AlphaFold predicts protein structures from the amino acid sequence at or near experimental resolution, solving the 50-year-old protein folding challenge, leading to progress by transforming large-scale genomics data into protein structures. AlphaFold will also greatly change the scientific research model from low-throughput to high-throughput manner. The AlphaFold framework is a mixture of two types of workloads: MSA construction based on CPUs and model inference on GPUs. The first CPU stage dominates the overall runtime, taking hours for a single protein due to the large database sizes and I/O bottlenecks. However, GPUs in this CPU stage remain idle, resulting in low GPU utilization and restricting the capacity of large-scale structure predictions. Therefore, we proposed ParaFold, an open-source parallel version of AlphaFold for high throughput protein structure predictions. ParaFold separates the CPU and GPU parts to enable large-scale structure predictions. ParaFold also effectively reduces the CPU and GPU runtime with two optimizations without compromising the quality of prediction results: using multi-threaded parallelism on CPUs and using optimized JAX compilation on GPUs. We evaluated ParaFold with three datasets of different size and protein lengths. We evaluated the accuracy and efficiency of optimizations on CPUs and GPUs, and showed the large-scale prediction capability by running ParaFold inferences of 19,704 small proteins in five hours on one NVIDIA DGX-2. Using the JAX compile optimization, ParaFold attained a 13.8X average speedup over AlphaFold. ParaFold offers a rapid and effective approach for high-throughput structure predictions, leveraging the predictive power by running on supercomputers, with shorter time, and at a lower cost. The development of ParaFold will greatly speed up high-throughput studies and render the protein "structure-omics" feasible.
While there have been many studies on hardware acceleration for deep learning on images, there has been a rather limited focus on accelerating deep learning applications involving graphs. The unique characteristics of graphs, such as the irregular memory access and dynamic parallelism, impose several challenges when the algorithm is mapped to a CPU or GPU. To address these challenges while exploiting all the available sparsity, we propose a flexible architecture called SPA-GCN for accelerating Graph Convolutional Networks (GCN), the core computation unit in deep learning algorithms on graphs. The architecture is specialized for dealing with many small graphs since the graph size has a significant impact on design considerations. In this context, we use SimGNN, a neural-network-based graph matching algorithm, as a case study to demonstrate the effectiveness of our architecture. The experimental results demonstrate that SPA-GCN can deliver a high speedup compared to a multi-core CPU implementation and a GPU implementation, showing the efficiency of our design.
Graph Neural Networks (GNNs) have shown success in learning from graph-structured data, with applications to fraud detection, recommendation, and knowledge graph reasoning. However, training GNN efficiently is challenging because: 1) GPU memory capacity is limited and can be insufficient for large datasets, and 2) the graph-based data structure causes irregular data access patterns. In this work, we provide a method to statistical analyze and identify more frequently accessed data ahead of GNN training. Our data tiering method not only utilizes the structure of input graph, but also an insight gained from actual GNN training process to achieve a higher prediction result. With our data tiering method, we additionally provide a new data placement and access strategy to further minimize the CPU-GPU communication overhead. We also take into account of multi-GPU GNN training as well and we demonstrate the effectiveness of our strategy in a multi-GPU system. The evaluation results show that our work reduces CPU-GPU traffic by 87-95% and improves the training speed of GNN over the existing solutions by 1.6-2.1x on graphs with hundreds of millions of nodes and billions of edges.
