The growing prominence of LLMs has led to an increase in the development of AI tutoring systems.These systems are crucial in providing underrepresented populations with improved access to valuable education.One important area of education that is unavailable to many learners is strategic bargaining related to negotiation.To address this, we develop a LLM-based Assistant for Coaching nEgotiation (ACE).ACE not only serves as a negotiation partner for users but also provides them with targeted feedback for improvement.To build our system, we collect a dataset of negotiation transcripts between MBA students.These transcripts come from trained negotiators and emulate realistic bargaining scenarios.We use the dataset, along with expert consultations, to design an annotation scheme for detecting negotiation mistakes.ACE employs this scheme to identify mistakes and provide targeted feedback to users.To test the effectiveness of ACE-generated feedback, we conducted a user experiment with two consecutive trials of negotiation and found that it improves negotiation performances significantly compared to a system that doesn't provide feedback and one which uses an alternative method of providing feedback.

Maintaining a consistent persona is a key quality for any open domain dialogue system.Current state-of-the-art systems do this by training agents with supervised learning or online reinforcement learning (RL).However, systems trained with supervised learning often lack consistency as they are never punished for uttering contradictions.Additional training with RL can alleviate some of these issues, however the training process is expensive.Instead, we propose an offline RL framework to improve the persona consistency of dialogue systems.Our framework allows us to combine the advantages of previous methods as we can inexpensively train our model on existing data as in supervised learning, while punishing and rewarding specific utterances as in RL.We also introduce a simple importance sampling method to reduce the variance of importance weights in offline RL training which we call Variance-Reducing MLE-Initialized (VaRMI) importance sampling.Our automatic and human evaluations show that our framework improves both the persona consistency and dialogue quality of a state-of-the-art social chatbot.

Protecting large language models from privacy leakage is becoming increasingly crucial with their wide adoption in real-world products.Yet applying differential privacy (DP), a canonical notion with provable privacy guarantees for machine learning models, to those models remains challenging due to the trade-off between model utility and privacy loss.Utilizing the fact that sensitive information in language data tends to be sparse, Shi et al. (2021) formalized a DP notion extension called Selective Differential Privacy (SDP) to protect only the sensitive tokens defined by a policy function.However, their algorithm only works for RNN-based models.In this paper, we develop a novel framework, Just Fine-tune Twice (JFT), that achieves SDP for state-of-the-art large transformer-based models.Our method is easy to implement: it first finetunes the model with redacted in-domain data, and then fine-tunes it again with the original in-domain data using a private training mechanism.Furthermore, we study the scenario of imperfect implementation of policy functions that misses sensitive tokens and develop systematic methods to handle it.Experiments show that our method achieves strong utility compared to previous baselines.We also analyze the SDP privacy guarantee empirically with the canary insertion attack 1 .

Virtualization has attained mainstream status in enterprise IT industry. Despite its widespread adoption, it is known that virtualization also introduces non-trivial overhead when tasks are executed on a virtual machine (VM). In particular, a combined effect from device virtualization overhead and CPU scheduling latency can cause performance degradation when computation intensive tasks and I/O intensive tasks are co-located on a VM. Such an interference also causes extra energy consumption. In this paper, we present Hylics, a novel solution that enables efficient data traverse paths for both I/O and computation intensive workloads. This is achieved with the provision of in-memory file system and network service at the hypervisor level. Several important design issues are pinpointed and addressed during our prototype implementation, including efficient intermediate data sharing, network service offloading, and QoS-aware memory usage management. Based on our real-world deployment on KVM, we show that Hylics can significantly improve computation and I/O performance for hybrid workloads. Moreover, this design also alleviates the existing virtualization overhead and naturally optimizes the overall energy efficiency.

