More Like this

1. Stash: A comprehensive stall-centric characterization of public cloud VMs for distributed deep learning

   Sharma, Aakash ; Bhasi, Vivek ; Singh, Sonali ; Jain, Rishabh ; Raj, Jashwant ; Mitra, Subrata ; Kandemir, Mahmut Taylan ; Kesidis, George ; Das, Chita ( January 2023 , Proceedings of the International Conference on Distributed Computing Systems)

   Deep neural networks (DNNs) are increasingly popular owing to their ability to solve complex problems such as image recognition, autonomous driving, and natural language processing. Their growing complexity coupled with the use of larger volumes of training data (to achieve acceptable accuracy) has warranted the use of GPUs and other accelerators. Such accelerators are typically expensive, with users having to pay a high upfront cost to acquire them. For infrequent use, users can, instead, leverage the public cloud to mitigate the high acquisition cost. However, with the wide diversity of hardware instances (particularly GPU instances) available in public cloud, it becomes challenging for a user to make an appropriate choice from a cost/performance standpoint. In this work, we try to address this problem by (i) introducing a comprehensive distributed deep learning (DDL) profiler Stash, which determines the various execution stalls that DDL suffers from, and (ii) using Stash to extensively characterize various public cloud GPU instances by running popular DNN models on them. Specifically, it estimates two types of communication stalls, namely, interconnect and network stalls, that play a dominant role in DDL execution time. Stash is implemented on top of prior work, DS-analyzer, that computes only the CPU and disk stalls. Using our detailed stall characterization, we list the advantages and shortcomings of public cloud GPU instances for users to help them make an informed decision(s). Our characterization results indicate that the more expensive GPU instances may not be the most performant for all DNN models and that AWS can sometimes sub-optimally allocate hardware interconnect resources. Specifically, the intra-machine interconnect can introduce communication overheads of up to 90% of DNN training time and the network-connected instances can suffer from up to 5× slowdown compared to training on a single instance. Furthermore, (iii) we also model the impact of DNN macroscopic features such as the number of layers and the number of gradients on communication stalls, and finally, (iv) we briefly discuss a cost comparison with existing work. 

   more » « less
2. MArk: Exploiting Cloud Services for Cost-Effective, SLO-Aware Machine Learning Inference Serving

   Zhang, Chengliang ; Yu, Minchen ; Wang, Wei ; Yan, Feng ( May 2019 , Proceedings of the USENIX Conference)

   The advances of Machine Learning (ML) have sparked a growing demand of ML-as-a-Service: developers train ML models and publish them in the cloud as online services to provide low-latency inference at scale. The key challenge of ML model serving is to meet the response-time Service-Level Objectives (SLOs) of inference workloads while minimizing the serving cost. In this paper, we tackle the dual challenge of SLO compliance and cost effectiveness with MArk (Model Ark), a general-purpose inference serving system built in Amazon Web Services (AWS). MArk employs three design choices tailor-made for inference workload. First, MArk dynamically batches requests and opportunistically serves them using expensive hardware accelerators (e.g., GPU) for improved performance-cost ratio. Second, instead of relying on feedback control scaling or over-provisioning to serve dynamic workload, which can be too slow or too expensive for inference serving, MArk employs predictive autoscaling to hide the provisioning latency at low cost. Third, given the stateless nature of inference serving, MArk exploits the flexible, yet costly serverless instances to cover the occasional load spikes that are hard to predict. We evaluated the performance of MArk using several state-of-the-art ML models trained in popular frameworks including TensorFlow, MXNet, and Keras. Compared with the premier industrial ML serving platform SageMaker, MArk reduces the serving cost up to 7.8× while achieving even better latency performance. 

   more » « less
3. AccPar: Tensor Partitioning for Heterogeneous Deep Learning Accelerators

   https://doi.org/10.1109/HPCA47549.2020.00036  

   Song, Linghao ; Chen, Fan ; Zhuo, Youwei ; Qian, Xuehai ; Li, Hai ; Chen, Yiran ( February 2020 , 2020 IEEE International Symposium on High Performance Computer Architecture (HPCA))