Deep generative models applied to audio have improved by a large margin the state-of-the-art in many speech and music related tasks. However, as raw waveform modelling remains an inherently difficult task, audio generative models are either computationally intensive, rely on low sampling rates, are complicated to control or restrict the nature of possible signals. Among those models, Variational AutoEncoders (VAE) give control over the generation by exposing latent variables, although they usually suffer from low synthesis quality. In this paper, we introduce a Realtime Audio Variational autoEncoder (RAVE) allowing both fast and high-quality audio waveform synthesis. We introduce a novel two-stage training procedure, namely representation learning and adversarial fine-tuning. We show that using a post-training analysis of the latent space allows a direct control between the reconstruction fidelity and the representation compactness. By leveraging a multi-band decomposition of the raw waveform, we show that our model is the first able to generate 48kHz audio signals, while simultaneously running 20 times faster than real-time on a standard laptop CPU. We evaluate synthesis quality using both quantitative and qualitative subjective experiments and show the superiority of our approach compared to existing models. Finally, we present applications of our model for timbre transfer and signal compression. All of our source code and audio examples are publicly available.
Model quantization has emerged as an indispensable technique to accelerate deep learning inference. While researchers continue to push the frontier of quantization algorithms, existing quantization work is often unreproducible and undeployable. This is because researchers do not choose consistent training pipelines and ignore the requirements for hardware deployments. In this work, we propose Model Quantization Benchmark (MQBench), a first attempt to evaluate, analyze, and benchmark the reproducibility and deployability for model quantization algorithms. We choose multiple different platforms for real-world deployments, including CPU, GPU, ASIC, DSP, and evaluate extensive state-of-the-art quantization algorithms under a unified training pipeline. MQBench acts like a bridge to connect the algorithm and the hardware. We conduct a comprehensive analysis and find considerable intuitive or counter-intuitive insights. By aligning the training settings, we find existing algorithms have about the same performance on the conventional academic track. While for the hardware-deployable quantization, there is a huge accuracy gap which remains unsettled. Surprisingly, no existing algorithm wins every challenge in MQBench, and we hope this work could inspire future research directions.
We present a careful comparison of two model-free control algorithms, Evolution Strategies (ES) and Proximal Policy Optimization (PPO), with receding horizon model predictive control (MPC) for operating simulated, price responsive water heaters. Four MPC variants are considered: a one-shot controller with perfect forecasting yielding optimal control; a limited-horizon controller with perfect forecasting; a mean forecasting-based controller; and a two-stage stochastic programming controller using historical scenarios. In all cases, the MPC model for water temperature and electricity price are exact; only water demand is uncertain. For comparison, both ES and PPO learn neural network-based policies by directly interacting with the simulated environment under the same scenarios used by MPC. All methods are then evaluated on a separate one-week continuation of the demand time series. We demonstrate that optimal control for this problem is challenging, requiring more than 8-hour lookahead for MPC with perfect forecasting to attain the minimum cost. Despite this challenge, both ES and PPO learn good general purpose policies that outperform mean forecast and two-stage stochastic MPC controllers in terms of average cost and are more than two orders of magnitude faster at computing actions. We show that ES in particular can leverage parallelism to learn a policy in under 90 seconds using 1150 CPU cores.
Graph convolutional networks (GCNs) have been introduced to effectively process non-euclidean graph data. However, GCNs incur large amounts of irregularity in computation and memory access, which prevents efficient use of traditional neural network accelerators. Moreover, existing dedicated GCN accelerators demand high memory volumes and are difficult to implement onto resource limited edge devices. In this work, we propose LW-GCN, a lightweight FPGA-based accelerator with a software-hardware co-designed process to tackle irregularity in computation and memory access in GCN inference. LW-GCN decomposes the main GCN operations into sparse-dense matrix multiplication (SDMM) and dense matrix multiplication (DMM). We propose a novel compression format to balance workload across PEs and prevent data hazards. Moreover, we apply data quantization and workload tiling, and map both SDMM and DMM of GCN inference onto a uniform architecture on resource limited hardware. Evaluation on GCN and GraphSAGE are performed on Xilinx Kintex-7 FPGA with three popular datasets. Compared to existing CPU, GPU, and state-of-the-art FPGA-based accelerator, LW-GCN reduces latency by up to 60x, 12x and 1.7x and increases power efficiency by up to 912x., 511x and 3.87x, respectively. Furthermore, compared with NVIDIA's latest edge GPU Jetson Xavier NX, LW-GCN achieves speedup and energy savings of 32x and 84x, respectively.
