Y. Charlie Hu

Sibendu Paul, Kunal Rao, Giuseppe Coviello et al.

It is a common practice to think of a video as a sequence of images (frames), and re-use deep neural network models that are trained only on images for similar analytics tasks on videos. In this paper, we show that this leap of faith that deep learning models that work well on images will also work well on videos is actually flawed. We show that even when a video camera is viewing a scene that is not changing in any human-perceptible way, and we control for external factors like video compression and environment (lighting), the accuracy of video analytics application fluctuates noticeably. These fluctuations occur because successive frames produced by the video camera may look similar visually, but these frames are perceived quite differently by the video analytics applications. We observed that the root cause for these fluctuations is the dynamic camera parameter changes that a video camera automatically makes in order to capture and produce a visually pleasing video. The camera inadvertently acts as an unintentional adversary because these slight changes in the image pixel values in consecutive frames, as we show, have a noticeably adverse impact on the accuracy of insights from video analytics tasks that re-use image-trained deep learning models. To address this inadvertent adversarial effect from the camera, we explore the use of transfer learning techniques to improve learning in video analytics tasks through the transfer of knowledge from learning on image analytics tasks. In particular, we show that our newly trained Yolov5 model reduces fluctuation in object detection across frames, which leads to better tracking of objects(40% fewer mistakes in tracking). Our paper also provides new directions and techniques to mitigate the camera's adversarial effect on deep learning models used for video analytics applications.

Sibendu Paul, Kunal Rao, Giuseppe Coviello et al.

IoT sensors, especially video cameras, are ubiquitously deployed around the world to perform a variety of computer vision tasks in several verticals including retail, healthcare, safety and security, transportation, manufacturing, etc. To amortize their high deployment effort and cost, it is desirable to perform multiple video analytics tasks, which we refer to as Analytical Units (AUs), off the video feed coming out of every camera. In this paper, we first show that in a multi-AU setting, changing the camera setting has disproportionate impact on different AUs performance. In particular, the optimal setting for one AU may severely degrade the performance for another AU, and further the impact on different AUs varies as the environmental condition changes. We then present Elixir, a system to enhance the video stream quality for multiple analytics on a video stream. Elixir leverages Multi-Objective Reinforcement Learning (MORL), where the RL agent caters to the objectives from different AUs and adjusts the camera setting to simultaneously enhance the performance of all AUs. To define the multiple objectives in MORL, we develop new AU-specific quality estimator values for each individual AU. We evaluate Elixir through real-world experiments on a testbed with three cameras deployed next to each other (overlooking a large enterprise parking lot) running Elixir and two baseline approaches, respectively. Elixir correctly detects 7.1% (22,068) and 5.0% (15,731) more cars, 94% (551) and 72% (478) more faces, and 670.4% (4975) and 158.6% (3507) more persons than the default-setting and time-sharing approaches, respectively. It also detects 115 license plates, far more than the time-sharing approach (7) and the default setting (0).

Sibendu Paul, Kunal Rao, Giuseppe Coviello et al.

Cameras are increasingly being deployed in cities, enterprises and roads world-wide to enable many applications in public safety, intelligent transportation, retail, healthcare and manufacturing. Often, after initial deployment of the cameras, the environmental conditions and the scenes around these cameras change, and our experiments show that these changes can adversely impact the accuracy of insights from video analytics. This is because the camera parameter settings, though optimal at deployment time, are not the best settings for good-quality video capture as the environmental conditions and scenes around a camera change during operation. Capturing poor-quality video adversely affects the accuracy of analytics. To mitigate the loss in accuracy of insights, we propose a novel, reinforcement-learning based system APT that dynamically, and remotely (over 5G networks), tunes the camera parameters, to ensure a high-quality video capture, which mitigates any loss in accuracy of video analytics. As a result, such tuning restores the accuracy of insights when environmental conditions or scene content change. APT uses reinforcement learning, with no-reference perceptual quality estimation as the reward function. We conducted extensive real-world experiments, where we simultaneously deployed two cameras side-by-side overlooking an enterprise parking lot (one camera only has manufacturer-suggested default setting, while the other camera is dynamically tuned by APT during operation). Our experiments demonstrated that due to dynamic tuning by APT, the analytics insights are consistently better at all times of the day: the accuracy of object detection video analytics application was improved on average by ~ 42%. Since our reward function is independent of any analytics task, APT can be readily used for different video analytics tasks.

Omar Basit, Yunzhao Liu, Z. Jonny Kong et al.

Prefill/decode disaggregation is increasingly adopted in LLM serving to improve the latency-throughput tradeoff and meet strict TTFT and TPOT SLOs. However, LLM inference remains energy-hungry: autoscaling alone is too coarse-grained to track fast workload fluctuations, and applying fine-grained DVFS under disaggregation is complicated by phase-asymmetric dynamics and coupling between provisioning and frequency control. We present DualScale, a two-tier energy optimization framework for disaggregated LLM serving. DualScale jointly optimizes placement and DVFS across prefill and decode using predictive latency and power models. At coarse timescales, DualScale computes phase-aware placement and baseline frequencies that minimize energy while satisfying SLO constraints. At fine timescales, DualScale dynamically adapts GPU frequency per iteration using stage-specific control: model predictive control (MPC) for prefill to account for queue evolution and future TTFT impact, and lightweight slack-aware adaptation for decode to exploit its smoother, memory-bound dynamics. This hierarchical design enables coordinated control across timescales while preserving strict serving SLOs. Evaluation on a 16x H100 cluster serving Llama 3.3 70B with production-style traces shows that DualScale meets TTFT/TPOT SLOs while reducing energy by up to 39% in prefill and 48% in decode relative to DistServe.

