Mixed reality (MR) telepresence provides a shared common 3D space for networked users to enjoy real-time immersive experiences with natural interactivity and movement. Due to its high hardware and compute demands, a large-scale cloud-edge MR solution is necessary for achieving critical mass adoption, where users engage in the MR environment using low-cost cameras at the edge and complex 3D world processing is offloaded to the cloud servers. However, cloud-edge MR solutions have distinct challenges such as concurrent multiple camera support, unpredictable compute demands, and CPU/GPU contention. In this paper, we present PointPresence, an edge-compute framework for large-scale MR applications. PointPresence is deployed at an edge node, such as an O-RAN RIC or a local enterprise server, and provides low-latency MR experiences to a large number of users by incorporating intelligent camera selection for compute reduction, reactive pipeline design for adaptation to compute demand changes, and context-aware GPU sharing. Our comprehensive evaluation on an MR testbed shows that PointPresence reduces end-to-end latency by up to 3.5× and improves end user perceived visual quality by 30%.

Remote Direct Memory Access (RDMA) has been widely adopted in modern data centers thanks to its high-throughput, low-latency data transfer capability and reduced CPU overhead. However, traditional network-flow-based monitoring is poor at interpreting RDMA communication and hence inadequate for gaining insights. In this paper, we present INSERT, an end-to-end RDMA telemetry system that enables seamless visibility into RDMA communication from the network layer all the way to the application layer. To this end, INSERT combines (i) eBPF-based transparent RDMA tracing on end-hosts and (ii) stateful RDMA network telemetry on programmable data plane. We implement RDMA network telemetry on programmable SmartNICs, where we address practical challenges for maintaining fine-grained state on massively-parallel packet processing pipelines. We demonstrate that INSERT can perform reasonably accurate telemetry at line-rate for different types of RDMA traffic even in the presence of out-of-order packets, and finally showcase two practical use cases that can benefit from it.

It is commonly assumed that the end-to-end networking performance of edge offloading is purely dictated by that of the network connectivity between end devices and edge computing facilities, where ongoing innovation in 5G/6G networking can help. However, with the growing complexity of edge-offloaded computation and dynamic load balancing requirements, an offloaded task often goes through a multi-stage pipeline that spans across multiple compute nodes and proxies interconnected via a dedicated network fabric within a given edge computing facility. As the latest hardware-accelerated transport technologies such as RDMA and GPUDirect RDMA are adopted to build such network fabric, there is a need for good understanding of the full potential of these technologies in the context of computation offload and the effect of different factors such as GPU scheduling and characteristics of computation on the net performance gain achievable by these technologies. This paper unveils detailed insights into the latency overhead in typical machine learning (ML)-based computation pipelines and analyzes the potential benefits of adopting hardware-accelerated communication. To this end, we build a model-serving framework that supports various communication mechanisms. Using the framework, we identify performance bottlenecks in state-of-the-art model-serving pipelines and show how hardware-accelerated communication can alleviate them. For example, we show that GPUDirect RDMA can save 15-50% of model-serving latency, which amounts to 70–160 ms.

Increasingly, smart Network Interface Cards (sNICs) are being used in data centers to offload networking functions (NFs) from host processors thereby making these processors available for tenant applications. Modern sNICs have fully programmable, energy-efficient multi-core processors on which many packet processing functions, including a full-blown programmable switch, can run. However, having multiple switch instances deployed across the host hypervisor and the attached sNICs makes controlling them difficult and data plane operations more complex.

Location based social networks (LBSNs) are becoming increasingly popular with the fast deployment of broadband mobile networks and the growing prevalence of versatile mobile devices. This success has attracted great interest in studying and measuring the characteristics of LBSNs, such as Facebook Places, Yelp, and Google+ Local. However, it is often prohibitive, and sometimes too costly, to obtain a detailed and complete snapshot of a LBSN due to its usually massive scale. In this work, taking Foursquare as an example, we focus on sampling and estimating restricted geographic regions in LBSNs, such as a city or a country. By exploiting the application programming interfaces (APIs) provided by Foursquare for geographic search, we first introduce how to obtain the “ground truth”, namely, a complete set of all venues (i.e., places) in a specified region. Then, we propose random region sampling algorithms that allow us to draw representative samples of venues, and design unbiased estimators of regional characteristics of venues. We validate the efficiency of our sampling algorithms on Foursquare using complete datasets obtained from 12 regions, such as Switzerland, New York City and Los Angeles. Our results are applicable to perform sampling and estimation in all GeoDatabases, such as Facebook Places, Yelp, and Google+ Local, which have similar venue search APIs as Foursquare. These location service providers can also benefit from our results to enable efficient online statistic estimation.