In recent years BERT shows apparent advantages and great potential in natural language processing tasks. However, both training and applying BERT requires intensive time and resources for computing contextual language representations, which hinders its universality and applicability. To overcome this bottleneck, we propose a deep bidirectional language model by using window masking mechanism at attention layer. This work computes contextual language representations without random masking as does in BERT and maintains the deep bidirectional architecture like BERT. To compute the same sentence representation, our method shows O(n) complexity less compared to other transformer-based models with O($n^2$). To further demonstrate its superiority, computing context language representations on CPU environments is conducted, by using the embeddings from the proposed method, logistic regression shows much higher accuracy in terms of SMS classification. Moverover, the proposed method also achieves significant higher performance in semantic similarity tasks.
Optimizing compilers exploit the memory hierarchy using loop tiling and fusion, but these two transformations usually interfere with each other due to the oversight of transformations on data in memories. We present a novel composition of loop tiling and fusion in this paper. Unlike existing tiling-after-fusion algorithms that only transform computation spaces, our approach first applies rectangular/parallelogram tiling to live-out computation spaces for fitting the memory hierarchy, followed by the computation of the memory footprints required by each tile. The upwards exposed data extracted from the memory footprints are used to determine the tile shapes of intermediate computation spaces, allowing the construction of arbitrary tile shapes. Finally, our technique implements a post-tiling fusion strategy for maximizing data locality without losing tilability or parallelism of live-out computation spaces, thereby enabling storage reduction and reuse, and optimizing the memory hierarchy. We demonstrate that our approach can achieve superior performance on both CPU and GPU architectures over the state of the art by experimenting on 11 benchmarks extracted from numerous domains including neural networks, image processing, sparse matrix computation and linear algebra. Also, the results of the ResNet-50 model on an AI accelerator show that our approach can obtain 16% performance improvement.
Emerging graph neural networks (GNNs) have extended the successes of deep learning techniques against datasets like images and texts to more complex graph-structured data. By leveraging GPU accelerators, existing frameworks combine both mini-batch and sampling for effective and efficient model training on large graphs. However, this setup faces a scalability issue since loading rich vertices features from CPU to GPU through a limited bandwidth link usually dominates the training cycle. In this paper, we propose PaGraph, a system that supports general and efficient sampling-based GNN training on single-server with multi-GPU. PaGraph significantly reduces the data loading time by exploiting available GPU resources to keep frequently accessed graph data with a cache. It also embodies a lightweight yet effective caching policy that takes into account graph structural information and data access patterns of sampling-based GNN training simultaneously. Furthermore, to scale out on multiple GPUs, PaGraph develops a fast GNN-computation-aware partition algorithm to avoid cross-partition access during data parallel training and achieves better cache efficiency. Evaluations on two representative GNN models, GCN and GraphSAGE, show that PaGraph achieves up to 96.8% data loading time reductions and up to 4.8X performance speedup over the state-of-the-art baselines. Together with preprocessing optimization, PaGraph further delivers up to 16.0X end-to-end speedup.
Existing tensor compilers have proven their effectiveness in deploying deep neural networks on general-purpose hardware like CPU and GPU, but optimizing for neural processing units (NPUs) is still challenging due to the heterogeneous compute units and complicated memory hierarchy. In this paper, we present AKG, a tensor compiler for NPUs. AKG first lowers the tensor expression language to a polyhedral representation, which is used to automate the memory management of NPUs. Unlike existing approaches that resort to manually written schedules, AKG leverages polyhedral schedulers to perform a much wider class of transformations, and extends the semantics of the polyhedral representation to combine complex tiling techniques and hierarchical fusion strategies. We also implement the domain-specific optimization of convolution in AKG. Moreover, to achieve the optimal performance, we introduce complementary optimizations in code generation, which is followed by an auto-tuner. We conduct extensive experiments on benchmarks ranging from single operators to end-to-end networks. The experimental results show that AKG can obtain superior performance to both manual scheduling approaches and vendor provided libraries. We believe AKG will cast a light on the follow-up compiler works on NPUs.
