Cloud service providers heavily colocate high-priority, latency sensitive (LS), and low-priority, best-effort (BE) DNN inference services on the same GPU to improve resource utilization in data centers. Among the critical shared GPU resources, there has been very limited analysis on the dynamic allocation of compute units and VRAM bandwidth, mainly for two reasons: (1) The native GPU resource management solutions are either hardware-specific, or unable to dynamically allocate resources to different tenants, or both; (2) NVIDIA doesn't expose interfaces for VRAM bandwidth allocation, and the software stack and VRAM channel architectures are black-box, both of which limit the software-level resource management. These drive prior work to design either conservative sharing policies detrimental to throughput, or static resource partitioning only applicable to a few GPU models.

The rapid expansion of video streaming content in our daily lives has rendered the real-time processing and analysis of these video streams a critical capability. However, existing deep video analytics systems primarily support only simple queries, such as selection and aggregation. Considering the inherent temporal nature of video streams, queries capable of matching patterns of events could enable a wider range of applications. In this paper, we present Bobsled, a novel video stream processing system designed to efficiently support complex event queries. Experimental results demonstrate that Bobsled can achieve a throughput improvement over state-of-the-art ranging from 2.4× to 11.6×, without any noticeable loss in accuracy.

Model quantization has been a popular approach to trade accuracy for speed during model serving, especially since new models are getting bigger and bigger. Traditionally, low-precision quantized models have better inference speed than their high-precision counterparts. However, the story may change with the advent of new machine learning instructions in modern processors. In this paper, we make a case for that using ARM processors. Our quantitative analysis shows that low-precision quantized models can be inferior in both accuracy and inference speed on modern processors. In response, we present, MiL, a new quantized neural network package. Operators in MiL are optimized for inferences based on the newly available machine learning instructions on the target platform. Experiments show that serving neural networks using PyTorch with MiL outperforms all state-of-the-art, including Riptide, TFLite with Ruy, and PyTorch with QNNPACK.

The impressive accuracy of deep neural networks (DNNs) has created great demands on practical analytics over video data. Although efficient and accurate, the latest video analytic systems have not supported analytics beyond selection and aggregation queries. In data analytics, Top-K is a very important analytical operation that enables analysts to focus on the most important entities. In this demonstration, we present Everest, the first system that supports efficient and accurate Top-K video analytics. Everest ranks and identifies the most interesting frames/clips from videos with probabilistic guarantees. Furthermore, it supports user-defined functions to rank frames/clips based on different semantics using different deep vision models. Everest leverages techniques from computer vision, uncertain databases, and Top-K query processing to return results quickly.

The impressive accuracy of deep neural networks (DNNs) has created great demands on practical analytics over video data. Although efficient and accurate, the latest video analytic systems have not supported analytics beyond selection and aggregation queries. In data analytics, Top-K is a very important analytical operation that enables analysts to focus on the most important entities. In this paper, we present Everest, the first system that supports efficient and accurate Top-K video analytics. Everest ranks and identifies the most interesting frames/moments from videos with probabilistic guarantees. Everest is a system built with a careful synthesis of deep computer vision models, uncertain data management, and Top-K query processing. Evaluations on real-world videos and the latest Visual Road benchmark show that Everest achieves between 14.3x to 20.6x higher efficiency than baseline approaches with high result accuracy.

In this paper, we explore the idea of distilling small networks for object detection task. More specifically, we propose a two-stage approach to learn more compact and efficient detectors under the single-shot object detection framework by leveraging knowledge distillation. During the 1st stage, we learn the feature maps of the student model for each of the prediction head from the teacher model. Instead of fitting the whole feature map directly, here we propose the mask guided structure including not only the entire feature map (i.e. global features) but also region features covered by the object (i.e. local features), which can significantly improve the performance of the student network. For the 2nd stage, the ground-truth is used to further refine the performance. Experimental results on PASCAL VOC and KITTI dataset demonstrate the effectiveness of our proposed approach. We achieve 56.88% mAP on VOC2007 at 143 FPS with the backbone of 1/8 VGG16.

Object detection and instance recognition play a central role in many AI applications like autonomous driving, video surveillance and medical image analysis. However, training object detection models on large scale datasets remains computationally expensive and time consuming. This paper presents an efficient and open source object detection framework called SimpleDet which enables the training of state-of-the-art detection models on consumer grade hardware at large scale. SimpleDet covers a wide range of models including both high-performance and high-speed ones. SimpleDet is well-optimized for both low precision training and distributed training and achieves 70% higher throughput for the Mask R-CNN detector compared with existing frameworks. Codes, examples and documents of SimpleDet can be found at https://github.com/tusimple/simpledet.