Zongpu Zhang, Pranab Dash, Y. Charlie Hu et al.

Large Language Models (LLMs) are increasingly being integrated into various applications and services running on billions of mobile devices. However, deploying LLMs on resource-limited mobile devices faces a significant challenge due to their high demand for computation, memory, and ultimately energy. While current LLM frameworks for mobile use three power-hungry components-CPU, GPU, and Memory-even when running primarily-GPU LLM models, optimized DVFS governors for CPU, GPU, and memory featured in modern mobile devices operate independently and are oblivious of each other. Motivated by the above observation, in this work, we first measure the energy-efficiency of a SOTA LLM framework consisting of various LLM models on mobile phones which showed the triplet mobile governors result in up to 40.4% longer prefilling and decoding latency compared to optimal combinations of CPU, GPU, and memory frequencies with the same energy consumption for sampled prefill and decode lengths. Second, we conduct an in-depth measurement study to uncover how the intricate interplay (or lack of) among the mobile governors cause the above inefficiency in LLM inference. Finally, based on these insights, we design FUSE - a unified energy-aware governor for optimizing the energy efficiency of LLM inference on mobile devices. Our evaluation using a ShareGPT dataset shows FUSE reduces the time-to-first-token and time-per-output-token latencies by 7.0%-16.9% and 25.4%-36.8% on average with the same energy-per-token for various mobile LLM models.

Sibendu Paul, Kunal Rao, Giuseppe Coviello et al.

In Video Analytics Pipelines (VAP), Analytics Units (AUs) such as object detection and face recognition running on remote servers critically rely on surveillance cameras to capture high-quality video streams in order to achieve high accuracy. Modern IP cameras come with a large number of camera parameters that directly affect the quality of the video stream capture. While a few of such parameters, e.g., exposure, focus, white balance are automatically adjusted by the camera internally, the remaining ones are not. We denote such camera parameters as non-automated (NAUTO) parameters. In this paper, we first show that environmental condition changes can have significant adverse effect on the accuracy of insights from the AUs, but such adverse impact can potentially be mitigated by dynamically adjusting NAUTO camera parameters in response to changes in environmental conditions. We then present CamTuner, to our knowledge, the first framework that dynamically adapts NAUTO camera parameters to optimize the accuracy of AUs in a VAP in response to adverse changes in environmental conditions. CamTuner is based on SARSA reinforcement learning and it incorporates two novel components: a light-weight analytics quality estimator and a virtual camera that drastically speed up offline RL training. Our controlled experiments and real-world VAP deployment show that compared to a VAP using the default camera setting, CamTuner enhances VAP accuracy by detecting 15.9% additional persons and 2.6%-4.2% additional cars (without any false positives) in a large enterprise parking lot and 9.7% additional cars in a 5G smart traffic intersection scenario, which enables a new usecase of accurate and reliable automatic vehicle collision prediction (AVCP). CamTuner opens doors for new ways to significantly enhance video analytics accuracy beyond incremental improvements from refining deep-learning models.

Sibendu Paul, Utsav Drolia, Y. Charlie Hu et al.

Millions of cameras at edge are being deployed to power a variety of different deep learning applications. However, the frames captured by these cameras are not always pristine - they can be distorted due to lighting issues, sensor noise, compression etc. Such distortions not only deteriorate visual quality, they impact the accuracy of deep learning applications that process such video streams. In this work, we introduce AQuA, to protect application accuracy against such distorted frames by scoring the level of distortion in the frames. It takes into account the analytical quality of frames, not the visual quality, by learning a novel metric, classifier opinion score, and uses a lightweight, CNN-based, object-independent feature extractor. AQuA accurately scores distortion levels of frames and generalizes to multiple different deep learning applications. When used for filtering poor quality frames at edge, it reduces high-confidence errors for analytics applications by 17%. Through filtering, and due to its low overhead (14ms), AQuA can also reduce computation time and average bandwidth usage by 25%.

Qiang Xu, James C. Davis, Y. Charlie Hu et al.

Improving software performance through configuration parameter tuning is a common activity during software maintenance. Beyond traditional performance metrics like latency, mobile app developers are interested in reducing app energy usage. Some mobile apps have centralized locations for parameter tuning, similar to databases and operating systems, but it is common for mobile apps to have hundreds of parameters scattered around the source code. The correlation between these "deep" parameters and app energy usage is unclear. Researchers have studied the energy effects of deep parameters in specific modules, but we lack a systematic understanding of the energy impact of mobile deep parameters. In this paper we empirically investigate this topic, combining a developer survey with systematic energy measurements. Our motivational survey of 25 Android developers suggests that developers do not understand, and largely ignore, the energy impact of deep parameters. To assess the potential implications of this practice, we propose a deep parameter energy profiling framework that can analyze the energy impact of deep parameters in an app. Our framework identifies deep parameters, mutates them based on our parameter value selection scheme, and performs reliable energy impact analysis. Applying the framework to 16 popular Android apps, we discovered that deep parameter-induced energy inefficiency is rare. We found only 2 out of 1644 deep parameters for which a different value would significantly improve its app's energy efficiency. A detailed analysis found that most deep parameters have either no energy impact, limited energy impact, or an energy impact only under extreme values. Our study suggests that it is generally safe for developers to ignore the energy impact when choosing deep parameter values in mobile apps.