Cloud auto-scaling mechanisms are typically based on reactive automation rules that scale a cluster whenever some metric, e.g., the average CPU usage among instances, exceeds a predefined threshold. Tuning these rules becomes particularly cumbersome when scaling-up a cluster involves non-negligible times to bootstrap new instances, as it happens frequently in production cloud services. To deal with this problem, we propose an architecture for auto-scaling cloud services based on the status in which the system is expected to evolve in the near future. Our approach leverages on time-series forecasting techniques, like those based on machine learning and artificial neural networks, to predict the future dynamics of key metrics, e.g., resource consumption metrics, and apply a threshold-based scaling policy on them. The result is a predictive automation policy that is able, for instance, to automatically anticipate peaks in the load of a cloud application and trigger ahead of time appropriate scaling actions to accommodate the expected increase in traffic. We prototyped our approach as an open-source OpenStack component, which relies on, and extends, the monitoring capabilities offered by Monasca, resulting in the addition of predictive metrics that can be leveraged by orchestration components like Heat or Senlin. We show experimental results using a recurrent neural network and a multi-layer perceptron as predictor, which are compared with a simple linear regression and a traditional non-predictive auto-scaling policy. However, the proposed framework allows for the easy customization of the prediction policy as needed.
Deep Q-Network (DQN) marked a major milestone for reinforcement learning, demonstrating for the first time that human-level control policies could be learned directly from raw visual inputs via reward maximization. Even years after its introduction, DQN remains highly relevant to the research community since many of its innovations have been adopted by successor methods. Nevertheless, despite significant hardware advances in the interim, DQN's original Atari 2600 experiments remain costly to replicate in full. This poses an immense barrier to researchers who cannot afford state-of-the-art hardware or lack access to large-scale cloud computing resources. To facilitate improved access to deep reinforcement learning research, we introduce a DQN implementation that leverages a novel concurrent and synchronized execution framework designed to maximally utilize a heterogeneous CPU-GPU desktop system. With just one NVIDIA GeForce GTX 1080 GPU, our implementation reduces the training time of a 200-million-frame Atari experiment from 25 hours to just 9 hours. The ideas introduced in our paper should be generalizable to a large number of off-policy deep reinforcement learning methods.
Recently, Graph Neural Networks (GNNs) have become state-of-the-art algorithms for analyzing non-euclidean graph data. However, to realize efficient GNN training is challenging, especially on large graphs. The reasons are many-folded: 1) GNN training incurs a substantial memory footprint. Full-batch training on large graphs even requires hundreds to thousands of gigabytes of memory. 2) GNN training involves both memory-intensive and computation-intensive operations, challenging current CPU/GPU platforms. 3) The irregularity of graphs can result in severe resource under-utilization and load-imbalance problems. This paper presents a GNNear accelerator to tackle these challenges. GNNear adopts a DIMM-based memory system to provide sufficient memory capacity. To match the heterogeneous nature of GNN training, we offload the memory-intensive Reduce operations to in-DIMM Near-Memory-Engines (NMEs), making full use of the high aggregated local bandwidth. We adopt a Centralized-Acceleration-Engine (CAE) to process the computation-intensive Update operations. We further propose several optimization strategies to deal with the irregularity of input graphs and improve GNNear's performance. Comprehensive evaluations on 16 GNN training tasks demonstrate that GNNear achieves 30.8$\times$/2.5$\times$ geomean speedup and 79.6$\times$/7.3$\times$(geomean) higher energy efficiency compared to Xeon E5-2698-v4 CPU and NVIDIA V100 GPU.
