Subscribe to the gold package and get unlimited access to Shamra Academy
Register a new userIrregular Accesses Reorder Unit: Improving GPGPU Memory Coalescing for Graph-Based Workloads
68
0
0.0
(
0
)
Added by
Albert Segura
Publication date
2020
fields
Informatics Engineering
and research's language is
English
Ask ChatGPT about the research
No Arabic abstract
GPGPU architectures have become established as the dominant parallelization and performance platform achieving exceptional popularization and empowering domains such as regular algebra, machine learning, image detection and self-driving cars. However, irregular applications struggle to fully realize GPGPU performance as a result of control flow divergence and memory divergence due to irregular memory access patterns. To ameliorate these issues, programmers are obligated to carefully consider architecture features and devote significant efforts to modify the algorithms with complex optimization techniques, which shift programmers priorities yet struggle to quell the shortcomings. We show that in graph-based GPGPU irregular applications these inefficiencies prevail, yet we find that it is possible to relax the strict relationship between thread and data processed to empower new optimizations. Based on this key idea, we propose the Irregular accesses Reorder Unit (IRU), a novel hardware extension tightly integrated in the GPGPU pipeline. The IRU reorders data processed by the threads on irregular accesses which significantly improves memory coalescing, and allows increased performance and energy efficiency. Additionally, the IRU is capable of filtering and merging duplicated irregular access which further improves graph-based irregular applications. Programmers can easily utilize the IRU with a simple API, or compiler optimized generated code with the extended ISA instructions provided. We evaluate our proposal for state-of-the-art graph-based algorithms and a wide selection of applications. Results show that the IRU achieves a memory coalescing improvement of 1.32x and a 46% reduction in the overall traffic in the memory hierarchy, which results in 1.33x and 13% improvement in performance and energy savings respectively, while incurring in a small 5.6% area overhead.
rate research
Read More
Over the past two decades, the storage capacity and access bandwidth of main memory have improved tremendously, by 128x and 20x, respectively. These improvements are mainly due to the continuous technology scaling of DRAM (dynamic random-access memory), which has been used as the physical substrate for main memory. In stark contrast with capacity and bandwidth, DRAM latency has remained almost constant, reducing by only 1.3x in the same time frame. Therefore, long DRAM latency continues to be a critical performance bottleneck in modern systems. Increasing core counts, and the emergence of increasingly more data-intensive and latency-critical applications further stress the importance of providing low-latency memory access. In this dissertation, we identify three main problems that contribute significantly to long latency of DRAM accesses. To address these problems, we present a series of new techniques. Our new techniques significantly improve both system performance and energy efficiency. We also examine the critical relationship between supply voltage and latency in modern DRAM chips and develop new mechanisms that exploit this voltage-latency trade-off to improve energy efficiency. The key conclusion of this dissertation is that augmenting DRAM architecture with simple and low-cost features, and developing a better understanding of manufactured DRAM chips together lead to significant memory latency reduction as well as energy efficiency improvement. We hope and believe that the proposed architectural techniques and the detailed experimental data and observations on real commodity DRAM chips presented in this dissertation will enable development of other new mechanisms to improve the performance, energy efficiency, or reliability of future memory systems.
GPUs offer orders-of-magnitude higher memory bandwidth than traditional CPU-only systems. However, GPU device memory tends to be relatively small and the memory capacity can not be increased by the user. This paper describes Buddy Compression, a scheme to increase both the effective GPU memory capacity and bandwidth while avoiding the downsides of conventional memory-expanding strategies. Buddy Compression compresses GPU memory, splitting each compressed memory entry between high-speed device memory and a slower-but-larger disaggregated memory pool (or system memory). Highly-compressible memory entries can thus be accessed completely from device memory, while incompressible entries source their data using both on and off-device accesses. Increasing the effective GPU memory capacity enables us to run larger-memory-footprint HPC workloads and larger batch-sizes or models for DL workloads than current memory capacities would allow. We show that our solution achieves an average compression ratio of 2.2x on HPC workloads and 1.5x on DL workloads, with a slowdown of just 1~2%.
In recent years, machine intelligence (MI) applications have emerged as a major driver for the computing industry. Optimizing these workloads is important but complicated. As memory demands grow and data movement overheads increasingly limit performance, determining the best GPU caching policy to use for a diverse range of MI workloads represents one important challenge. To study this, we evaluate 17 MI applications and characterize their behaviors using a range of GPU caching strategies. In our evaluations, we find that the choice of caching policy in GPU caches involves multiple performance trade-offs and interactions, and there is no one-size-fits-all GPU caching policy for MI workloads. Based on detailed simulation results, we motivate and evaluate a set of cache optimizations that consistently match the performance of the best static GPU caching policies.
With the increasing adoption of graph neural networks (GNNs) in the machine learning community, GPUs have become an essential tool to accelerate GNN training. However, training GNNs on very large graphs that do not fit in GPU memory is still a challenging task. Unlike conventional neural networks, mini-batching input samples in GNNs requires complicated tasks such as traversing neighboring nodes and gathering their feature values. While this process accounts for a significant portion of the training time, we find existing GNN implementations using popular deep neural network (DNN) libraries such as PyTorch are limited to a CPU-centric approach for the entire data preparation step. This all-in-CPU approach has negative impact on the overall GNN training performance as it over-utilizes CPU resources and hinders GPU acceleration of GNN training. To overcome such limitations, we introduce PyTorch-Direct, which enables a GPU-centric data accessing paradigm for GNN training. In PyTorch-Direct, GPUs are capable of efficiently accessing complicated data structures in host memory directly without CPU intervention. Our microbenchmark and end-to-end GNN training results show that PyTorch-Direct reduces data transfer time by 47.1% on average and speeds up GNN training by up to 1.6x. Furthermore, by reducing CPU utilization, PyTorch-Direct also saves system power by 12.4% to 17.5% during training. To minimize programmer effort, we introduce a new unified tensor type along with necessary changes to the PyTorch memory allocator, dispatch logic, and placement rules. As a result, users need to change at most two lines of their PyTorch GNN training code for each tensor object to take advantage of PyTorch-Direct.
Even with generational improvements in DRAM technology, memory access latency still remains the major bottleneck for application accelerators, primarily due to limitations in memory interface IPs which cannot fully account for variations in target applications, the algorithms used, and accelerator architectures. Since developing memory controllers for different applications is time-consuming, this paper introduces a modular and programmable memory controller that can be configured for different target applications on available hardware resources. The proposed memory controller efficiently supports cache-line accesses along with bulk memory transfers. The user can configure the controller depending on the available logic resources on the FPGA, memory access pattern, and external memory specifications. The modular design supports various memory access optimization techniques including, request scheduling, internal caching, and direct memory access. These techniques contribute to reducing the overall latency while maintaining high sustained bandwidth. We implement the system on a state-of-the-art FPGA and evaluate its performance using two widely studied domains: graph analytics and deep learning workloads. We show improved overall memory access time up to 58% on CNN and GCN workloads compared with commercial memory controller IPs.
Log in to be able to interact and post comments
comments
Fetching comments
Sign in to be able to follow your search criteria
