Transformer-based large language models (LLMs) exhibit remarkable generative capabilities, but their inference throughput is limited by the autoregressive decoding process, which generates only one token per iteration. Speculative decoding mitigates this bottleneck by using a lightweight draft language model (DLM) to generate multiple draft tokens, which are then verified in parallel by a more accurate target language model (TLM). To accommodate the differing computational patterns of the DLM and TLM, prior work has leveraged heterogeneous systems combining xPUs and processing-in-memory (PIM) units to offload compute- and memory-intensive operators, respectively. However, existing systems often adopt a fixed draft sequence length, leading to excessive rejection of draft tokens during verification-especially under large-batch scenarios-resulting in redundant computation and reduced efficiency. This paper proposes a runtime adaptive draft length adjustment technique that dynamically tailors the draft length for each request by monitoring cumulative acceptance probabilities, thereby minimizing the generation and verification of invalid tokens. Yet, integrating adaptive draft lengths into existing PIM-enabled heterogeneous systems introduces two new challenges: (1) sequential execution of the DLM and TLM becomes inefficient due to synchronization bubbles caused by request-wise variability in draft lengths, and (2) static operator mappings become suboptimal as draft length variability alters operator arithmetic intensities dynamically. To address these issues, we introduce SADDLE, a PIM-enabled heterogeneous system designed to exploit adaptive draft lengths effectively. SADDLE incorporates two key mechanisms: (1) an asynchronous speculative decoding pipeline that decouples DLM prediction and TLM verification to reduce idle time, and (2) an arithmetic intensity-aware operator scheduler that dynamically assigns operators to the most suitable hardware units. Experimental results show that SADDLE achieves average speedups of$\mathbf{2. 8 8} \times$over a state-of-the-art GPU-only solution and$\mathbf{1. 7 1} \times$over the best-performing GPU+PIM baseline.

Multi-Scale Deformable Attention (MSDAttn) has become a fundamental component in various vision tasks due to its effective multi-scale grid sampling (MSGS). However, its reliance on random sampling results in highly irregular memory access patterns, making it a memory-intensive operation inefficient for GPUs. Near-memory processing (NMP) offers a promising solution for accelerating memory-bound kernels, yet existing NMP-based attention accelerators remain suboptimal for MSDAttn due to incompatible load balancing and data reuse strategies. Specifically, current NMP solutions uniformly distribute processing elements (PEs) across all banks, leading to significant PE underutilization and excessive cross-bank data transfers. Moreover, most rely on locality-based reuse, which fails under MSDAttn’s unpredictable sampling patterns.

Hypergraph Neural Networks (HGNNs) are increasingly utilized to analyze complex inter-entity relationships. Traditional HGNN systems, based on a hyperedge-centric dataflow model, independently process aggregation tasks for hyperedges and vertices, leading to significant computational redundancy. This redundancy arises from recalculating shared information across different tasks. For the first time, we identify and harness implicit dataflows (i.e., dependencies) within HGNNs, introducing the microedge concept to effectively capture and reuse intricate shared information among aggregation tasks, thereby minimizing redundant computations. We have developed a new microedge-centric dataflow model that processes shared information as fine-grained microedge aggregation tasks. This dataflow model is supported by the Read-Process-Activate-Generate execution model, which aims to optimize parallelism among these tasks. Furthermore, our newly developed MeHyper, a microedge-centric HGNN accelerator, incorporates a decoupled pipeline for improved computational parallelism and a hierarchical feature management strategy to reduce off-chip memory accesses for large volumes of intermediate feature vectors generated. Our evaluation demonstrates that MeHyper substantially outperforms the leading CPUbased system PyG-CPU and the GPU-based system HyperGef, delivering performance improvements of $1,032.23 \times$ and $10.51 \times$, and energy efficiencies of $1,169.03 \times$ and $9.96 \times$, respectively.