Graph neural networks (GNNs) are emerging as a powerful technique for modeling graph structures. Due to the sparsity of real-world graph data, GNN performance is limited by extensive sparse matrix multiplication (SpMM) operations involved in computation. While the right sparse matrix storage format varies across input data, existing deep learning frameworks employ a single, static storage format, leaving much room for improvement. This paper investigates how the choice of sparse matrix storage formats affect the GNN performance. We observe that choosing a suitable sparse matrix storage format can significantly improve the GNN training performance, but the right format depends on the input workloads and can change as the GNN iterates over the input graph. We then develop a predictive model to dynamically choose a sparse matrix storage format to be used by a GNN layer based on the input matrices. Our model is first trained offline using training matrix samples, and the trained model can be applied to any input matrix and GNN kernels with SpMM computation. We implement our approach on top of PyTorch and apply it to 5 representative GNN models running on a multi-core CPU using real-life and synthetic datasets. Experimental results show that our approach gives an average speedup of 1.17x (up to 3x) for GNN running time.
The better accuracy and efficiency trade-off has been a challenging problem in object detection. In this work, we are dedicated to studying key optimizations and neural network architecture choices for object detection to improve accuracy and efficiency. We investigate the applicability of the anchor-free strategy on lightweight object detection models. We enhance the backbone structure and design the lightweight structure of the neck, which improves the feature extraction ability of the network. We improve label assignment strategy and loss function to make training more stable and efficient. Through these optimizations, we create a new family of real-time object detectors, named PP-PicoDet, which achieves superior performance on object detection for mobile devices. Our models achieve better trade-offs between accuracy and latency compared to other popular models. PicoDet-S with only 0.99M parameters achieves 30.6% mAP, which is an absolute 4.8% improvement in mAP while reducing mobile CPU inference latency by 55% compared to YOLOX-Nano, and is an absolute 7.1% improvement in mAP compared to NanoDet. It reaches 123 FPS (150 FPS using Paddle Lite) on mobile ARM CPU when the input size is 320. PicoDet-L with only 3.3M parameters achieves 40.9% mAP, which is an absolute 3.7% improvement in mAP and 44% faster than YOLOv5s. As shown in Figure 1, our models far outperform the state-of-the-art results for lightweight object detection. Code and pre-trained models are available at https://github.com/PaddlePaddle/PaddleDetection.
The strong demand for efficient and performant deployment of Deep Learning (DL) applications prompts the rapid development of a rich DL ecosystem. To keep up with this fast advancement, it is crucial for modern DL frameworks to efficiently integrate a variety of optimized tensor algebra libraries and runtimes as their backends and generate the fastest possible executable using these backends. However, current DL frameworks require significant manual effort and expertise to integrate every new backend while failing to unleash its full potential. Given the fast-evolving nature of the DL ecosystem, this manual approach often slows down continuous innovations across different layers; it prevents hardware vendors from the fast deployment of their cutting-edge libraries, DL framework developers must repeatedly adjust their hand-coded rules to accommodate new versions of libraries, and machine learning practitioners need to wait for the integration of new technologies and often encounter unsatisfactory performance. In this paper, we propose Collage, a DL framework that offers seamless integration of DL backends. Collage provides an expressive backend registration interface that allows users to precisely specify the capability of various backends. By leveraging the specifications of available backends, Collage automatically searches for an optimized backend placement strategy for a given workload and execution environment. Our evaluation shows that Collage outperforms the best existing framework for each hardware by $1.26\times$, $1.43\times$, $1.40\times$ on average on NVIDIA's RTX 2070 GPU, V100 GPU, and Intel's Xeon 8259CL CPU, respectively. Collage has been open-sourced and deployed in Apache TVM.