Today's explosively growing Internet video traffics and viewers' ever-increasing quality of experience (QoE) demands for video streaming bring tremendous pressures to the backbone network. As a new network paradigm, mobile edge caching provides a promising alternative by pushing video content closer at the network edge rather than the remote CDN servers so as to reduce both content access latency and redundant network traffic. However, our large-scale trace analysis shows that different from CDN based caching, edge caching environment is much more complicated with massively dynamic and diverse request patterns, which renders that existing rule-based and model-based caching solutions may not well fit such complicated edge environments. Moreover, although cooperative caching has been proposed to better afford limited storage on each individual edge server, our trace analysis also shows that the request similarity among neighboring edges can be highly dynamic and diverse, which is drastically different from CDN based caching environment, and can easily compromise the benefits from traditional cooperative caching mostly designed based on CDN environment. In this paper, we propose MacoCache, an intelligent edge caching framework that is carefully designed to afford the massively diversified and distributed caching environment to minimize both content access latency and traffic cost. Specifically, MacoCache leverages a multi-agent deep reinforcement learning (MADRL) based solution, where each edge is able to adaptively learn its own best policy in conjunction with other edges for intelligent caching. The real trace-driven evaluation further demonstrates that MacoCache is able to reduce an average of 21% latency and 26% cost compared with the state-of-the-art caching solution.

Mobile augmented reality (MAR) augments a real-world environment (probably surrounding or close to the mobile user) by computer-generated perceptual information. Utilizing the emerging edge computing paradigm in MAR systems can reduce the power consumption and computation load for the mobile devices and improve responsiveness of the MAR service. Different from existing studies that mainly explored how to better enable the MAR services utilizing edge computing resources, our focus is to optimize the video generation stage of the edge-based MAR services-efficiently using the available edge computing resources to render and encode the augmented reality as video streams to the mobile clients. Specifically, for multi-party AR applications, we identify the advantages and disadvantages of two encoding schemes, namely colocated encoding and spilt encoding, and examine the trade-off between performance and scalability when the rendering and encoding tasks are colocated or split. Towards optimally placing AR video rendering and encoding in the edge, we formulate and solve the rendering and encoding task assignment problem for multi-party edge-based MAR services to maximize the QoS for the users and the edge computing efficiency. The proposed task assignment scheme is proved to be superior through extensive trace-driven simulations and experiments on our prototype system.

Data enrichment is the act of extending a local database with new attributes from external data sources. In this paper, we study a novel problem-how to progressively crawl the deep web (i.e., a hidden database) through a keyword-search API to enrich a local database in an e ective way. This is chal- lenging because these interfaces often limit the data access by enforcing the top-k constraint or limiting the number of queries that can be issued within a time window. In response, we propose SmartCrawl, a new framework to collect re- sults e ectively. Given a query budget b, SmartCrawl rst constructs a query pool based on the local database, and then iteratively issues a set of most bene cial queries to the hidden database such that the union of the query results can cover the maximum number of local records. The key technical challenge is how to estimate query bene t, i.e., the number of local records that can be covered by a given query. A simple approach is to estimate it as the query frequency in the local database. We nd that this is ine ective due to i) the impact of |ΔD|, where |ΔD| represents the number of local records that cannot be found in the hidden database, and ii) the top-k constraint enforced by the hidden database. We study how to mitigate the negative impacts of the two factors and propose e ective optimization techniques to improve performance. The experimental results show that on both simulated and real-world hidden databases, SmartCrawl signi cantly increases coverage over the local database as compared to the baselines.

The past decade has witnessed significant breakthroughs in the world of computer vision. Recent deep learning-based computer vision algorithms exhibit strong performance on recognition, detection, and segmentation. While the development of vision algorithms elicits promising applications, it also presents immense computational challenge to the underlying hardware due to its complex nature, especially when attempting to process the data at line-rate.

Data scientists often spend more than 80% of their time on data preparation. Data enrichment, the act of extending a local database with new attributes from external data sources, is among the most time-consuming tasks. Existing data enrichment works are resource intensive: data-intensive by relying on web tables or knowledge bases, monetarily-intensive by purchasing entire datasets, or time-intensive by fully crawling a web-based data source. In this work, we explore a more targeted alternative that uses resources (in terms of web API calls) proportional to the size of the local database of interest. We build Deeper, a data enrichment system powered by the deep web. The goal of Deeper is to help data scientists to link a local database to a hidden database so that they can easily enrich the local database with the attributes from the hidden database. We find that a challenging problem is how to crawl a hidden database. This is different from a typical deep web crawling problem, whose goal is to crawl the entire hidden database rather than only the content relating to the data enrichment task. We demonstrate the limitations of straightforward solutions and propose an effective new crawling strategy. We also present the Deeper system architecture and discuss how to implement each component. During the demo, we will use Deeper to enrich a publication database and aim to show that (1) Deeper is an end-to-end data enrichment solution, and (2) the proposed crawling strategy is superior to the straightforward ones.

