Neural networks of ads systems usually take input from multiple resources, e.g., query-ad relevance, ad features and user portraits. These inputs are encoded into one-hot or multi-hot binary features, with typically only a tiny fraction of nonzero feature values per example. Deep learning models in online advertising industries can have terabyte-scale parameters that do not fit in the GPU memory nor the CPU main memory on a computing node. For example, a sponsored online advertising system can contain more than $10^{11}$ sparse features, making the neural network a massive model with around 10 TB parameters. In this paper, we introduce a distributed GPU hierarchical parameter server for massive scale deep learning ads systems. We propose a hierarchical workflow that utilizes GPU High-Bandwidth Memory, CPU main memory and SSD as 3-layer hierarchical storage. All the neural network training computations are contained in GPUs. Extensive experiments on real-world data confirm the effectiveness and the scalability of the proposed system. A 4-node hierarchical GPU parameter server can train a model more than 2X faster than a 150-node in-memory distributed parameter server in an MPI cluster. In addition, the price-performance ratio of our proposed system is 4-9 times better than an MPI-cluster solution.
We propose a server-based approach to manage a general-purpose graphics processing unit (GPU) in a predictable and efficient manner. Our proposed approach introduces a GPU server that is a dedicated task to handle GPU requests from other tasks on their behalf. The GPU server ensures bounded time to access the GPU, and allows other tasks to suspend during their GPU computation to save CPU cycles. By doing so, we address the two major limitations of the existing real-time synchronization-based GPU management approach: busy waiting and long priority inversion. We have implemented a prototype of t…
Imperial College London（帝国理工学院）；Microsoft Research（微软研究院）；Humboldt-Universität zu Berlin（柏林洪堡大学）
Deep learning models are trained on servers with many GPUs, and training must scale with the number of GPUs. Systems such as TensorFlow and Caffe2 train models with parallel synchronous stochastic gradient descent: they process a batch of training data at a time, partitioned across GPUs, and average the resulting partial gradients to obtain an updated global model. To fully utilise all GPUs, systems must increase the batch size, which hinders statistical efficiency. Users tune hyper-parameters such as the learning rate to compensate for this, which is complex and model-specific. We describe CR…
University of California Merced（加州大学默塞德分校）；Lawrence Berkeley National Laboratory（劳伦斯伯克利国家实验室）
Motivated by extreme multi-label classification applications, we consider training deep learning models over sparse data in multi-GPU servers. The variance in the number of non-zero features across training batches and the intrinsic GPU heterogeneity combine to limit accuracy and increase the time to convergence. We address these challenges with Adaptive SGD, an adaptive elastic model averaging stochastic gradient descent algorithm for heterogeneous multi-GPUs that is characterized by dynamic scheduling, adaptive batch size scaling, and normalized model merging. Instead of statically partition…
In cloud machine learning (ML) inference systems, providing low latency to end-users is of utmost importance. However, maximizing server utilization and system throughput is also crucial for ML service providers as it helps lower the total-cost-of-ownership. GPUs have oftentimes been criticized for ML inference usages as its massive compute and memory throughput is hard to be fully utilized under low-batch inference scenarios. To address such limitation, NVIDIA's recently announced Ampere GPU architecture provides features to "reconfigure" one large, monolithic GPU into multiple smaller "GPU p…
Deep neural networks (DNNs) have grown exponentially in size over the past decade, leaving only those who have massive datacenter-based resources with the ability to develop and train such models. One of the main challenges for the long tail of researchers who might have only limited resources (e.g., a single multi-GPU server) is limited GPU memory capacity compared to model size. The problem is so acute that the memory requirement of training massive DNN models can often exceed the aggregate capacity of all available GPUs on a single server; this problem only gets worse with the trend of ever…
Cloud GPU servers have become the de facto way for deep learning practitioners to train complex models on large-scale datasets. However, it is challenging to determine the appropriate cluster configuration---e.g., server type and number---for different training workloads while balancing the trade-offs in training time, cost, and model accuracy. Adding to the complexity is the potential to reduce the monetary cost by using cheaper, but revocable, transient GPU servers. In this work, we analyze distributed training performance under diverse cluster configurations using CM-DARE, a cloud-based mea…
UCLA（加州大学洛杉矶分校）；University of Wisconsin（威斯康星大学）；CMU（卡内基梅隆大学）；Google Brain（谷歌大脑）；Simon Fraser（西蒙弗雷泽大学）；Princeton University（普林斯顿大学）
A graph neural network (GNN) enables deep learning on structured graph data. There are two major GNN training obstacles: 1) it relies on high-end servers with many GPUs which are expensive to purchase and maintain, and 2) limited memory on GPUs cannot scale to today's billion-edge graphs. This paper presents Dorylus: a distributed system for training GNNs. Uniquely, Dorylus can take advantage of serverless computing to increase scalability at a low cost. The key insight guiding our design is computation separation. Computation separation makes it possible to construct a deep, bounded-asynchron…
Worcester Polytechnic Institute（伍斯特理工学院）；Institute of Software, Chinese Academy of Sciences（中国科学院软件研究所）