Knowledge graphs (KGs) capture knowledge in the form of head--relation--tail triples and are a crucial component in many AI systems. There are two important reasoning tasks on KGs: (1) single-hop knowledge graph completion, which involves predicting individual links in the KG; and (2), multi-hop reasoning, where the goal is to predict which KG entities satisfy a given logical query. Embedding-based methods solve both tasks by first computing an embedding for each entity and relation, then using them to form predictions. However, existing scalable KG embedding frameworks only support single-hop knowledge graph completion and cannot be applied to the more challenging multi-hop reasoning task. Here we present Scalable Multi-hOp REasoning (SMORE), the first general framework for both single-hop and multi-hop reasoning in KGs. Using a single machine SMORE can perform multi-hop reasoning in Freebase KG (86M entities, 338M edges), which is 1,500x larger than previously considered KGs. The key to SMORE's runtime performance is a novel bidirectional rejection sampling that achieves a square root reduction of the complexity of online training data generation. Furthermore, SMORE exploits asynchronous scheduling, overlapping CPU-based data sampling, GPU-based embedding computation, and frequent CPU--GPU IO. SMORE increases throughput (i.e., training speed) over prior multi-hop KG frameworks by 2.2x with minimal GPU memory requirements (2GB for training 400-dim embeddings on 86M-node Freebase) and achieves near linear speed-up with the number of GPUs. Moreover, on the simpler single-hop knowledge graph completion task SMORE achieves comparable or even better runtime performance to state-of-the-art frameworks on both single GPU and multi-GPU settings.
Propositional satisfiability (SAT) is an NP-complete problem that impacts many research fields, such as planning, verification, and security. Mainstream modern SAT solvers are based on the Conflict-Driven Clause Learning (CDCL) algorithm. Recent work aimed to enhance CDCL SAT solvers using Graph Neural Networks (GNNs). However, so far this approach either has not made solving more effective, or required substantial GPU resources for frequent online model inferences. Aiming to make GNN improvements practical, this paper proposes an approach called NeuroBack, which builds on two insights: (1) predicting phases (i.e., values) of variables appearing in the majority (or even all) of the satisfying assignments are essential for CDCL SAT solving, and (2) it is sufficient to query the neural model only once for the predictions before the SAT solving starts. Once trained, the offline model inference allows NeuroBack to execute exclusively on the CPU, removing its reliance on GPU resources. To train NeuroBack, a new dataset called DataBack containing 120,286 data samples is created. NeuroBack is implemented as an enhancement to a state-of-the-art SAT solver called Kissat. As a result, it allowed Kissat to solve up to 5.2% and 7.4% more problems on two recent SAT competition problem sets, SATCOMP-2022 and SATCOMP-2023, respectively. NeuroBack therefore shows how machine learning can be harnessed to improve SAT solving in an effective and practical manner.
Outlier detection (OD) is a key learning task for finding rare and deviant data samples, with many time-critical applications such as fraud detection and intrusion detection. In this work, we propose TOD, the first tensor-based system for efficient and scalable outlier detection on distributed multi-GPU machines. A key idea behind TOD is decomposing complex OD applications into a small collection of basic tensor algebra operators. This decomposition enables TOD to accelerate OD computations by leveraging recent advances in deep learning infrastructure in both hardware and software. Moreover, to deploy memory-intensive OD applications on modern GPUs with limited on-device memory, we introduce two key techniques. First, provable quantization speeds up OD computations and reduces its memory footprint by automatically performing specific floating-point operations in lower precision while provably guaranteeing no accuracy loss. Second, to exploit the aggregated compute resources and memory capacity of multiple GPUs, we introduce automatic batching, which decomposes OD computations into small batches for both sequential execution on a single GPU and parallel execution on multiple GPUs. TOD supports a diverse set of OD algorithms. Extensive evaluation on 11 real and 3 synthetic OD datasets shows that TOD is on average 10.9x faster than the leading CPU-based OD system PyOD (with a maximum speedup of 38.9x), and can handle much larger datasets than existing GPU-based OD systems. In addition, TOD allows easy integration of new OD operators, enabling fast prototyping of emerging and yet-to-discovered OD algorithms.