Memory Access Latency Research Articles

Ferroelectric Devices for Content-Addressable Memory.

In-memory computing is an attractive solution for reducing power consumption and memory access latency cost by performing certain computations directly in memory without reading operands and sending them to arithmetic logic units. Content-addressable memory (CAM) is an ideal way to smooth out the distinction between storage and processing, since each memory cell is a processing unit. CAM compares the search input with a table of stored data and returns the matched data address. The issues of constructing binary and ternary content-addressable memory (CAM and TCAM) based on ferroelectric devices are considered. A review of ferroelectric materials and devices is carried out, including on ferroelectric transistors (FeFET), ferroelectric tunnel diodes (FTJ), and ferroelectric memristors.

Towards Efficient Sparse Matrix Vector Multiplication on Real Processing-In-Memory Architectures

Several manufacturers have already started to commercialize near-bank Processing-In-Memory (PIM) architectures, after decades of research efforts. Near-bank PIM architectures place simple cores close to DRAM banks. Recent research demonstrates that they can yield significant performance and energy improvements in parallel applications by alleviating data access costs. Real PIM systems can provide high levels of parallelism, large aggregate memory bandwidth and low memory access latency, thereby being a good fit to accelerate the Sparse Matrix Vector Multiplication (SpMV) kernel. SpMV has been characterized as one of the most significant and thoroughly studied scientific computation kernels. It is primarily a memory-bound kernel with intensive memory accesses due its algorithmic nature, the compressed matrix format used, and the sparsity patterns of the input matrices given. This paper provides the first comprehensive analysis of SpMV on a real-world PIM architecture, and presents SparseP, the first SpMV library for real PIM architectures. We make two key contributions. First, we design efficient SpMV algorithms to accelerate the SpMV kernel in current and future PIM systems, while covering a wide variety of sparse matrices with diverse sparsity patterns. Second, we provide the first comprehensive analysis of SpMV on a real PIM architecture. Specifically, we conduct our rigorous experimental analysis of SpMV kernels in the UPMEM PIM system, the first publicly-available real-world PIM architecture. Our extensive evaluation provides new insights and recommendations for software designers and hardware architects to efficiently accelerate the SpMV kernel on real PIM systems. For more information about our thorough characterization on the SpMV PIM execution, results, insights and the open-source SparseP software package [21], we refer the reader to the full version of the paper [3, 4]. The SparseP software package is publicly and freely available at https://github.com/CMU-SAFARI/SparseP.

VMT: Virtualized Multi-Threading for Accelerating Graph Workloads on Commodity Processors

Modern-day graph workloads operate on huge graphs through pointer chasing which leads to high last-level cache (LLC) miss rates and limited memory-level parallelism (MLP). Simultaneous Multi-Threading (SMT) effectively hides the memory access latencies for multi-threaded graph workloads provided that sufficient threads are supported in hardware. Unfortunately, providing a sufficiently large number of physical threads incurs an unjustifiably high hardware cost for commodity SMT processors which typically implement only two physical hardware threads. Ideally, we would like to achieve aggressive-SMT performance when running graph workloads on modest commodity processors. In this paper, we propose Virtualized Multi-Threading (VMT), a low-overhead multi-threading paradigm for accelerating graph workloads on commodity processors. Unlike prior multi-threading paradigms, VMT virtualizes both the physical hardware threads and the architecture state: VMT maps a large number of logical software threads to a small number of physical hardware threads, while maintaining the architecture state of the logical threads in the processor's cache hierarchy. Implemented on top of a quad-core 2-way SMT processor, VMT achieves an average speedup of 1.74× for a set of representative graph workloads, while incurring minimal hardware cost (195 bytes per core to support up to 32 logical threads). VMT's low hardware cost paves the way for implementation in commodity processors.

Interconnect support for energy efficient and high bandwidth memory access in CMPs

The increasing number of on-chip cores and a limited number of memory controllers pose a critical problem for off-chip memory bandwidth. In this work, we propose an adaptive hybrid switching strategy with a dual crossbar router to provide low latency paths between cache and memory controllers. The performance is further improved by finding the optimal number and placement of memory controllers using machine learning approach with low overheads. The proposed architecture improves the average bandwidth of the network by 21.03% and reduces the network energy and memory access latency by 12.60% and 20.45%, respectively as compared to the traditional NoC architecture.