   Deep neural network (DNN) accelerators as an example of domain-specific architecture have demonstrated great success in DNN inference. However, the architecture acceleration for equally important DNN training has not yet been fully studied. With data forward, error backward and gradient calculation, DNN training is a more complicated process with higher computation and communication intensity. Because the recent research demonstrates a diminishing specialization return, namely, “accelerator wall”, we believe that a promising approach is to explore coarse-grained parallelism among multiple performance-bounded accelerators to support DNN training. Distributing computations on multiple heterogeneous accelerators to achieve high throughput and balanced execution, however, remaining challenging. We present ACCPAR, a principled and systematic method of determining the tensor partition among heterogeneous accelerator arrays. Compared to prior empirical or unsystematic methods, ACCPAR considers the complete tensor partition space and can reveal previously unknown new parallelism configurations. ACCPAR optimizes the performance based on a cost model that takes into account both computation and communication costs of a heterogeneous execution environment. Hence, our method can avoid the drawbacks of existing approaches that use communication as a proxy of the performance. The enhanced flexibility of tensor partitioning in ACCPAR allows the flexible ratio of computations to be distributed among accelerators with different performances. The proposed search algorithm is also applicable to the emerging multi-path patterns in modern DNNs such as ResNet. We simulate ACCPAR on a heterogeneous accelerator array composed of both TPU-v2 and TPU-v3 accelerators for the training of large-scale DNN models such as Alexnet, Vgg series and Resnet series. The average performance improvements of the state-of-the-art “one weird trick” (OWT) and HYPAR, and ACCPAR, normalized to the baseline data parallelism scheme where each accelerator replicates the model and processes different input data in parallel, are 2.98×, 3.78×, and 6.30×, respectively. 

   more » « less
4. Understanding Reuse, Performance, and Hardware Cost of DNN Dataflow: A Data-Centric Approach

   https://doi.org/10.1145/3352460.3358252  

   Kwon, Hyoukjun ; Chatarasi, Prasanth ; Pellauer, Michael ; Parashar, Angshuman ; Sarkar, Vivek ; Krishna, Tushar ( October 2019 , MICRO '52: Proceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture)

   null (Ed.)

   The data partitioning and scheduling strategies used by DNN accelerators to leverage reuse and perform staging are known as dataflow, which directly impacts the performance and energy efficiency of DNN accelerators. An accelerator micro architecture dictates the dataflow(s) that can be employed to execute layers in a DNN. Selecting a dataflow for a layer can have a large impact on utilization and energy efficiency, but there is a lack of understanding on the choices and consequences of dataflow, and of tools and methodologies to help architects explore the co-optimization design space. In this work, we first introduce a set of data-centric directives to concisely specify the DNN dataflow space in a compiler-friendly form. We then show how these directives can be analyzed to infer various forms of reuse and to exploit them using hardware capabilities. We codify this analysis into an analytical cost model, MAESTRO (Modeling Accelerator Efficiency via Patio-Temporal Reuse and Occupancy), that estimates various cost-benefit tradeoffs of a dataflow including execution time and energy efficiency for a DNN model and hardware configuration. We demonstrate the use of MAESTRO to drive a hardware design space exploration experiment, which searches across 480M designs to identify 2.5M valid designs at an average rate of 0.17M designs per second, including Pareto-optimal throughput- and energy-optimized design points. 

   more » « less
5. Primitives Enhancing GPU Runtime Support for improved DNN Performance

   Dhakal, Aditya ; Kulkarni, Sameer ; Ramakrishnan, K. K. ( September 2021 , IEEE International Conference on Cloud Computing)

   null (Ed.)

   Deep neural networks (DNNs) are increasingly used for real-time inference, requiring low latency, but require significant computational power as they continue to increase in complexity. Edge clouds promise to offer lower latency due to their proximity to end-users and having powerful accelerators like GPUs to provide the computation power needed for DNNs. But it is also important to ensure that the edge-cloud resources are utilized well. For this, multiplexing several DNN models through spatial sharing of the GPU can substantially improve edge-cloud resource usage. Typical GPU runtime environments have significant interactions with the CPU, to transfer data to the GPU, for CPU-GPU synchronization on inference task completions, etc. These result in overheads. We present a DNN inference framework with a set of software primitives that reduce the overhead for DNN inference, increase GPU utilization and improve performance, with lower latency and higher throughput. Our first primitive uses the GPU DMA effectively, reducing the CPU cycles spent to transfer the data to the GPU. A second primitive uses asynchronous ‘events’ for faster task completion notification. GPU runtimes typically preclude fine-grained user control on GPU resources, causing long GPU downtimes when adjusting resources. Our third primitive supports overlapping of model-loading and execution, thus allowing GPU resource re-allocation with very little GPU idle time. Our other primitives increase inference throughput by improving scheduling and processing more requests. Overall, our primitives decrease inference latency by more than 35% and increase DNN throughput by 2-3×. 

   more » « less