The emerging Mixture-of-Experts (MoE) model mitigates the high compute cost of large-scale LLMs by sparsely activating a subset of experts during inference. However, MoE requires storing massive expert parameters, creating a severe memory bottleneck on resource-constrained GPUs. Existing approaches offload parameters to host memory and prefetch activated experts to GPU memory with sophisticated policies, but these solutions are tailored to single-batch inference and suffer from communication bottlenecks at larger batch sizes, limiting throughput. We identify two forms of locality in expert activation: a small set of experts are frequently invoked across inference (global locality), while others recur within short decoding bursts (temporal locality). To exploit this, we propose Diff-MoE, which introduces a differential cache hierarchy in GPU memory. Globally hot experts reside in per-layer high-priority caches, locally hot ones are dynamically managed in per-layer medium-priority caches under a priority-driven replacement policy, and the remaining cold experts are cached temporarily and evicted on demand. Moreover, Diff-MoE incorporates a lightweight predictor that prefetches experts likely needed in the next MoE layer, overlapping migration with computation to further reduce latency. Our evaluation shows that Diff-MoE improves inference throughput by 2.74 ×, 2.22 ×, and 1.55 × over DeepSpeed, Pre-gated MoE, and MoE-Infinity, respectively.

With the merits of high productivity and ease of use, highlevel synthesis (HLS) tools bring hope to fast FPGA-based architecture development. However, their usability and popularity are still limited due to lack of support for dynamic memory management (DMM). Though HLS-compatible DMM solutions have been proposed recently, nevertheless, based on our investigation, none of them can hit high performance (i.e., minimal memory (de-)allocation latency) and resource efficiency (i.e., managing arbitrarily sized memory with minimal FPGA resource consumption) with one stone, seriously limiting their practicality. In response, we propose HeroDMM, a high-performance and resource-efficient dynamic memory manager for HLS. Specifically, HeroDMM organizes the managed memory area with a novel cartesian-like tree (CT) structure, a key to resolving the dilemma between (de-)allocation latency and resource efficiency standing in front of prior efforts. With the CT structure, HeroDMM further devises a delicate memory management algorithm and specializes the hardware implementation for achieving ever-higher performance while ensuring resource efficiency. Results show that HeroDMM outperforms state-of-the-art HLS-compatible DMM solutions by 61.69%~99.99% in performance improvement and 23.79%~97.22% in resource consumption savings.

Nowadays, high-level synthesis (HLS) has gained prominence for FPGA-based architecture prototyping, enhancing productivity significantly. Despite this advancement, HLS tools are impeded by a critical drawback: they lack support for dynamic memory management (DMM), leading to static mem-ory allocation and suboptimal use of memory resources. In response, numerous efforts have been made to develop DMM solutions compatible with HLS. However, our analysis indicates that existing solutions fail to concurrently meet the desired trifecta of scalability (efficient management of memory of any size), efficiency (minimal latency in memory (de-)allocation), and robustness (low allocation failure rates). This limitation hampers their applicability in real-world scenarios. In this paper, we introduce GraDMM, a “three-birds-one- stone” solution that comprehensively enhances the scalability, efficiency, and robustness of DMM. The key insight is to formulate memory (de-)allocation as graph analytics and lever-age sophisticated FPGA-based graph processing techniques. To achieve scalability, GraDMM specializes a simplified pipeline that significantly suppresses resource utilization expansion caused by managed memory scaling. This is crucial for managing arbitrarily sized memory on resource-limited FPGA platforms. For efficiency, GraDMM implements a data-centric concurrent traversal scheme and a shortcut-assisted fast traversal policy to accelerate (de-)allocation-guided graph traversal, reducing mem-ory (de-)allocation latency. To enhance robustness, GraDMM incorporates an adaptive memory defragmenter that defragments managed memory to minimize fragmentation-induced allocation failures. GraDMM is encapsulated as a library, providing high- level interfaces for users and ensuring synthesizability with Vi- vado HLS. Experimental results demonstrate that GraDMM out-performs three state-of-the-art HLS-compatible DMM solutions by significant margins: 56.71 %-85.59% in resource consumption savings, 78.94 % -99.99 % in (de-)allocation latency improvement, and 10.71 %-65.75% in allocation failure reduction.