Dependence-aware Slice Execution to Boost MLP in Slice-out-of-order Cores

Exploiting memory-level parallelism (MLP) is crucial to hide long memory and last-level cache access latencies. While out-of-order (OoO) cores, and techniques building on them, are effective at exploiting MLP, they deliver poor energy efficiency due to their complex and energy-hungry hardware. This work revisits slice-out-of-order (sOoO) cores as an energy-efficient alternative for MLP exploitation. sOoO cores achieve energy efficiency by constructing and executing slices of MLP-generating instructions out-of-order only with respect to the rest of instructions; the slices and the remaining instructions, by themselves, execute in-order. However, we observe that existing sOoO cores miss significant MLP opportunities due to their dependence-oblivious in-order slice execution, which causes dependent slices to frequently block MLP generation. To boost MLP generation, we introduce Freeway, a sOoO core based on a new dependence-aware slice execution policy that tracks dependent slices and keeps them from blocking subsequent independent slices and MLP extraction. The proposed core incurs minimal area and power overheads, yet approaches the MLP benefits of fully OoO cores. Our evaluation shows that Freeway delivers 12% better performance than the state-of-the-art sOoO core and is within 7% of the MLP limits of full OoO execution.

SparseP

Several manufacturers have already started to commercialize near-bank Processing-In-Memory (PIM) architectures, after decades of research efforts. Near-bank PIM architectures place simple cores close to DRAM banks. Recent research demonstrates that they can yield significant performance and energy improvements in parallel applications by alleviating data access costs. Real PIM systems can provide high levels of parallelism, large aggregate memory bandwidth and low memory access latency, thereby being a good fit to accelerate the Sparse Matrix Vector Multiplication (SpMV) kernel. SpMV has been characterized as one of the most significant and thoroughly studied scientific computation kernels. It is primarily a memory-bound kernel with intensive memory accesses due its algorithmic nature, the compressed matrix format used, and the sparsity patterns of the input matrices given. This paper provides the first comprehensive analysis of SpMV on a real-world PIM architecture, and presents SparseP, the first SpMV library for real PIM architectures. We make three key contributions. First, we implement a wide variety of software strategies on SpMV for a multithreaded PIM core, including (1) various compressed matrix formats, (2) load balancing schemes across parallel threads and (3) synchronization approaches, and characterize the computational limits of a single multithreaded PIM core. Second, we design various load balancing schemes across multiple PIM cores, and two types of data partitioning techniques to execute SpMV on thousands of PIM cores: (1) 1D-partitioned kernels to perform the complete SpMV computation only using PIM cores, and (2) 2D-partitioned kernels to strive a balance between computation and data transfer costs to PIM-enabled memory. Third, we compare SpMV execution on a real-world PIM system with 2528 PIM cores to an Intel Xeon CPU and an NVIDIA Tesla V100 GPU to study the performance and energy efficiency of various devices, i.e., both memory-centric PIM systems and conventional processor-centric CPU/GPU systems, for the SpMV kernel. SparseP software package provides 25 SpMV kernels for real PIM systems supporting the four most widely used compressed matrix formats, i.e., CSR, COO, BCSR and BCOO, and a wide range of data types. SparseP is publicly and freely available at https://github.com/CMU-SAFARI/SparseP. Our extensive evaluation using 26 matrices with various sparsity patterns provides new insights and recommendations for software designers and hardware architects to efficiently accelerate the SpMV kernel on real PIM systems.

