Abstract

As the models and the datasets to train deep learning (DL) models scale, system architects are faced with new challenges, one of which is the memory capacity bottleneck, where the limited physical memory inside the accelerator device constrains the algorithm that can be studied. We propose a memory-centric deep learning system that can transparently expand the memory capacity available to the accelerators while also providing fast inter-device communication for parallel training. Our proposal aggregates a pool of memory modules locally within the device-side interconnect, which are decoupled from the host interface and function as a vehicle for transparent memory capacity expansion. Compared to conventional systems, our proposal achieves an average 2.8x speedup on eight DL applications and increases the system-wide memory capacity to tens of TBs.

Abstract

Internet video streaming has experienced tremendous growth over the last few decades. However, the quality of existing video delivery critically depends on the bandwidth resource. Consequently, user quality of experience (QoE) suffers inevitably when network conditions become unfavorable. We present a new video delivery framework that utilizes client computation and recent advances in deep neural networks (DNNs) to reduce the dependency for delivering high-quality video. The use of DNNs enables us to enhance the video quality independent to the available bandwidth. We design a practical system that addresses several challenges, such as client heterogeneity, interaction with bitrate adaptation, and DNN transfer, in enabling the idea. Our evaluation using 3G and broadband network traces shows the proposed system outperforms the current state of the art, enhancing the average QoE by 43.08% using the same bandwidth budget or saving 17.13% of bandwidth while providing the same user QoE.

Abstract

We are approaching a fundamental shift in the computational era as the number of smart device users (e.g., smartphone and tablet users) is expected to exceed 6 billion (more than 50% of the global population) by 2020. To cope with the explosion of mobile devices coupled with a growing proliferation of their corresponding applications, best-effort quality-of-service (QoS) is no longer a satisfactory solution and a new breed of intelligent networks is required. We argue that both network wireless-edges and end-devices, which construct Network Edge, need to become more intelligent with respect to networking in order to facilitate a quality user. We believe that an SDN-like paradigm needs to be pushed to wireless-edges and mobile clients (ExtremeSDN) to provide optimal network performance between the cloud and wirelessly connected clients. In this talk, I will present SMILE - SMart and Intelligent wireLess Edge framework that supports SDN-like paradigm at user smart devices and network wireless-edges. SMILE enables network wireless-edges to become more active and to host several services (including partial cloud services) to enhance users quality of experience. Building on top of SMILE framework, I will present two edge-based services: 1) FlexStream - an edge-based flexible adaptive video streaming framework that allows fine-grained management of bandwidth based on real-time context-awareness and specified policy, and 2) PrivacyGuard is a light-weight programmable security framework that enables flexible policies to protect the wireless network communication of sensitive IoT and mobile applications.

Bio

Tamer Nadeem is an associate professor and the founder of Mobile Systems and Intelligent Communication (MuSIC) Lab (https://music.lab.vcu.edu/) in the Department of Computer Science at Virginia Commonwealth University (VCU). Before that, he spent a few years at Old Dominion University (ODU) and Siemens Corporate Research (SCR). His research interests cover several aspects of wireless networking and mobile computing systems including smart wireless systems, mobile & edge computing, software-defined networks, network security and privacy, Internet-of-things & smart city systems, vehicular networks, and intelligent transportation systems. Dr. Nadeem research is funded by several federal agencies and industries including National Science Foundation (NSF), National Institute of Standards and Technology (NIST), Federal Highway Administration (FHWA), Siemens Corporate Research, AT&T Labs, Microsoft, Nokia-Bell Labs, and Google. Dr. Nadeem holds 5 US patents has over 70 publications in peer-reviewed top scholarly journals and conference proceedings. Dr. Nadeem serves as a member of the technical and organizing committees of various ACM and IEEE conferences. He is serving as an associate editor of IET Communications journal and as a guest editor of multiple journals. In addition, he is serving as program chair of the 15th International Wireless Communications & Mobile Computing Conference (IWCMC 2019) - Mobile Computing Symposium and as program co-chair of The Third International Workshop on Smart Edge Computing and Networking (SmartEdge 2019), in conjunction with PerCom 2019.

Abstract

Cloud computing is becoming increasingly popular. Particularly, the public cloud and commercial virtual desktop infrastructure (VDI) are widely used. It is thus becoming increasingly important to understand their workloads and network characteristics on cloud networks. Such an understanding can improve the cloud systems and help to design the next generation of cloud systems. However, little is known about workloads on the VDI and the impact of software-based virtual network on latency in the public cloud. In this talk, we cover the topics related to both the workloads on the VDI and the latency in the public cloud. Our analysis results, discussions, and implications can not only help cloud researchers and developers design the next generation of cloud systems but can also help cloud operators improve the performance of cloud system.

Bio

Chunghan Lee received the Ph.D. degree in the Dept. of Electronic and Information Engineering from Toyohashi University of Technology in 2013. He is currently a researcher at Fujitsu Laboratories, Japan. He developed on SDN controller based on GBP to control both virtual and legacy physical switches, and focused on cloud system analysis in the wild. He has been contributing to both academic and industrial sides, such as SNIA, OpenDaylight, and OPNFV communities. His research interests include NFV/SDN, cloud system monitoring, network measurement, and virtualization technology.

Abstract

Minimizing response time of mobile applications is critical for user experience. Existing work predominantly focuses on reducing mobile Web latency, whereas users spend more time on native mobile apps than mobile Web. Similar to Web, mobile apps contain a chain of dependencies between successive requests. However, unlike Web acceleration where object dependencies can easily be identified by parsing Web documents, App acceleration is much more difficult because the dependency is encoded in the app binary. Motivated by recent advances in program analysis, this paper presents a system that utilizes static program analysis to automatically generate acceleration proxies for mobile apps. Our framework takes Android app binary as input, performs program analysis to identify resource dependencies, and outputs an acceleration proxy that performs dynamic prefetching. Our evaluation using a user study from 30 participants and an in-depth analysis of popular commercial apps shows that an acceleration proxy reduces the median user-perceived latency by up to 64% (1,471 ms).