Graph Convolutional Networks (GCNs) hold great promise in facilitating machine learning on graph-structured data. However, the sparsity of graphs often results in a significant number of irregular memory accesses, leading to inefficient data movement for existing GCNs accelerators. With the advancement of 3D stacked technology, the processing-in-memory (PIM) architecture has emerged as a promising solution for graph processing. Nevertheless, existing PIM accelerators are confronted with the challenges of irregular remote access in the aggregation phase of GCNs and dynamic workload variations between phases. In this paper, we present GCNim, a PIM accelerator based on 3D stacked memory, which features two key innovations in terms of the computation model and hardware designs. First, we present a PIM-based hybrid computation model, which employs a remote merging strategy to achieve the outer product in aggregation and the row-wise product in combination. Second, GCNim builds a three-stage aggregation and combination pipeline and integrates unified processing elements (PEs) supporting these three stages at the bank level, achieving load balance among PEs through a lightweight data placement algorithm. Compared with the state-of-the-art software frameworks running on CPUs and GPUs, GCNim achieves an average speedup of 3,736.06× and 76.56×, respectively. Moreover, GCNim outperforms the state-of-the-art GCN hardware accelerators, I-GCN, PEDAL, FlowGNN, and GCIM, with average speedups of 3.35×, 8.97×, 2.24×, and 5.58×, respectively.

Hypergraph processing has emerged as an effective approach to analyze complex multilateral relationships in real-world scenarios. Existing hypergraph processing solutions based on conventional architectures are severely bottlenecked by off-chip memory accesses. In this paper, we propose the first Processing-In-Memory (PIM)-featured ReRAM-based hypergraph accelerator, dubbed PhGraph, which facilitates performance-and energy-efficient hypergraph processing. On the hardware level, PhGraph integrates analog memristor-based PIM (with high matrix-grained parallelism) and digital memristor-based PIM (for high bipartite-edge-grained efficiency) into one standalone solution. On the software level, an overlap-aware hypergraph partitioning mechanism is proposed to polarize hypergraph workloads into matrix-formatted dense and bipartite-edge-formatted sparse partitions for performance acceleration using analog memristor-based PIM and digital ones, respectively. In addition, PhGraph is equipped with load-balanced partition scheduling and algorithm mapping co-designs to boost hardware utilization and efficiency. Experimental results show that PhGraph outperforms the state-of-the-art CPU-, FPGA-, and ASIC-based solutions by up to 4,309.81×, 547.13×, and 166.76× in terms of performance, and 36,416.11×, 924.12×, and 41.44× in terms of energy-savings, respectively.

Genome graphs analysis has emerged as an effective means to enable mapping DNA fragments (known as reads) to the reference genome. It replaces the traditional linear reference with a graph-based representation to augment the genetic variations and diversity information, significantly improving the quality of genotyping. The in-depth characterization of genome graphs analysis uncovers that it is bottlenecked by the irregular seed index access and the intensive alignment operation, stressing both the memory system and computing resources.Based on these observations, we propose MeG2, a lightweight, commodity DRAM-compliant, processing-in-memory architecture to accelerate genome graphs analysis. MeG2is specifically integrated with the capabilities of both near-memory processing and bitwise in-situ computation. Specifically, MeG2leverages the low access latency of near-memory processing with the index-centric offload mechanism to alleviate the irregular memory access in the seeding procedure, and harnesses the row-parallel capacity of in-situ computation with the distance-aware technique to exploit the intensive computational parallelism in the alignment process. Results show that MeG2outperforms the CPU-, GPU-, and ASIC-based genome graphs analysis solutions by 502× (30.2×), 272× (15.1× ), and 5.5× (8.3×) for short (long) reads, while reducing energy consumption by 1628× (85.6×), 1443× (77.1×), and 7.8× (11.7×), respectively. We also demonstrate that MeG2offers significant improvements over existing PIM-based genome sequence analysis accelerators.

