Embedding models are effective for learning high-dimensional sparse data. Traditionally, they are deployed in DRAM parameter servers (PS) for online inference access. However, the ever-increasing model capacity makes this practice suffer from both high storage costs and long recovery time. Rapidly developing Persistent Memory (PM) offers new opportunities to PSs owing to its large capacity at low costs, as well as its persistence, while the application of PM also faces two challenges including high read latency and heavy CPU burden. To provide a low-cost but still high-performance parameter service for online inferences, we introduce PetPS, the first production-deployed PM parameter server. (1) To escape with high PM latency, PetPS introduces a PM hash index tailored for embedding model workloads, to minimize PM access. (2) To alleviate the CPU burden, PetPS offloads parameter gathering to NICs, to avoid CPU stalls when accessing parameters on PM and thus improve CPU efficiency. Our evaluation shows that PetPS can boost throughput by 1.3 -- 1.7X compared to PSs that use state-of-the-art PM hash indexes, or get 2.9 -- 5.5X latency reduction with the same throughput. Since 2020, PetPS has been deployed in Kuaishou, one world-leading short video company, and successfully reduced TCO by 30% without performance degradation.

Video accounts for a large proportion of traffic on the Internet. Understanding its geographical viewing patterns is extremely valuable for the design of Internet ecosystems for content delivery, recommendation and ads. While previous works have addressed this problem at coarse-grain scales (e.g., national), the urban-scale geographical patterns of video access have never been revealed. To this end, this article aims to investigate the problem that whether there exists distinct viewing patterns among the neighborhoods of a large-scale city. To achieve this, we need to address several challenges including unknown of patterns profiles, complicate urban neighborhoods, and comprehensive viewing features. The contributions of this article include two aspects. First, we design a framework to automatically identify geographical video viewing patterns in urban neighborhoods. Second, by using a dataset of two months real video requests in Shanghai collected from one major ISP of China, we make a rigorous analysis of video viewing patterns in Shanghai. Our study reveals the following important observations. First, there exists four prevalent and distinct patterns of video access behavior in urban neighborhoods, which are corresponding to four different geographical contexts: downtown residential, office, suburb residential and hybrid regions. Second, there exists significant features that distinguish different patterns, e.g., the probabilities of viewing TV plays at midnight, and viewing cartoons at weekends can distinguish the two viewing patterns corresponding to downtown and suburb regions.

Combining Network Functions Virtualization (NFV) with Software-Defined Networking (SDN) is an emerging and promising solution to provide scalable and elastic network control and service. In such a system, virtualized Network Functions (NFs) need to be consistently migrated from one instance to another for various purposes, such as resource optimization, fault tolerance, load balancing, etc. These migrations involve simultaneously coordinating updates to the NF state and SDN forwarding state. To solve this problem, we design two consistent NF state update schemes: a controller-forwarding based scheme and a tagging-based scheme. Through analysis of the update process, we demonstrate that they both guarantee loss-free and order-preserving migrations. We further implement a prototype and carry out experiments with diverse traffic settings. Results demonstrate that the controller-forwarding based solution achieves 77 percent migration time compared with the state-of-the-art solution OpenNF, while correcting an error of it. Moreover, the tagging-based solution not only achieves 4.4 percent migration time, but also reduces up to 75 percent controller overhead compared with OpenNF at the cost of adding a tag in the unused fields of packet header.

Accurate and comprehensive end-to-end network performance measurement is critical for the automatic troubleshooting and optimized provision of various services in Cloud. However, cloud providers and tenants still rely on rudimentary and separate tools for end-to-end performance measurement, which are inflexible, tedious, and error-prone. In this paper, we present NetWatch, a system that provides measurement as a service through open APIs for both cloud providers and tenants to measure end-to-end performance on-demand. In this system, measurement requests are first delivered to NetWatch Controller by open APIs, which transforms the request to configure specific Probes to fulfill the requests by active measurement. We make delicate design choices and address several challenges to enable NetWatch offering accurate and low-overhead measurement service for multiple users simultaneously and efficiently. A prototype implementation and experiments with diverse network settings link and traffic demonstrate that NetWatch can support flexible and accurate measurement of end-to-end network performance with small overhead.

Recent proliferation of mobile networks and smart devices drives the rapid growth of mobile video traffic. Caching popular video content at any possible place of the network near to users could significantly increase their delivery efficiency. However, fundamental problems of how cache behaves and what is the principle for cache deployment in a mobile network under large-scale video views are still unknown, which include three closely relevant problems: 1) what is the best scale of regions to deploy cache appliances; 2) how many contents should be cached; and 3) which contents should be cached. In this paper, we synthetically study these problems by analyzing 10 million video view requests of six most popular content providers, in the city of Shanghai, China. We first aggregate videos from different providers by topics to measure user interests, and divide the city into nonoverlapping regions of different sizes to investigate the influence of scale. Then, we define metrics of view concentration, popular topic number, cache revenue, and popular topic similarity to quantitatively characterize cache behaviors and consequently answer the three problems. Our studies reveal that: 1) it is effective to deploy cache in regions of a wide range of different scales; 2) the larger scale region and the regions with more views should cache more contents; and 3) different regions, especially small scale ones, should cache different contents. Furthermore, based on trace-driven evaluation, we show that the overall cache hit ratio can increase by up to 30% when we apply above guidelines for cache deployment.

