC platform.

GPUs have become a very popular platform for accelerating the processing involved in deep learning applications. One class of popular variants, Convolutional Neural Networks (CNNs), have been widely deployed to run on GPUs. In many application settings, a GPU has sufficient computing power and memory space to accommodate the dense matrix operations performed during CNN training. However, few characterization studies have considered how CNNs can impact microarchitectural structures in a GPU. In this paper, we perform a characterization of one selected CNN workload as run on two different NVIDIA GPUs from distinct microarchitecture families, highlighting the impact that microarchitecture plays on this important class of workload. First, we analyze the performance implications of a CNN model using microarchitectural details on a layer-by-layer basis, and characterize the memory access behavior in the context of a typical GPU memory hierarchy, considering hardware resource utilization associated with each primitive in the CNN model. We identify major bottlenecks by considering the potential limits of using a single GPU. Additionally, we evaluate a number of optimization approaches, such as L1 cache bypassing and kernel fusion. L1 cache bypassing can achieve up to a 6.2% speedup for a single layer, but manipulating L1 cache provides very limited benefits in terms of application speedup, while kernel fusion provides an overall application speedup of 4.0%, on average.