Distributed training frameworks, like TensorFlow, have been proposed as a means to reduce the training time of deep learning models by using a cluster of GPU servers. While such speedups are often desirable---e.g., for rapidly evaluating new model designs---they often come with significantly higher monetary costs due to sublinear scalability. In this paper, we investigate the feasibility of using training clusters composed of cheaper transient GPU servers to get the benefits of distributed training without the high costs. We conduct the first large-scale empirical analysis, launching more than…
As machine learning techniques are applied to a widening range of applications, high throughput machine learning (ML) inference servers have become critical for online service applications. Such ML inference servers pose two challenges: first, they must provide a bounded latency for each request to support consistent service-level objective (SLO), and second, they can serve multiple heterogeneous ML models in a system as certain tasks involve invocation of multiple models and consolidating multiple models can improve system utilization. To address the two requirements of ML inference servers, …
Recommendation systems are of crucial importance for a variety of modern apps and web services, such as news feeds, social networks, e-commerce, search, etc. To achieve peak prediction accuracy, modern recommendation models combine deep learning with terabyte-scale embedding tables to obtain a fine-grained representation of the underlying data. Traditional inference serving architectures require deploying the whole model to standalone servers, which is infeasible at such massive scale. In this paper, we provide insights into the intriguing and challenging inference domain of online recommendat…
Multi-accelerator servers are increasingly being deployed in shared multi-tenant environments (such as in cloud data centers) in order to meet the demands of large-scale compute-intensive workloads. In addition, these accelerators are increasingly being inter-connected in complex topologies and workloads are exhibiting a wider variety of inter-accelerator communication patterns. However, existing allocation policies are ill-suited for these emerging use-cases. Specifically, this work identifies that multi-accelerator workloads are commonly fragmented leading to reduced bandwidth and increased …
Performance Characterization of Distributed Deep Learning Strategies: A Quantitative Evaluation of DDP, FSDP, and Parameter Server Architectures on GPU Clusters
Science Information Network (SINET) is a Japanese academic backbone network for more than 800 research institutions and universities. In this paper, we present a multi-GPU-driven pipeline for handling huge session data of SINET. Our pipeline consists of ELK stack, multi-GPU server, and Splunk. A multi-GPU server is responsible for two procedures: discrimination and histogramming. Discrimination is dividing session data into ingoing/outgoing with subnet mask calculation and network address matching. Histogramming is grouping ingoing/outgoing session data into bins with map-reduce. In our archit…
The widely-adopted practice is to train deep learning models with specialized hardware accelerators, e.g., GPUs or TPUs, due to their superior performance on linear algebra operations. However, this strategy does not employ effectively the extensive CPU and memory resources -- which are used only for preprocessing, data transfer, and scheduling -- available by default on the accelerated servers. In this paper, we study training algorithms for deep learning on heterogeneous CPU+GPU architectures. Our two-fold objective -- maximize convergence rate and resource utilization simultaneously -- make…
As deep neural networks become more complex and input datasets grow larger, it can take days or even weeks to train a deep neural network to the desired accuracy. Therefore, distributed Deep Learning at a massive scale is a critical capability, since it offers the potential to reduce the training time from weeks to hours. In this paper, we present a software-hardware co-optimized distributed Deep Learning system that can achieve near-linear scaling up to hundreds of GPUs. The core algorithm is a multi-ring communication pattern that provides a good tradeoff between latency and bandwidth and ad…
Because of the superior feature representation ability of deep learning, various deep Click-Through Rate (CTR) models are deployed in the commercial systems by industrial companies. To achieve better performance, it is necessary to train the deep CTR models on huge volume of training data efficiently, which makes speeding up the training process an essential problem. Different from the models with dense training data, the training data for CTR models is usually high-dimensional and sparse. To transform the high-dimensional sparse input into low-dimensional dense real-value vectors, almost all …
Institute of Science and Technology Austria（奥地利科学技术研究所）；University of British Columbia（不列颠哥伦比亚大学）
The ability to scale out training workloads has been one of the key performance enablers of deep learning. The main scaling approach is data-parallel GPU-based training, which has been boosted by hardware and software support for highly efficient point-to-point communication, and in particular via hardware bandwidth overprovisioning. Overprovisioning comes at a cost: there is an order of magnitude price difference between "cloud-grade" servers with such support, relative to their popular "consumer-grade" counterparts, although single server-grade and consumer-grade GPUs can have similar comput…
GPUs in High-Performance Computing systems remain under-utilised due to the unavailability of schedulers that can safely schedule multiple applications to share the same GPU. The research reported in this paper is motivated to improve the utilisation of GPUs by proposing a framework, we refer to as schedGPU, to facilitate intra-node GPU co-scheduling such that a GPU can be safely shared among multiple applications by taking memory constraints into account. Two approaches, namely a client-server and a shared memory approach are explored. However, the shared memory approach is more suitable due …