With the prevalence of broadband network and wireless mobile network accesses, distributed interactive applications (DIAs) such as online gaming have attracted a vast number of users over the Internet. The deployment of these systems, however, comes with peculiar hardware/software requirements on the user consoles. Recently, such industrial pioneers as Gaikai, Onlive, and Ciinow have offered a new generation of cloud-based DIAs (CDIAs), which shifts the necessary computing loads to cloud platforms and largely relieves the pressure on individual user's consoles. In this paper, we aim to understand the existing CDIA framework and highlight its design challenges. Our measurement reveals the inside structures as well as the operations of real CDIA systems and identifies the critical role of cloud proxies. While its design makes effective use of cloud resources to mitigate client's workloads, it may also significantly increase the interaction latency among clients if not carefully handled. Besides the extra network latency caused by the cloud proxy involvement, we find that computation-intensive tasks (e.g., game video encoding) and bandwidth-intensive tasks (e.g., streaming the game screens to clients) together create a severe bottleneck in CDIA. Our experiment indicates that when the cloud proxies are virtual machines (VMs) in the cloud, the computation-intensive and bandwidth-intensive tasks may seriously interfere with each other. We accordingly capture this feature in our model and present an interference-aware solution. This solution not only smartly allocates workloads but also dynamically assigns capacities across VMs based on their arrival/departure patterns.

Combining advanced sensors and powerful processing capabilities smart-phone based augmented reality (AR) is becoming increasingly prolific. The increase in prominence of these resource hungry AR applications poses significant challenges to energy constrained environments such as mobile-phones.; [email protected] that end we present a platform for offloading AR applications to powerful cloud servers. We implement this system using a thin-client design and explore its performance using the real world application Pokemon Go as a case study. We show that with careful design a thin client is capable of offloading much of the AR processing to a cloud server, with the results being streamed back. Our initial experiments show substantial energy savings, low latency and excellent image quality even at relatively low bit-rates.

Unmanned aerial vehicles (UAVs) are experiencing a major increase in popularity in both consumer and industrial markets as prices fall and the technology matures. No longer are drones limited to military purposes as manufacturers begin to mass produce civilian models, ushering in a new era of transportation technology. While consumer drones are still in their infancy stage, there is little in the way of rules and regulations regarding privacy issues of these new devices. In this paper, we design an energy efficient off-the-shelf hardware system capable of detecting and selectively disabling video feeds of WiFi based consumer drones if they enter a defended area.

Recent years have witnessed the rapid growth of new live broadcast services, represented by Twitch.tv and YouTube live events, where videos are crowdsourced from amateur users (e.g., game players), rather than from commercial and professional TV broadcaster or content providers. The viewers also actively contribute to the content through embedded open-chat channels. Such community interactions among viewers, or even between broadcasters and viewers, make content generation highly diversified and engaging, particularly for the young generation. In this context, cross-viewer synchronization is highly desirable; otherwise the viewers with shorter broadcast latency may act as spoilers, significantly affecting the user experience of other viewers. In this paper, we show that the end-to-end delay has a dramatically amplified impact on the broadcast latency for individual viewers. We suggest smart rate adaptation to achieve cross-viewer synchronization, and develop distributed algorithms based on dual decomposition. We further extend our solution to the cloud environment, and present the concept of ShadowCast, which moves broadcasters to the cloud to provide high-quality streams beyond broadcasters' network bandwidth constraint. Its practicability and effectiveness is demonstrated by our implementation and test bed experiments.

Virtualization has become a building block for modern IT industry, and many datacenters are now highly virtualized. It is known that virtualization also introduces non-trivial overhead, which can cause severe self-interference inside a VM when CPU intensive tasks and bandwidth intensive tasks are co-located. Energy efficiency of the server can be affected as well. While such overhead is well-studied in application/protocol specific context, a more comprehensive solution is yet to be explored for general cloud services. In this paper, we present MemNet, a novel protocol-independent solution that enables para-virtualized memory sharing between host and guest VMs. This design successfully decouples I/O and computation operations and lifts the offered interface from the physical devices to high-level network services. Our real-world implementation on KVM indicates that, MemNet can achieve 27% and 70% gain in terms of computing and networking performance, respectively, by resolving the self-interference. It also provides 32% improvement in terms of energy efficiency.

