Generative image editing using diffusion models has become a prevalent application in today's AI cloud services. In production environments, image editing typically involves a mask that specifies the regions of an image template to be edited. The use of mask provides direct control over the editing process and introduces sparsity in the model inference. In this paper, we present FlashPS, a system that efficiently serves image editing requests. The key insight behind FlashPS is that image editing only modifies the masked regions of image templates, while preserving the original content in the unmasked areas. Driven by this insight, FlashPS judiciously skips redundant computations associated with the unmask areas by reusing cached intermediate activations from previous inferences. To mitigate the high cache loading overhead, FlashPS employs a bubble-free pipeline scheme that overlaps computation with cache loading. Additionally, to reduce queuing latency in online serving while improving the GPU utilization, FlashPS proposes a novel continuous batching strategy for diffusion model serving, allowing newly arrived requests to join the running batch in just one step of denoising computation, without waiting for the entire batch to complete. As heterogenous masks induce imbalanced load, FlashPS also develops a load balancing strategy that takes into account the loads of both computation and cache loading. Collectively, FlashPS outperforms state-of-the-art diffusion serving systems for image editing, achieving up to 3× higher throughput and reducing average request latency by up to 14.7× while ensuring image quality.

The increase in radio link capacities has brought significant growth in number and variety of mobile applications. To support these applications, network providers and researchers need an up-to- date, thorough understanding of the composition of 4G LTE traffic. In this paper, we conduct a deep study over a tier-1 U.S. wireless carrier to measure and analyze 4G LTE traffic. We collect data on 2.5+ million flows from 5000+ user devices, analyzing cross-layer characteristics for IPv4/6, HTTP/S and newly emerging Quick UDP Internet Connection (QUIC) traffic, with analysis extracted for the most prolific domains. To the best of our knowledge, our QUIC traffic analysis is the first published from a carrier's perspective. Results show data from multimedia sources (audio, video and images) dominate link capacities, with video, in particular, being the "king of the hill". Detailed analysis suggests methods for identifying encrypted video traffic, potentially leading to improved services that can provide network treatments for video.

With the development of wireless communication network, Vehicular Ad-hoc Network (VANET) has received considerable attentions. US Department of Transportation (USDOT) is piloting a deployment concept to encourage embedding and retrofitting vehicles with DSRC interface to support varieties of applications for both safety and non-safety purposes. In addition to traditional data service, many real-time video streaming applications are increasingly being developed over VANET. Although it is desirable to delivery the packets directly through single hop communications, such condition is not always available in the real environment. To support real-time video applications in VANET, we have to deeply understand their performance under multihop conditions and even with others' interference. To the best of our knowledge, no work has been done on measuring the multihop video streaming transmission performance with interference over real environment. In this paper, a DSRC device based test-bed is developed to investigate the performance of multihop real-time video streaming transmission and the impact of interference. To overcome the high packet loss rate which is a main shortage of video streaming transmission over VANET, we improve the application by adding retransmission and startup caching schemes. We also evaluate the performance of different retransmission strategies and analyze how to optimize the retransmission scheme for minimizing the number of collided and late packets based on the transmission rate of interference.