Optimizing Depthwise Separable Convolution Operations on GPUs

The depthwise separable convolution is commonly seen in convolutional neural networks (CNNs), and is widely used to reduce the computation overhead of a standard multi-channel 2D convolution. Existing implementations of depthwise separable convolutions target accelerating model training with large batch sizes with a large number of samples to be processed at once. Such approaches are inadequate for small-batch-sized model training and the typical scenario of model inference where the model takes in a few samples at once. This article aims to bridge the gap of optimizing depthwise separable convolutions by targeting the GPU architecture. We achieve this by designing two novel algorithms to improve the column and row reuse of the convolution operation to reduce the number of memory operations performed on the width and the height dimensions of the 2D convolution. Our approach employs a dynamic tile size scheme to adaptively distribute the computational data across GPU threads to improve GPU utilization and to hide the memory access latency. We apply our approach on two GPU platforms: an NVIDIA RTX 2080Ti GPU and an embedded NVIDIA Jetson AGX Xavier GPU, and two data types: 32-bit floating point (FP32) and 8-bit integer (INT8). We compared our approach against cuDNN that is heavily tuned for the NVIDIA GPU architecture. Experimental results show that our approach delivers over 2× (up to 3×) performance improvement over cuDNN. We show that, when using a moderate batch size, our approach averagely reduces the end-to-end training time of MobileNet and EfficientNet by 9.7 and 7.3 percent respectively, and reduces the end-to-end inference time of MobileNet and EfficientNet by 12.2 and 11.6 percent respectively.

Graph Connectivity in Log Steps Using Label Propagation

The fastest deterministic algorithms for connected components take logarithmic time and perform superlinear work on a Parallel Random Access Machine (PRAM). These algorithms maintain a spanning forest by merging and compressing trees, which requires pointer-chasing operations that increase memory access latency and are limited to shared-memory systems. Many of these PRAM algorithms are also very complicated to implement. Another popular method is “leader-contraction” where the challenge is to select a constant fraction of leaders that are adjacent to a constant fraction of non-leaders with high probability, but this can require adding more edges than were in the original graph. Instead we investigate label propagation because it is deterministic, easy to implement, and does not rely on pointer-chasing. Label propagation exchanges representative labels within a component using simple graph traversal, but it is inherently difficult to complete in a sublinear number of steps. We are able to overcome the problems with label propagation for graph connectivity. We introduce a surprisingly simple framework for deterministic, undirected graph connectivity using label propagation that is easily adaptable to many computational models. It achieves logarithmic convergence independently of the number of processors and without increasing the edge count. We employ a novel method of propagating directed edges in alternating direction while performing minimum reduction on vertex labels. We present new algorithms in PRAM, Stream, and MapReduce. Given a simple, undirected graph [Formula: see text] with [Formula: see text] vertices, [Formula: see text] edges, our approach takes O(m) work each step, but we can only prove logarithmic convergence on a path graph. It was conjectured by Liu and Tarjan (2019) to take [Formula: see text] steps or possibly [Formula: see text] steps. Our experiments on a range of difficult graphs also suggest logarithmic convergence. We leave the proof of convergence as an open problem.

Heterogeneity-aware Multicore Synchronization for Intermittent Systems

Intermittent systems enable batteryless devices to operate through energy harvesting by leveraging the complementary characteristics of volatile (VM) and non-volatile memory (NVM). Unfortunately, alternate and frequent accesses to heterogeneous memories for accumulative execution across power cycles can significantly hinder computation progress. The progress impediment is mainly due to more CPU time being wasted for slow NVM accesses than for fast VM accesses. This paper explores how to leverage heterogeneous cores to mitigate the progress impediment caused by heterogeneous memories. In particular, a delegable and adaptive synchronization protocol is proposed to allow memory accesses to be delegated between cores and to dynamically adapt to diverse memory access latency. Moreover, our design guarantees task serializability across multiple cores and maintains data consistency despite frequent power failures. We integrated our design into FreeRTOS running on a Cypress device featuring heterogeneous dual cores and hybrid memories. Experimental results show that, compared to recent approaches that assume single-core intermittent systems, our design can improve computation progress at least 1.8x and even up to 33.9x by leveraging core heterogeneity.