Breadth-First Search (BFS) is a building block used in a wide array of graph analytics and is used in various network analysis domains: social, road, transportation, communication, and much more. Over the last two decades, network sizes have continued to grow. The popularity of BFS has brought with it a need for significantly faster traversals. Thus, BFS algorithms have been designed to exploit shared-memory and shared-nothing systems -- this includes algorithms for accelerators such as the GPU. GPUs offer extremely fast traversals at the cost of processing smaller graphs due to their limited …
Real-world node embedding applications often contain hundreds of billions of edges with high-dimension node features. Scaling node embedding systems to efficiently support these applications remains a challenging problem. In this paper we present a high-performance multi-GPU node embedding system. It uses model parallelism to split node embeddings onto each GPU's local parameter server, and data parallelism to train these embeddings on different edge samples in parallel. We propose a hierarchical data partitioning strategy and an embedding training pipeline to optimize both communication and m…
Deep learning frameworks have been widely deployed on GPU servers for deep learning applications in both academia and industry. In training deep neural networks (DNNs), there are many standard processes or algorithms, such as convolution and stochastic gradient descent (SGD), but the running performance of different frameworks might be different even running the same deep model on the same GPU hardware. In this study, we evaluate the running performance of four state-of-the-art distributed deep learning frameworks (i.e., Caffe-MPI, CNTK, MXNet, and TensorFlow) over single-GPU, multi-GPU, and m…
Distributed deep learning is becoming increasingly popular due to the expanding demand for computing resources for deep learning models with a larger amount of parameters. Different from traditional training approaches, data-parallel training allows multiple compute nodes to train large deep learning models simultaneously in order to boost the training efficiency. In this paper, we present and compare six strategies for data-parallel training using PyTorch on the language model GPT-2 with 100M parameters using a qualitative approach. These strategies are Single GPU, Single Parameter Server, Di…
Optimizing Resource Allocation for Geographically-Distributed Inference by Large Language Models
Department of Computer Science and Engineering, Pennsylvania State University, University Park, PA, USA（计算机科学与工程系，宾夕法尼亚州立大学，大学公园，PA，USA）；Department of Computer Science, Virginia Tech, Blacksburg, VA, USA（计算机科学系，弗吉尼亚理工大学，布莱克堡，VA，USA）；Department of Electrical Communication Engineering, Indian Institute of Science, Bangalore, India（电子通信工程系，印度科学研究院，班加罗尔，印度）
High performance multi-GPU computing becomes an inevitable trend due to the ever-increasing demand on computation capability in emerging domains such as deep learning, big data and planet-scale simulations. However, the lack of deep understanding on how modern GPUs can be connected and the real impact of state-of-the-art interconnect technology on multi-GPU application performance become a hurdle. In this paper, we fill the gap by conducting a thorough evaluation on five latest types of modern GPU interconnects: PCIe, NVLink-V1, NVLink-V2, NVLink-SLI and NVSwitch, from six high-end servers and…
Hash tables are used in a plethora of applications, including database operations, DNA sequencing, string searching, and many more. As such, there are many parallelized hash tables targeting multicore, distributed, and accelerator-based systems. We present in this work a multi-GPU hash table implementation that can process keys at a throughput comparable to that of distributed hash tables. Distributed CPU hash tables have received significantly more attention than GPU-based hash tables. We show that a single node with multiple GPUs offers roughly the same performance as a 500-1,000-core CPU-ba…
Most work in the deep learning systems community has focused on faster inference, but arriving at a trained model requires lengthy experiments. Accelerating training lets developers iterate faster and come up with better models. DNN training is often seen as a compute-bound problem, best done in a single large compute node with many GPUs. As DNNs get bigger, training requires going distributed. Distributed deep neural network (DDNN) training constitutes an important workload on the cloud. Larger DNN models and faster compute engines shift the training performance bottleneck from computation to…
Hash tables are ubiquitous. Properties such as an amortized constant time complexity for insertion and querying as well as a compact memory layout make them versatile associative data structures with manifold applications. The rapidly growing amount of data emerging in many fields motivated the need for accelerated hash tables designed for modern parallel architectures. In this work, we exploit the fast memory interface of modern GPUs together with a parallel hashing scheme tailored to improve global memory access patterns, to design WarpCore -- a versatile library of hash table data structure…
Collaborative deep learning inference between low-resource endpoint devices and edge servers has received significant research interest in the last few years. Such computation partitioning can help reducing endpoint device energy consumption and improve latency, but equally importantly also contributes to privacy-preserving of sensitive data. This paper describes Edge-PRUNE, a flexible but light-weight computation framework for distributing machine learning inference between edge servers and one or more client devices. Compared to previous approaches, Edge-PRUNE is based on a formal dataflow c…