In this article, we have classified cloud gaming platforms into three types based on how games are integrated with platforms. We have also reviewed the history of cloud gaming services, and noted that it is a key moment for cloud gaming services to increase their penetration rates. Last, built upon our extensive research experience in cloud gaming, we share several of our visions into future cloud gaming technologies, business models, and social impacts, in the format of forecasts. While our forecasts may not be an exhausted list, we firmly believe this article will stimulate more discussions among cloud gaming researchers and practitioners, resulting in a sustainable cloud gaming ecosystem.

Motivated by our systematic study on the diverse aspects of migrating gaming services to a virtualized cloud environment, we designed and implemented a fully virtualized cloud gaming platform, Rhizome, utilizing the latest hardware support for both remote servers and local clients. Our platform takes the first step towards bridging online gaming systems and public clouds. To accomplish ultra-low latency and a low power consumption gaming experience, we further optimized Rhizome's thin-client configuration and its interaction modules. In this proposed demo, we demonstrate that gaming over a virtualized cloud can be made possible with careful optimization and integration of different modules. It also helps us reveal the critical challenges towards full-fledged deployment of gaming services over public virtualized cloud.

Recent years have witnessed the emergence and growth of Cloud Gaming, where players interact with the remote game instance and receive rendered game scenes in video stream. Meanwhile, broadcasting and viewing games through live streaming platforms, e.g., Twitch.tv, have become increasingly popular. The interaction and performance of the many modules involved in this new generation of gaming and streaming platforms have yet to be closely investigated. In this paper, we present an initial experiment-based performance study, in which we profile the architecture of realworld gaming and streaming platforms, namely the Open Broadcast Software (OBS) module and its connection to the Twitch server. Our investigation shows that the recording operation can greatly increase the CPU utilization and the power consumption can increase over 60% on the game streaming computer. The use of advanced hardware encoding found on modern GPUs can greatly alleviate these performance issues. Yet, through profiling, we show that hardware encoding can introduce remarkable delays to the whole pipeline. We track this to a complicated interplay between the CPUs power saving methods and the implementation of hardware encoders.

Existing cloud gaming platforms have mainly focused on private nonvirtualized environments with proprietary hardware. Modern public cloud platforms heavily rely on virtualization for efficient resource sharing, the potentials of which have yet to be explored. Migrating gaming to a public cloud is nontrivial, however, particularly considering the overhead for virtualization and that the graphics processing units (GPUs) for game rendering has long been an obstacle in virtualization. This paper takes a first step toward bridging the online gaming system and the public cloud platforms. We present the design and implementation of a fully virtualized cloud gaming platform with the latest hardware support for both remote servers and local clients. We explore many critical design issues inherent in cloud gaming, including the choice of hardware or software video encoding, and the configuration and the detailed power consumption of thin client. We demonstrate that with the latest hardware and virtualization support, gaming over virtualized cloud can be made possible with careful optimization and integration of the different modules. We also highlight critical challenges toward full-fledged deployment of gaming services over the public virtualized cloud.

Distributed interactive applications (DIAs) such as online gaming have attracted a vast number of users over the Internet. It is however known that the deployment of DIA systems comes with peculiar hardware/software requirements on the users' consoles. Recently, such industrial pioneers as Gaikai, Onlive and Ciinow have offered a new based distributed interactive applications generation of cloud (CDIAs), which shift the necessary computing loads to cloud platforms and largely relieve the pressure on individual user consoles. In this paper, we take a first step towards understanding the CDIA framework and highlight its design challenges. Our measurement reveals the inside structure as well as the operations of real CDIA systems and identifies the critical role of the cloud proxies. While this design makes effective use of cloud resources to mitigate the clients' workloads, it can also significantly increase the interaction latency among clients if not carefully handled. Besides the extra network latency due to the involvement of cloud proxies, we find that the computation-intensive tasks (e.g., Game rendering) and bandwidth-intensive tasks (e.g., Streaming the game screen to the clients) together create a severe bottleneck in CDIA. Our experiment indicates that when the cloud proxies are virtual machines (VMs) in the cloud, the computation-intensive and bandwidth-intensive tasks will seriously interfere with each other if not handled carefully. We accordingly capture this feature in our model and present an interference-aware solution. This approach not only smartly allocates the workloads but also dynamically assigns the capacities across VMs.