Efficient classification of private memory blocks

Shared memory architectures are pervasive in the multicore technology era. Still, sequential and parallel applications use most of the data as private in a multicore system. Recent proposals using this observation and driven by a classification of private/shared memory data can reduce the coherence directory area or the memory access latency. The effectiveness of these proposals depends on the accuracy of the classification. The existing proposals perform the private/shared classification at page granularity, leading to a miss-classification and reducing the number of detected private memory blocks.We propose a mechanism able to accurately classify memory blocks using the existing translation lookaside buffers (TLB), which increases the effectiveness of proposals relying on a private/shared classification. Our experimental results show that the proposed scheme reduces L1 cache misses by 25% compared to a page-grain classification approach, which translates into an improvement in system performance by 8.0% with respect to a page-grain approach.

ThunderRW

As random walk is a powerful tool in many graph processing, mining and learning applications, this paper proposes an efficient in-memory random walk engine named ThunderRW. Compared with existing parallel systems on improving the performance of a single graph operation, ThunderRW supports massive parallel random walks. The core design of ThunderRW is motivated by our profiling results: common RW algorithms have as high as 73.1% CPU pipeline slots stalled due to irregular memory access, which suffers significantly more memory stalls than the conventional graph workloads such as BFS and SSSP. To improve the memory efficiency, we first design a generic step-centric programming model named Gather-Move-Update to abstract different RW algorithms. Based on the programming model, we develop the step interleaving technique to hide memory access latency by switching the executions of different random walk queries. In our experiments, we use four representative RW algorithms including PPR, DeepWalk, Node2Vec and MetaPath to demonstrate the efficiency and programming flexibility of ThunderRW. Experimental results show that ThunderRW outperforms state-of-the-art approaches by an order of magnitude, and the step interleaving technique significantly reduces the CPU pipeline stall from 73.1% to 15.0%.

Data-Level Parallelism Oriented Memory Access and On-Chip Buffering Mechanisms for a Loop Accelerator

Memory access latency is always a bottleneck for the performance improvement of data-intensive applications. Exploiting the memory access patterns of Data-Level Parallelism(DLP) is a promising way for loop accelerators to reduce the latency significantly. This paper proposes two DLP-oriented data provisioning mechanisms to alleviate memory access latency: 1) DLP-oriented memory access(DoMA) for efficiently utilizes the available memory bandwidth. 2) a data access patterns aware on-chip buffer(PABUF) for exploiting reuse in a user-transparent manner. Unlike those loop accelerators using traditional DMA to access global memory, DoMA efficiently reduces the transmission of useless data by adjusting the size of requests intelligently. In addition, PABUF, which manages data using DLP’s memory access patterns without software engineering efforts, allows the loop accelerators to access data in parallel. Experiments show that when our mechanisms are integrated into a loop accelerator based on Rocket Chip Coprocessor(RoCC), it can achieve 4.20x-10.65x(6.81x on average) speedups with negligible overhead of power and area compared to L1 Cache.

CLU

Convolutional/Deep Neural Networks (CNNs/DNNs) are rapidly growing workloads for the emerging AI-based systems. The gap between the processing speed and the memory-access latency in multi-core systems affects the performance and energy efficiency of the CNN/DNN tasks. This article aims to alleviate this gap by providing a simple and yet efficient near-memory accelerator-based system that expedites the CNN inference. Towards this goal, we first design an efficient parallel algorithm to accelerate CNN/DNN tasks. The data is partitioned across the multiple memory channels (vaults) to assist in the execution of the parallel algorithm. Second, we design a hardware unit, namely the convolutional logic unit (CLU), which implements the parallel algorithm. To optimize the inference, the CLU is designed, and it works in three phases for layer-wise processing of data. Last, to harness the benefits of near-memory processing (NMP), we integrate homogeneous CLUs on the logic layer of the 3D memory, specifically the Hybrid Memory Cube (HMC). The combined effect of these results in a high-performing and energy-efficient system for CNNs/DNNs. The proposed system achieves a substantial gain in the performance and energy reduction compared to multi-core CPU- and GPU-based systems with a minimal area overhead of 2.37%.