Multi-version graph processing has been widely used to solve many real-world problems. The process of the multi-version graph processing typically includes: (1) a history graph version switching at a specific time and (2) graph processing on this history graph. Existing multi-version graph systems assume ideally that every request for a particular graph version at a particular time will have a corresponding snapshot available. However, in most cases, this is not true. Then existing solutions usually have to settle with an "approximating" version as a substitute, leading to unexpected results for the underlying graph algorithm and thus reducing the practicality of a multi-version graph system for many application scenarios significantly.In this paper, we observe that only a few graph updates have a great impact on the final results. We therefore present AFaVS, a novel multi-version graph system that can improve accuracy effectively in both time- and memory-efficient manners. The cornerstone of AFaVS lies in a novel concept "value" that characterizes the importance of graph updates. AFaVS proposes differential management of updates based on their values and achieves higher accuracy while preserving processing and memory efficiency. AFaVS is also equipped with value-guided version switching and locality-aware optimizations to boost its overall efficiency. Our results on a variety of real-world datasets show that AFaVS outperforms four state-of-the-art multi-version graph systems by 74.35%~95.72% in terms of accuracy improvement and 57.03%~90.44% in terms of memory reduction while introducing less than 2.96% extra computing time. We have deployed AFaVS in a disaster recovery system on the production cluster of Alibaba, achieving 78.8%~90.1% fewer error rates than advanced systems at a comparable efficiency.

Graph convolutional networks (GCNs) are promising to enable machine learning on graphs. GCNs exhibit mixed computational kernels, involving regular neural-network-like computing and irregular graph-analytics-like processing. Existing GCN accelerators obey a divide-and-conquer philosophy to architect two separate types of hardware to accelerate these two types of GCN kernels, respectively. This hybrid architecture improves intra-kernel efficiency but considers little inter-kernel interactions in a holistic view for improving overall efficiency.In this paper, we present a new GCN accelerator, RE-FLIP, with three key innovations in terms of architecture design, algorithm mappings, and practical implementations. First, ReFlip leverages PIM-featured crossbar architectures to build a unified architecture for supporting the two types of GCN kernels simultaneously. Second, ReFlip adopts novel algorithm mappings that can maximize potential performance gains reaped from the unified architecture by exploiting the massive crossbar-structured parallelism. Third, ReFlip assembles software/hardware co-optimizations to process real-world graphs efficiently. Compared to the state-of-the-art software frameworks running on Intel Xeon E5-2680v4 CPU and NVIDIA Tesla V100 GPU, ReFlip achieves the average speedups of 6,432× and 86.32× and the average energy savings of 9,817× and 302.44×, respectively. In addition, ReFlip also outperforms a state-of-the-art GCN hardware accelerator, AWB-GCN, by achieving an average speedup of 5.06× and an average energy saving of 15.63×.

Beyond ordinary graphs, hypergraphs are a graph representation to flexibly express complex multilateral relationships between entities. Hypergraph processing can be used to solve many real-world problems, e.g., machine learning, VLSI design, and image retrieval. Existing hypergraph processing systems handle a hypergraph in order of its hyperedge and vertex indices. This makes processing hypergraphs on generalpurpose architectures suffer significantly from excessive offchip memory accesses, most of which however are redundant in frequently accessing overlapped hyperedges and vertices, but the index-ordered scheduling destroys this potential locality.In this paper, we propose a novel Generate-Load-Apply (GLA) execution model to improve locality in hypergraph processing. The key insight of GLA is to use a concept of chain to characterize the overlapped feature of a hypergraph, exposing data reuse opportunities missed in existing hypergraph systems. The precondition of driving GLA model is to generate expected chains on the fly, but the software solution is so expensive that its overheads may outweigh the benefits achieved from the chain-driven scheduling. We further present ChGraph, the first hardware-accelerated hypergraph processing engine near each core. ChGraph is specialized in accelerating the chain generation and the chain-guided data loading (to hide memory access latency) while the general-purpose cores are responsible only for handling the apply operations of GLA. We evaluate ChGraph against a state-of-the-art hypergraph processing system Hygra on six hypergraph algorithms using five large real-world hypergraphs. Results on a simulated 16core system show that ChGraph reduces the number of offchip memory accesses by up to 4.56× and achieves up to 4.73× speedup while introducing only 0.26% area overhead.