There have been significant studies on virtual machines (VMs), including their power consumption in performing different types of tasks. The VM's power consumption with network transactions, however, has seldom been examined. This paper presents an empirical study on the power consumption of typical virtualization packages while performing network tasks. We find that both Hardware Virtualization and Paravirtualization add considerable energy overhead, affecting both sending and receiving, and a busy virtualized web-server may consume 40% more energy than its non-virtualized counterparts. Our detailed profiling on packet path reveals that a VM can take 5 times more cycles to deliver a packet than a bare-metal machine, and is also much less efficient on caching. Without fundamental changes to the hypervisor-based VM architecture, we show that the use of adaptive packet buffering potentially reduces the overhead. Its practicality and effectiveness in power saving are validated through driver-level implementation and experiments.

Existing research on cloud network (in)stability has primarily focused on communications between Virtual Machines (VMs) inside a cloud, leaving that of VM communications over higher-latency wide-area networks largely unexplored. Through measurement in real-world cloud platforms, we find that there are prevalent and significant degradation and variation for such VM communications with both TCP and UDP traffic, even over lightly utilized networks. Our in-depth measurement and detailed system analysis reveal that the performance variation and degradation are mainly due to the dual-role of the CPU in both computation and network communication in a VM, and they can be dramatically affected by the CPU's scheduling policy. We provide strong evidence that such issues can be addressed in the hypervisor level and present concrete solutions. Such remedies have been implemented and evaluated in our cloud testbed, showing noticeable improvement for long-haul network communications with VMs.

Virtualization, which allows multiple Virtual Machines (VMs) to reside on a single physical machine, has become an indispensable technology for today's IT infrastructure. It is known that the overhead for virtualization affects system performance; yet it remains largely unknown whether VMs are more vulnerable to networked Denial of Service (DoS) attacks than conventional physical machines. A clear understanding here is obviously critical to such networked virtualization system as cloud computing platforms. In this paper, we present an initial study on the performance of modern virtualization solutions under DoS attacks. We experiment with the full spectrum of modern virtualization techniques, from paravirtualization, hardware virtualization, to container virtualization, with a comprehensive set of benchmarks. Our results reveal severe vulnerability of modern virtualization: even with relatively light attacks, the file system and memory access performance of VMs degrades at a much higher rate than their non-virtualized counterparts, and this is particularly true for hypervisor-based solutions. We further examine the root causes, with the goal of enhancing the robustness and security of these virtualization systems. Inspired by the findings, we implement a practical modification to the VirtIO drivers in the Linux KVM package, which effectively mitigates the overhead of a DoS attack by up to 40%.

Powered by cloud computing, Dropbox not only provides reliable file storage but also enables effective file synchronization and user collaboration. This new generation of service, beyond conventional client/server or peer-to-peer file hosting with storage only, has attracted a vast number of Internet users. It is however known that the synchronization delay of Dropbox-like systems is increasing with their expansion, often beyond the accepted level for practical collaboration. In this paper, we present an initial measurement to understand the design and performance bottleneck of the proprietary Dropbox system. Our measurement identifies the cloud servers/instances utilized by Dropbox, revealing its hybrid design with both Amazon's S3 (for storage) and Amazon's EC2 (for computation). The mix of bandwidth-intensive tasks (such as content delivery) and computation-intensive tasks (such as compare hash values for the contents) in Dropbox enables seamless collaboration and file synchronization among multiple users; yet their interference, revealed in our experiments, creates a severe bottleneck that prolongs the synchronization delay with virtual machines in the cloud, which has not seen in conventional physical machines. We thus re-model the resource provisioning problem in the Dropbox-like systems and present an interference-aware solution that smartly allocates the Dropbox tasks to different cloud instances. Evaluation results show that our solution remarkably reduces the synchronization delay for this new generation of file hosting service.