New Two-Level L1 Data Cache Bypassing Technique for High Performance GPUs

On-chip caches of graphics processing units (GPUs) have contributed to improved GPU performance by reducing long memory access latency. However, cache efficiency remains low despite the facts that recent GPUs have considerably mitigated the bottleneck problem of L1 data cache. Although the cache miss rate is a reasonable metric for cache efficiency, it is not necessarily proportional to GPU performance. In this study, we introduce a second key determinant to overcome the problem of predicting the performance gains from L1 data cache based on the assumption that miss rate only is not accurate. The proposed technique estimates the benefits of the cache by measuring the balance between cache efficiency and throughput. The throughput of the cache is predicted based on the warp occupancy information in the warp pool. Then, the warp occupancy is used for a second bypass phase when workloads show an ambiguous miss rate. In our proposed architecture, the L1 data cache is turned off for a long period when the warp occupancy is not high. Our two-level bypassing technique can be applied to recent GPU models and improves the performance by 6% on average compared to the architecture without bypassing. Moreover, it outperforms the conventional bottleneck-based bypassing techniques.

Energy Efficient Placement of Workloads in Composable Data Center Networks

This paper studies the energy efficiency of composable data center (DC) infrastructures over network topologies. Using a mixed integer linear programming (MILP) model, we compare the performance of disaggregation at rack-scale and pod-scale over selected electrical, optical and hybrid network topologies relative to a traditional DC. Relative to a pod-scale DC, the results show that physical disaggregation at rack-scale is sufficient for optimal efficiency when the optical network topology is adopted, and resource components are allocated in a suitable manner. The optical network topology also enables optimal energy efficiency in composable DCs. The paper also studies logical disaggregation of traditional DC servers over an optical network topology. Relative to physical disaggregation at rack-scale, logical disaggregation of server resources within each rack enables marginal fall in the total DC power consumption (TDPC) due to improved resource demands placement. Hence, an adaptable composable infrastructure that can support both in memory (access) latency sensitive and insensitive workloads is enabled. We also conduct a study of the adoption of micro-service architecture in both traditional and composable DCs. Our results show that increasing the modularity of workloads improves the energy efficiency in traditional DCs, but disproportionate utilization of DC resources persists. A combination of disaggregation and micro-services achieved up to 23% reduction in the TDPC of the traditional DC by enabling optimal resources utilization and energy efficiencies. Finally, we propose a heuristic for energy efficient placement of workloads in composable DCs which replicates the trends produced by the MILP model formulated in this paper.

Implementation and Optimization of Data Prefetching Algorithm Based on LLVM Compilation System

In order to reduce the problem of mismatch between high-performance processors and DRAM speeds, current processors have added a cache structure, but the low cache hit rate also seriously affects the actual performance of the program. Data prefetching technology can alleviate the problems of memory access latency and low hit rate caused by the speed difference between high-performance processors and DRAM. Based on the LLVM open source compiler, this article first implements the data prefetch module on the Shenwei platform. This paper improves the prefetch distance algorithm, proposes a new prefetch scheduling algorithm, introduces a cost model to evaluate the prefetch revenue, and accurately determines the insertion timing of the prefetch instruction to improve the cache hit rate. SPEC2006 performance test results show that after optimization, Shenwei 1621 processor single-core can achieve a maximum performance improvement of 50%, and an average performance improvement of 11%.

Evaluation of Static Mapping for Dynamic Space-Shared Multi-task Processing on FPGAs

Whilst FPGAs have been used in cloud ecosystems, it is still extremely challenging to achieve high compute density when mapping heterogeneous multi-tasks on shared resources at runtime. This work addresses this by treating the FPGA resource as a service and employing multi-task processing at the high level, design space exploration and static off-line partitioning in order to allow more efficient mapping of heterogeneous tasks onto the FPGA. In addition, a new, comprehensive runtime functional simulator is used to evaluate the effect of various spatial and temporal constraints on both the existing and new approaches when varying system design parameters. A comprehensive suite of real high performance computing tasks was implemented on a Nallatech 385 FPGA card and show that our approach can provide on average 2.9 × and 2.3 × higher system throughput for compute and mixed intensity tasks, while 0.2 × lower for memory intensive tasks due to external memory access latency and bandwidth limitations. The work has been extended by introducing a novel scheduling scheme to enhance temporal utilization of resources when using the proposed approach. Additional results for large queues of mixed intensity tasks (compute and memory) show that the proposed partitioning and scheduling approach can provide higher than 3 × system speedup over previous schemes.