Graph processing is promising to extract valuable insights in graphs. Nowadays, emerging 3D-stacked memories and silicon technologies can provide over terabytes per second memory bandwidth and thousands of processing elements (PEs) to meet the high hardware demand of graph applications. However, this leap in hardware capability does not result in a huge increase but even a degradation sometimes in performance for graph processing. In this paper, we discover that the centralized on-chip memory hierarchy adopted in existing graph accelerators is the villain causing poor scalability due to its quadratic increase of hardware overheads with respect to the number of PEs.We present a novel distributed on-chip memory hierarchy by leveraging the network-on-chip (NoC) to enable massively parallel graph processing. We architect ScalaGraph, a brand new graph processing accelerator, to exploit this insight. ScalaGraph adopts a software-hardware co-design to minimize NoC communication overheads via an efficient row-oriented dataflow mapping and runtime aggregation. A specialized scheduling mechanism is also proposed to improve load imbalance. Our results on a Xilinx Alveo U280 FPGA card show that ScalaGraph on a modest configuration of 512 PEs achieves 2.2× and 3.2× speedups over a state-of-theart graph accelerator GraphDyns and a GPU-based graph system Gunrock, respectively. Moreover, ScalaGraph enables supporting at least 1,024 PEs with nearly linear performance scaling while GraphDyns fails to work.

Processing-in-memory (PIM) is promising to solve the well-known data movement challenge by performing in-situ computations near the data. Leveraging PIM features is pretty profitable to boost the energy efficiency of applications. Early studies mainly focus on improving the programmability for computation offloading on PIM architectures. They lack a comprehensive analysis of computation locality and hence fail to accelerate a wide variety of applications. In this paper, we present a general-purpose instruction-level offloading technique for near-DRAM PIM architectures, namely IOTPIM, to exploit PIM features comprehensively. IOTPIM is novel with two technical advances: 1) a new instruction offloading policy that fully considers the locality of the whole on-chip cache hierarchy, and 2) an offloading performance benefit prediction model that directly predicts offloading performance benefits of an instruction based on the input dataset characterizes, preserving low analysis overheads. The evaluation demonstrates that IOTPIM can be applied to accelerate a wide variety of applications, including graph processing, machine learning, and image processing. IOT-PIM outperforms the state-of-the-art PIM offloading techniques by 1.28×-1.51× while ensuring offloading accuracy as high as 91.89% on average.

Hypergraph processing has emerged as a powerful approach for analyzing complex multilateral relationships among multiple entities. Past research on building hypergraph systems suggests that changing the scheduling order of bipartite edge tasks can improve the overlap-induced data locality in hypergraph processing. However, due to the complex intertwined connections between vertices and hyperedges, it is almost impossible to find a locality-optimal scheduling order. Thus, these task-centric hypergraph systems often suffer from substantial off-chip communications. In this paper, we first propose a novel data-centric Load-Trigger-Reduce (LTR) execution model to exploit fully the locality in hypergraph processing. Unlike a task-centric model that loads the required data along with a task, our LTR model invokes tasks as per the data used. Specifically, once the hypergraph data is loaded into the on-chip memory, all of its relevant computation tasks will be triggered simultaneously to output intermediate results, which are finally reduced to update the final results. Our LTR model enables all hypergraph data to be accessed once in each iteration. To fully exploit the LTR performance potential, we further architect an LTR-driven hypergraph accelerator, XuLin, which features with an adaptive data loading mechanism to minimize the loading cost via chunk merging at runtime. XuLin is also equipped with a priority-based differential data reduction scheme to reduce the impact of conflicting updates on performance. We have implemented XuLin both on a Xilinx Alveo U250 FPGA card and using a cycle-accurate simulator. The results show that XuLin outperforms the state-of-the-art hypergraph processing solutions Hygra and ChGraph by $20.47 \times$ and $8.77 \times$ on average, respectively.