With the popularity of watching mobile videos, a major form of multimedia contents, many works focus on the geographic features of user viewing behaviors, but few study them in the context of an entire metropolitan city. Different regions of a large city have different intensity of economy activities with respect to their different distances to the downtown, and how this will influence video popularity and similarity is still unclear. To quantitatively study the spatial popularity and similarity of watching videos in a large urban environment, we collect a dataset with two-month video view requests from the largest network provider in Shanghai, containing the top six content providers, and study the spatial features of video access in regions of different scales. We find that: 1) video popularity and similarity exist at different scales of city division; 2) the concentration of video popularity becomes higher as the region is closer to downtown; and 3) when comparing the regions of same scale, the similarity of popular videos becomes lower as the region is farther away from the downtown. Finally, we correlate our findings with cache deployment, advertising, and video recommendation to illustrate the implications.

Previous works have proposed various approaches to implement service chaining by routing traffic through the desired middleboxes according to pre-defined policies. However, no matter what routing scheme is used, the performance of service chaining depends on where these middleboxes are placed. Thus, in this paper, we study middlebox placement problem, i.e., given network information and policy specifications, we attempt to determine the optimal locations to place the middleboxes so that the performance is optimized. The performance metrics studied in this paper include the end-to-end delay and the bandwidth consumption, which cover both users’ and network providers’ interests. We first formulate it as 0-1 programming problem, and prove it is NP-hard. We then propose two heuristic algorithms to obtain the sub-optimal solutions. The first algorithm is a greedy algorithm, and the second algorithm is based on simulated annealing. Through extensive simulations, we show that in comparison with a baseline algorithm, the proposed algorithms can reduce 22 percent end-to-end delay and save 38 percent bandwidth consumption on average. The formulation and proposed algorithms have no special assumption on network topology or policy specifications, therefore, they have broad range of applications in various types of networks such as enterprise, data center and broadband access networks.

With the popularity of watching mobile videos, many works focus on the geographic features of user viewing behaviors, but few study them in the context of an entire metropolitan city. Different regions of a large city have different intensity of economy activities with respect to their different distances to the downtown, and how this will influence video popularity and similarity is still unclear. To quantitatively study the spatial popularity and similarity of watching videos in a large urban environment, we collect a dataset with two-month video view requests from the largest network provider in Shanghai, containing top six content providers, and study the spatial features of video access in regions of different scales. We find that 1) video popularity and similarity exist at different scales of city division; 2) the concentration of video popularity becomes higher as the region is closer to downtown; 3) when comparing the regions of same scale, the similarity of popular videos becomes lower as the region is farther away from the downtown. Finally, we correlate our findings with cache deployment, advertising and video recommendation to illustrate the implications.

Various network optimization and management goals in cloud data centers can be achieved by re-mapping VMs to the substrate servers. Live VM migration is used to implement such re-mapping by moving VMs from the initial servers to the target ones. For efficiency and usability considerations, these migration tasks are expected to be completed as soon as possible, which can be achieved by planning multiple VMs to be migrated simultaneously. However, the available resources of computation and bandwidth limits the number of VMs that can be migrated at the same time. Besides, the available resources change with the progress of migration process, which makes VM migration planning a challenge problem. Considering both resources constraints and the dependence between the migration plan and the total migration time, in this paper, we investigate the problem of virtual machine migration planning to minimize the total migration time. We formulate VM migration planning as an optimization problem considering both computation and bandwidth constraints. Our formulation is based on a step-by-step migration scheme, where multiple VMs can be migrated simultaneously as long as the resource constraints are satisfied. The solution of the optimization problem outputs the migration plan which achieves the minimum time to finish the migration task. Moreover, we verify the effectiveness of the optimal migration plan through extensive simulations and comparison with the best algorithm in existing works.

No abstract available.

Wireless Virtualization enables multiple concurrent virtual networks running on shared wireless substrate resource, which makes networks more flexible, efficient and customizable. Efficiently allocating physical wireless spectrum to multiple virtual networks to enhance the resource utilization is a fundamental challenge. Different from previous works focusing on only one physical wireless network, we study the problem under the multiple physical networks scenario, which makes the problem more general and practical. Furthermore, after analyzing the fluctuation feature in wireless virtualization, we introduce an opportunistic spectrum sharing method to increase spectrum allocation efficiency. First, we formulate the problem of opportunistic spectrum sharing with multiple physical wireless networks and prove it as an NP-Hard problem. Then, we propose a genetic algorithm and heuristic algorithm to solve the problem. Simulations show that opportunistic spectrum sharing conspicuously improves the performance of spectrum utilization and confirm the advantage of our proposed algorithms.

Despite of various benefits such as load balance and energy saving virtual machine (VM) migration promises to provide, its application in realistic data centers is still limited due to the limitation of migration in the LAN environment and the unpredictable performance. Through separation of control plane and data plane, software defined network (SDN) provides the possibility for an alternate solution to overcome these limitations. In this work, we aim at designing and implementing a software defined approach for live VM migration to experiment the possibility. This paper presents the key challenges and our preliminary design result to address these challenges.