In this paper, we provide a detailed analysis on the venue popularity in Foursquare, a leading location-based social network. By collecting 2.4 million venues from 14 geographic regions all over the world, we study the common characteristics of popular venues, and make the following observations. First, venues with more complete profile information are more likely to be popular. Second, venues in the Food category attract the most (43%) public tips (comments) by users, and the Travel & Transport category is the most popular category with the highest per venue check-ins, i.e., each venue in this category attracts on average 376 check-ins. Moreover, the stickiness of users checking in venues in the residence, office, and school categories is higher than in other categories. Last but not least, in general, old venues created at the early stage of Foursquare are more popular than new venues. Our results help to understand the factors that cause venues to become popular, and have applications in venue recommendations and advertisement in location based social networks.

Behavioral targeting of content to users is a huge and lucrative business, valued as a $20 billion industry that is growing rapidly. So far, the dominant players in this field like Google and Yahoo! examine the user requests coming to their servers and place appropriate ads based on the user's search keywords. Triple-play service providers have access to all the traffic generated by the users and can generate more comprehensive profiles of users based on their TV, broadband, and mobile usage. Using such multisource profile information, they can generate new revenue streams by smart targeting of ads to their users over multiple screens (computer, TV, and mobile handset). This paper proposes methods to place targeted ads to a TV based on user's interests. It proposes an ad auction model that can leverage multisource profile and can handle dynamic profile-based targeting like Google's AdWords vis-à-vis static demography-based targeting of legacy TV. We then present a 0.502-competitive revenue maximizing scheduling algorithm that chooses a set of ads in each time slot and assigns users to one of these selected ads.

Behavioral targeting of content to users is a huge and lucrative business, valued as a $20 billion industry that is growing rapidly. So far dominant players in this field like Google and Yahoo examine the user requests coming to their servers and place appropriate ads based on the user's search keywords. Triple play service providers have access to all the traffic generated by the users and can generate more comprehensive profiles of users based on their TV, broadband and mobile usage. Using such multi-source profile information they can generate new revenue streams by smart targeting of ads to their users over multiple screens (computer, TV and mobile handset). This paper proposes methods to place targeted ads to a TV based on user's interests. It proposes an ad auction model that can leverage multi-source profile and can handle dynamic profile-based targeting like Google's AdWords vis-a-vis static demography-based targeting of legacy TV. We then propose a 0.502-competitive revenue maximizing scheduling algorithm that chooses a set of ads in each time slot and assigns users to one of these selected ads.

We describe DipZoom (for "deep Internet performance zoom"), an approach to provide focused, on-demand Internet measurements. Unlike existing approaches that face a difficult challenge of building a measurement platform with sufficiently diverse measurements and measuring hosts, DipZoom offers a matchmaking service instead, which uses P2P concepts to bring together experimenters in need of measurements with external measurement providers. It then harnesses market forces to orchestrate the supply and demand sides in the resulting open eco-system. This paper outlines the overall design of DipZoom, and discusses payment, trust and security issues in the resulting open system.

Emerging multi-hop wireless mesh networks have much different characteristics than the Internet. They have low dimensionality and large diameters. Content and service replication can greatly improve their scalability. However, replication strategies in such networks have not been well studied. This paper studies the optimality of replication strategies and explores it in multi-hop wireless mesh networks for the first time. We start with the problem of determining the optimal numbers of replicas for a set of objects which have distinct probabilities of being requested in large 2-D mesh networks. We reveal the structure of the optimal replication strategy to minimize object access cost. To minimize average cost to access an object in 2-D mesh networks, the optimal strategy replicates an object such that the number of its replicas is proportional to p0.667, where p is the access probability of the object. This result indicates the inefficiency of demand-driven content and service replication in 2-D mesh networks, where an object is replicated such that the number of its replicas is proportional to p. We further study practical, online algorithms to approximate the optimal strategy. Interestingly, the optimal replication can be approximated well by a localized replacement algorithm. The algorithm utilizes only handy information and incurs no communication overhead. The paper demonstrates a significant performance gain by the optimal strategy, and the effectiveness of the online replacement algorithm.

We view congestion control as a distributed primal--dual algorithm carried out by sources and links over a network to solve a global optimization problem. We describe a multilink multisource model of the TCP Vegas congestion control mechanism. The model provides a fundamental understanding of delay, fairness and loss properties of TCP Vegas. It implies that Vegas stabilizes around a weighted proportionally fair allocation of network capacity when there is sufficient buffering in the network. It clarifies the mechanism through which persistent congestion may arise and its consequences, and suggests how we might use REM active queue management to prevent it. We present simulation results that validate our conclusions.