Dynamic graph convolutional networks (DGCNs) have emerged as an effective approach to analyzing graph data that is constantly changing. The typical DGCNs incorporate not only graph convolutional networks (GCNs) to extract the structural information but also with recurrent neural networks (RNNs) to capture the temporal information from evolving graph data. These two alternative execution kernels of DGCNs impose unique architecture challenges for both types of kernels to be implemented efficiently. The presence of complex execution patterns of DGCNs renders existing architectures unsuitable. In this article, we present the first DGCN accelerator with an integrated architecture, named ReaDy, to accelerate DGCNs based on emerging PIM-featured ReRAM architectures. ReaDy is novel with an integrated architecture that enables running the GCN and RNN kernels of DGCNs simultaneously. Specifically, ReaDy is equipped with a redundancy-free scheduling mechanism to alleviate intrinsic dynamic irregularity for the GCN kernel, improving hardware utilization. In addition, ReaDy also includes a locality-aware dataflow strategy to exploit the inherent intervertex data locality for the RNN kernel, reducing superfluous data accesses to vertices and weight parameters. In a holistic view, ReaDy further enhances the entire system via an interkernel pipeline to reduce the off-chip accesses of intermediate results, boosting the overall efficiency of DGCNs significantly. Compared to the state-of-the-art software framework, PyGT, running on Intel Xeon E5-2680v4 CPU and NVIDIA Ampere A100 GPU, ReaDy achieves the average speedups of$955\times $and$27.33\times $, and the average energy savings of 1$093\times $and$80.21\times $, respectively. In addition, ReaDy outperforms ReFlip-ERA, which is obtained by combining a state-of-the-art GCN accelerator ReFlip and RNN accelerator ERA-LSTM, by an average speedup of$8.30\times $and an average energy saving of$7.29\times $.

Graph mining aims to explore interesting structural information of a graph. Pattern-centric systems typically transform a generic-purpose graph mining problem into a series of subgraph matching problems for high performance. Existing pattern-centric mining systems reduce the substantial search space towards a single pattern by exploring a highly-optimized matching order, but inherent computational redundancies of such a matching order itself still suffer severely, leading to significant performance degradation. The key innovation of this work lies in a general redundancy criterion that characterizes computational redundancies arising in not only handing a single pattern but also matching multiple patterns simultaneously. In this paper, we present SumPA, a high-performance pattern-centric graph mining system that can sufficiently remove redundant computations for any complex graph mining problems. SumPA features three key designs: (1) a pattern abstraction technique that can simplify numerous complex patterns into a few simple abstract patterns based on pattern similarity, (2) abstraction-guided pattern matching that completely eliminates (totally and partially) redundant computations during subgraph enumeration, and (3) a suite of system optimizations to maximize storage and computation efficiency. Our evaluation on a wide variety of real-world graphs shows that SumPA outperforms the two state-of-the-art systems Peregrine and GraphPi by up to 61.89× and 8.94×, respectively. For many mining problems on large graphs, Peregrine takes hours or even days while SumPA finishes in only a few minutes.