Efficient local locking for massively multithreaded in-memory hash-based operators

The join and group-by aggregation are two memory intensive operators that are affecting the performance of relational databases. Hashing is a common approach used to implement both operators. Recent paradigm shifts in multi-core processor architectures have reinvigorated research into how the join and group-by aggregation operators can leverage these advances. However, the poor spatial locality of the hashing approach has hindered performance on multi-core processor architectures which rely on using large cache hierarchies for latency mitigation. Multithreaded architectures can better cope with poor spatial locality by masking memory latency with many outstanding requests. Nevertheless, the number of parallel threads, even in the most advanced multithreaded processors, such as UltraSPARC, is not enough to fully cover the main memory access latency. In this paper, we explore the hardware re-configurability of FPGAs to enable deeper execution pipelines that maintain hundreds (instead of tens) of outstanding memory requests across four FPGAs-drastically increasing concurrency and throughput. We present two end-to-end in-memory accelerators for the join and group-by aggregation operators using FPGAs. Both accelerators use massive multithreading to mask long memory delays of traversing linked-list data structures, while concurrently managing hundreds of thread states across four FPGAs locally. We explore how content addressable memories can be intermixed within our multithreaded designs to act as a synchronizing cache, which enforces locks and merges jobs together before they are written to memory. Throughput results for our hash-join operator accelerator show a speedup between 2times and 3.4times over the best multi-core approaches with comparable memory bandwidths on uniform and skewed datasets. The accelerator for the hash-based group-by aggregation operator demonstrates that leveraging CAMs achieves average speedup of 3.3times with a best case of 9.4times in terms of throughput over CPU implementations across five types of data distributions.

A Performance Evaluation of DRAM Access for In-Memory Databases

The latest CPUs(computer cpu processors) employ multiple cores, massively superscalar pipelines, out-of-order execution of tons of instructions, and advanced SIMD capabilities, which can hide the memory access latency. And most of recent memory-oriented data structures have already benefit from these features. However, due to the complexity of data organization, these CPUs do not always work well in memory resident database systems (MMDBs), particularly regarding storing data in dynamic random-access memory (DRAM). This article studies memory-efficient data structures by analyzing the run time, access latency, cache misses, instructions per cycle (IPC), and DRAM reads (bytes). Then, we design and implement two data organization schemas in the main memory database: dispersing data block organization and clustering data block organization. Using algorithmic engineering and careful attention to internal parallelism, cache alignment can hide the memory access latency. However, we find that these data structures work well in some cases, though they have been eclipsed in the face of complex access paths. To determine the reasons, we study the impact of database techniques on memory access latency, such as data partitioning, storage models, and by processing algorithms. With the specific main memory database system, we estimate the performance of each data organization schema based on DRAM DDR4 and the latest Intel Haswell microarchitecture. In conclusion, this work will make DRAM access applicable in real-world situations by implementing the schema to systems, such as in-memory databases.

A Performance-Stable NUMA Management Scheme for Linux-Based HPC Systems

Linux is becoming the de-facto standard operating system for today’s high-performance computing (HPC) systems because it can satisfy the demands of many HPC systems for rich operating system (OS) features. However, owing to features intended for the general-purpose OS, Linux has many OS noise sources such as page faults or thread migrations that can result in the unstable performance of HPC application. Furthermore, in the case of the non-uniform memory access (NUMA) architecture, which has different memory access latencies to local and remote memory nodes, the performance stability of the application can be more exacerbated by the OS noise. In this paper, we address the OS noise caused by Linux in the NUMA architecture and propose a novel performance-stable NUMA management scheme called Stable-NUMA . Stable-NUMA comprises three techniques for improving performance stability: two-level thread clustering, state-based page placement, and selective page profiling. Our proposed Stable-NUMA scheme significantly alleviates OS noise and enhances the local memory access ratio of the NUMA system as compared to Linux. We implemented Stable-NUMA in Linux and experimented with various HPC workloads. The evaluation results demonstrated that Stable-NUMA outperforms Linux with and without its NUMA-aware feature by up to 25% in terms of average performance and 73% in terms of performance stability.