Existing FPGA-based graph accelerators, typically designed for static graphs, rarely handle dynamic graphs that often involve substantial graph updates (e.g., edge/node insertion and deletion) over time. In this paper, we aim to fill this gap. The key innovation of this work is to build an FPGA-based dynamic graph accelerator easily from any off-the-shelf static graph accelerator with minimal hardware engineering efforts (rather than from scratch). We observe \em spatial similarity of dynamic graph updates in the sense that most of graph updates get involved with only a small fraction of vertices. We therefore propose an FPGA library, called GraSU, to exploit spatial similarity for fast graph updates. GraSU uses a differential data management, which retains the high-value data (that will be frequently accessed) in the specialized on-chip UltraRAM while the overwhelming majority of low-value ones reside in the off-chip memory. Thus, GraSU can transform most of off-chip communications arising in dynamic graph updates into fast on-chip memory accesses. Our experiences show that GraSU can be easily integrated into existing state-of-the-art static graph accelerators with only 11 lines of code modifications. Our implementation atop AccuGraph using a Xilinx Alveo#8482; \ U250 board outperforms two state-of-the-art CPU-based dynamic graph systems, Stinger and Aspen, by an average of 34.24× and 4.42× in terms of update throughput, improving further overall efficiency by 9.80× and 3.07× on average.

Resistive random access memory (ReRAM) addresses the high memory bandwidth requirement challenge of graph analytics by integrating the computing logic in the memory. Due to the matrix-structured crossbar architecture, existing ReRAM-based accelerators, when handling real-world graphs that often have the skewed degree distribution, suffer from the severe sparsity problem arising from zero fillings and activation nondeterminism, incurring substantial ineffectual computations.In this paper, we observe that the sparsity sources lie in the consecutive mapping of source and destination vertex index onto the wordline and bitline of a crossbar. Although exhaustive graph reordering improves the sparsity-induced inefficiency, its totally-random (source and destination) vertex mapping leads to expensive overheads. This work exploits the insight in a mid-point vertex mapping with the random wordlines and consecutive bitlines. A cost-effective preprocessing is proposed to exploit the insight by rapidly exploring the crossbar-fit vertex reorderings but ignores the sparsity arising from activation dynamics. We present a novel ReRAM-based graph analytics accelerator, named Spara, which can maximize the workload density of crossbars dynamically by using a tightly-coupled bank parallel architecture further proposed. Results on real-world and synthesized graphs show that Spara outperforms GraphR and GraphSAR by 8.21 × and 5.01 × in terms of performance, and by 8.97 × and 5.68× in terms of energy savings (on average), while incurring a reasonable (<; 9.98%) pre-processing overhead.

FPGA-based graph processing accelerators are nowadays equipped with multiple pipelines for hardware acceleration of graph computations. However, their multi-pipeline efficiency can suffer greatly from the considerable overheads caused by the read/write conflicts in their on-chip BRAM from different pipelines, leading to significant performance degradation and poor scalability. In this article, we investigate the underlying causes behind such inter-pipeline read/write conflicts by focusing on multi-pipeline FPGAs for accelerating Sparse Matrix Vector Multiplication (SpMV) arising in graph processing. We exploit our key insight that the problem of eliminating inter-pipeline read/write conflicts for SpMV can be formulated as one of solving a row- and column-wise tiling problem for its associated adjacency matrix. However, how to partition a sparse adjacency matrix obtained from any graph with respect to a set of pipelines by both eliminating all the inter-pipeline read/write conflicts and keeping all the pipelines reasonably load-balanced is challenging. We present a conflict-free scheduler, WaveScheduler, that can dispatch different sub-matrix tiles to different pipelines without any read/write conflict. We also introduce two optimizations that are specifically tailored for graph processing, “degree-aware vertex index renaming” for improving load balancing and “data re-organization” for enabling sequential off-chip memory access, for all the pipelines. Our evaluation on Xilinx®Alveo™ U250 accelerator card with 16 pipelines shows that WaveScheduler can achieve up to 3.57 GTEPS, running much faster than native scheduling and two state-of-the-art FPGA-based graph accelerators (by 6.48× for “native,” 2.54× for HEGP, and 2.11× for ForeGraph), on average. In particular, these performance gains also scale up significantly as the number of pipelines increases.

Existing constraint-solving-based technique enables an efficient and high-coverage concurrency debugging. Yet, there remains a significant gap between the state of the art and the state of the programming practices for scaling to handle long-running execution of programs.